How to Scan a Handwritten Revision Pack into ChatGPT and Turn It into a 30-Minute Study Session

Current source date: 5 October 2026. You can use an iPhone to capture a finite pack of permitted handwritten notes, test whether ChatGPT can actually read each page, and then prepare a source-bounded 30-minute Study-mode session. The essential safeguard is to treat the scanned Portable Document Format (PDF)A fixed-layout document format used to preserve page appearance across systems. Open glossary entry file as an intake route, not as proof that the handwriting was understood: request a literal page-by-page legibility check first, replace any uncertain PDF page with a clear Joint Photographic Experts Group (JPEG)JPEG is a widely used compressed still-image format named after the Joint Photographic Experts Group committee that developed the standard. JPEG photos commonly have a .jpeg or .jpg filename extension; the two extensions do not imply different photographic standards. Open glossary entry or Portable Network Graphics (PNG)An image file format for losslessly compressed raster pictures, including pictures with transparent areas. It is not proof that an uploaded page is legible to a model. Open glossary entry image, and compare important claims with the visible original before practising.

Evidence checkpoints

Documented point: The ChatGPT camera has a Scan option under the three-dot menu. Source accessed 5 October 2026. [ChatGPT release notes]

Documented point: As documented on 5 October 2026, Study mode is available across ChatGPT plans and models, and on web, iOS and Android, subject to normal plan, message and upload limits. OpenAI says it is unavailable in Temporary Chats, custom GPTs (user-built ChatGPT assistants) and Projects. Source accessed 5 October 2026. [OpenAI documentation: Using study mode in ChatGPT]

Documented point: Still-image inputs are available on Free and paid plans, on web and mobile platforms, subject to usage limits and settings. Source accessed 5 October 2026. [OpenAI documentation: ChatGPT image inputs FAQ]

Documented point: File uploads are available to Free and paid ChatGPT plans on web and supported mobile apps, but usage depends on plan and account settings. Source accessed 5 October 2026. [File Uploads FAQ]

Documented point: ChatGPT Enterprise can read text and visual elements embedded in an uploaded PDF. Source accessed 5 October 2026. [OpenAI documentation: Visual retrieval with PDFs FAQ]

Documented point: Settings > Data controls provides the Improve the model for everyone control. Source accessed 5 October 2026. [Data controls in ChatGPT]

Documented point: For individual services such as ChatGPT, OpenAI may use content to train models unless the user chooses otherwise. Source accessed 5 October 2026. [OpenAI documentation: How your data is used to improve model performance]

Documented point: Regular and archived chats remain in an account until deleted or removed by applicable workspace policy. Source accessed 5 October 2026. [OpenAI documentation: Chat and file retention policies in ChatGPT]

1. What changed on 1 October—and what this tutorial is and is not

The new iPhone intake route

OpenAI’s 1 October 2026 ChatGPT release notes documented a Scan option in the ChatGPT phone camera’s three-dot menu. It can capture several pages consecutively and combine them into one PDF ready to upload to a chat. OpenAI described Scan as rolling out on iOS; that wording does not establish availability on every iPhone, account, plan or region, and it does not establish an Android or web equivalent.

Procedure: before beginning, choose a small, finite pack that fits one revision objective—for example, four pages covering simultaneous equations rather than an entire mathematics folder. In a regular ChatGPT conversation on an iPhone, open the camera and inspect its three-dot menu for Scan. If it is present, capture the pages in their intended order and review the resulting PDF before attaching it. If Scan is absent, do not spend the session troubleshooting an undocumented rollout schedule: use + > Add photos & files and attach clear page photographs instead. Follow the upload and usage limits shown by your own account because limits and settings can vary and may change.

Example: a permitted four-page pack might contain “Page 1: definitions”, “Page 2: worked example”, “Page 3: graph”, and “Page 4: practice questions”. Add those labels lightly in the page margin before capture if they are not already numbered. The labels make later verification precise without changing the educational content. Do not relabel a textbook’s printed page number; use a separate intake label such as “Pack page 2”.

Use Scan when it appears, because consolidating several pages makes capture orderly, but retain the original pages and be ready to provide individual photographs. Use individual images immediately when the pack contains faint pencil, dense formulae, diagrams whose layout matters, marginal annotations, or pages that cannot be captured sharply. The PDF is operationally convenient; individual images can be materially better evidence for a page that needs visual interpretation.

After capture, check the preview yourself. Confirm that the first and last lines, page number, headings, symbols, graph labels and all four margins are visible. Recapture any folded, shadowed, blurred, clipped or sideways page. A visually present page can still be misread by the model, so this review is necessary but not sufficient.

The first decision is PDF versus individual source photos

The Scan-created PDF has an important plan-dependent limitation. OpenAI’s File Uploads FAQ, as read on 5 October 2026, says that plans other than Enterprise use text-based retrieval for documents and PDFs: ChatGPT extracts digital text and discards embedded images. OpenAI separately documents visual retrieval of text and visual elements embedded in PDFs as an Enterprise capability. A PDF assembled from photographs of handwriting may contain little or no usable digital text. Therefore, a successful upload is not evidence that a non-Enterprise chat can see the handwriting, diagrams or layout inside it.

This differs from attaching a supported still image directly. OpenAI’s Image Inputs FAQ documents image input on Free and paid plans, subject to account limits and settings, and advises enlarging text without cropping important details. It also warns that unclear images are less accurate, images are resized, and rotated text can be misinterpreted. Image availability is not a guarantee of accurate handwriting recognition, but it provides the correct fallback when a PDF page has not passed a literal reading check.

Use this legibility gate before requesting any quiz:

1. Attach the Scan-generated PDF, if you used Scan.
2. Ask ChatGPT to identify every page separately and reproduce only a small set of discriminating details: the page label, exact heading, one representative handwritten line and each displayed formula.
3. Require it to mark anything unreadable as “unclear” rather than reconstructing likely wording.
4. Compare that response with every physical page or original photograph.
5. If a heading, line, sign, exponent, denominator, diagram label or page is missing or uncertain, attach that original page as one upright JPEG or PNG and repeat the check for that page.
6. Proceed to practise only after the pages needed for the session pass your comparison.

An example checking request is: Before teaching or quizzing me, inspect the supplied pack page by page. For each page, give its page label, exact heading, the first complete handwritten sentence and every displayed formula. Do not repair or infer missing text. Write “unclear” beside anything you cannot read, and write “page not available” if you cannot access a page. This is a suggested diagnostic method, not a guarantee that the response will expose every recognition error.

Worked routing example: suppose the PDF response correctly identifies pack pages 1, 2 and 4 but reports no content for page 3, which contains a hand-drawn graph. The correct action is not to ask a question about the graph and judge readability from the apparent quality of the answer. Attach a clear image of page 3 and ask: Check this source image literally. State the page label and heading, transcribe both axis labels, list each plotted coordinate you can read, and mark all doubtful characters as “unclear”. Do not solve the exercise. Compare every reported coordinate with the original. If one coordinate remains doubtful, recapture the page closer and more squarely or exclude that graph from the session.

When a page image still fails, place it flat under even light, hold the phone parallel to the sheet, use one page per image, keep all margins in frame, increase contrast where possible and ensure the writing is large enough to inspect. Do not submit a collage of four pages merely to reduce upload count; shrinking the writing undermines the legibility check. If the handwriting itself is too faint or compressed, type the relevant permitted line into the chat and label it as your manual transcription, then compare the typed version once more with the original.

Keep the PDF as the working source only where the literal check demonstrates access to the content required for this session. Replace a failed page with an individual source photo rather than assuming that a fluent summary means the page was read. If the model answers from general knowledge, that may conceal missing source access.

Personally compare headings, definitions, mathematical operators, units and diagram labels. Prioritise characters whose substitution changes meaning: “−” versus “+”, “x” versus “×”, “1” versus “7”, superscripts, subscripts, inequality signs and chemical coefficients. For consequential educational decisions—such as deciding that a topic is mastered, submitting assessed work or challenging a mark—consult the original material and, where appropriate, a teacher or qualified human reviewer.

The physical-page comparison remains the test for whether this particular handwritten pack is usable; for high-level background on vision, document processing and optical character recognition, see GPT-5.6 Sol Vision Capabilities, an explainer of GPT-5.6 Sol vision capabilities rather than a checklist for photographing handwriting or verifying literal transcription.

A 30-minute review, not a durable study system

The target is one bounded session based only on the verified pages. It is not a flashcard programme, a spaced-repetition schedule, a semester archive, a research workflow or a claim that ChatGPT has checked the notes against an external syllabus. Keeping the objective narrow reduces both intake time and the chance that unsupported outside knowledge will be presented as if it appeared in the pack.

Procedure: define one outcome before capture in the form “By the end of this session, I want to identify and practise…”. Suitable examples include “the three conditions for natural selection on these four pages” or “the method used in the two quadratic-equation examples”. Avoid a broad objective such as “revise biology”. Set aside only the pages needed for that outcome and remove unrelated sheets.

After the legibility gate, the later Study-mode request should impose explicit evidence boundaries: use only confirmed readable pages; attach an exact pack-page and heading reference to feedback; say “not in the supplied pages” when support is absent; and ask one question at a time. OpenAI’s Study mode guidance, as read on 5 October 2026, says that a learner can ask for one question at a time, have the system wait for an answer and receive an explanation of what to review next. Study mode can nevertheless make mistakes or sometimes answer directly, so the instruction remains a preference rather than a behavioural guarantee.

Example boundary: if the pack contains a teacher’s permitted notes on photosynthesis but does not state a wavelength for chlorophyll absorption, the session should return “not in the supplied pages” rather than importing a plausible value. If that omitted fact matters, verify it separately with an authorised course source or teacher; do not quietly expand the evidence boundary during the timed review.

Continue only when the selected pack is narrow enough to inspect and practise within 30 minutes. If page verification consumes most of the available time, reduce the pack rather than skipping verification. A verified two-page session is preferable to a six-page session whose source access is uncertain.

Before entering Study mode, write down the included pack-page range and topic. Check that ChatGPT’s stated topic matches your own reading of the pages. If it assigns the wrong subject, chapter or method, repair the source intake first; do not rely on later questions to correct a faulty identification.

Only after that page-by-page verification can the learner adapt the already verified material into quizzes, flashcards or a planned review sequence using 25 ChatGPT Prompts for Interactive Learning and Study, a 25-prompt collection covering interactive learning, spaced repetition, adaptive study plans and feedback; it does not replace this article’s source pointers or physical-page checks.

2. Before you capture: permission, privacy and the correct chat surface (2 minutes)

Minute 0–1: apply a permission and content filter

Use only material you are allowed to share: your own notes, permitted worksheets, authorised textbook excerpts and class materials for which the school, instructor and rights holder permit this use. School rules may distinguish private revision from assessed work, or prohibit artificial intelligence tools for particular courses. ChatGPT cannot determine whether your proposed upload complies with those rules.

Procedure: look through every page before opening the camera. Remove names, email addresses, student numbers, health information, access codes and unrelated personal details. Exclude private classmates’ notes, live or restricted assessment questions, answer keys, instructor-only material, and copyrighted or proprietary content you do not have permission to upload. Check the relevant school or instructor artificial-intelligence policy. If the rule is ambiguous, ask the responsible teacher before uploading.

Example: a worksheet may be permitted for personal revision but show another pupil’s name and teacher feedback at the top. Do not upload it unchanged. Use your own authorised copy, or cover and recapture the identifying area only if doing so is allowed and does not remove information required for the exercise. An answer key attached to the back should be separated and excluded unless its use is explicitly permitted.

If you cannot establish both permission and relevance, leave the page out. Redaction is suitable for incidental personal data, but it does not create permission to upload restricted questions or someone else’s work. Never place passwords, authentication codes, confidential school records or other secrets in an image, filename or prompt.

Inspect both the paper and the camera preview. Pay particular attention to margins, sticky notes, desk surfaces and sheets visible behind the target page. If the material could affect assessment integrity or another person’s privacy, require review by the teacher, school or other authorised human decision-maker rather than relying on ChatGPT’s judgement.

Minute 1–2: choose privacy settings before the upload

For a personal ChatGPT account, review Settings > Data controls > Improve the model for everyone before creating the conversation. OpenAI’s Data Controls guidance, read on 5 October 2026, says that when this setting is off, new conversations are not used to train OpenAI models, although they can still appear in chat history. The change concerns new conversations; it is not retroactive deletion of existing chats or files.

Procedure: on a personal account, decide whether to turn off Improve the model for everyone, make that choice before starting this revision chat, and then create a new regular conversation. Do not interpret the switch as an instruction to upload sensitive or prohibited material. Workspace accounts can have different controls and retention policies, so follow the school or organisation’s displayed settings and policy rather than assuming that personal-account behaviour applies.

Example: a learner who turns the setting off and then starts a new regular chat has opted out of model improvement for that new conversation under OpenAI’s documented control. The chat can still remain in history, and its uploaded file can still require later deletion. If the learner submits response feedback, OpenAI’s separate data-use guidance says the associated conversation may be used for training even after opt-out; therefore, feedback is not appropriate if the learner’s aim is to minimise that use.

Treat training preference, storage and deletion as three separate questions. Turning off model improvement addresses the first for new conversations; it does not answer whether the chat remains stored or whether a file has been saved separately. If the material is unsuitable for storage or disclosure under school rules, do not upload it merely because the training control is off.

Confirm the setting shown in your own account and check whether you are using a personal or managed workspace. Do not infer another learner’s configuration from yours. Account, workspace and regional settings can differ.

Use a regular conversation, not Temporary Chat

Study mode is unavailable in Temporary Chats. That creates a genuine trade-off. OpenAI documents Temporary Chat as outside normal history and not used for model improvement while temporary, with possible retention for up to 30 days for safety; however, it cannot provide the Study-mode surface required for this tutorial. It is therefore incorrect to combine “Temporary Chat” and “Study mode” in the same proposed session.

Procedure: create a regular ChatGPT conversation, then use the plus menu to select Study; if Study is not immediately visible, use the menu search described in OpenAI’s Study-mode guidance. Only after entering the correct surface should you attach the verified pack or replacement images. Memory is not required for this one-off session, and there is no need to store sensitive course details for personalisation.

Example: if your priority is the guided, one-question-at-a-time workflow, choose a regular chat, review Data Controls first and plan deletion afterwards. If your overriding requirement is to avoid a normal history entry, this tutorial’s Study-mode route is not compatible with that requirement; use Temporary Chat only as a different, non-Study workflow and accept that it may still be retained for up to 30 days for safety.

Choose a regular chat only when the permitted material and applicable policy allow it. Do not sacrifice a privacy or school requirement merely to use Study mode. Study mode also supplies no extra message, file, rate or upload allowance and does not unlock unavailable models; follow the current limits displayed by your account.

Check that the conversation is regular and that Study is active before uploading. If the interface does not offer Study, do not assume that a differently labelled chat has the same behaviour. Recheck the account surface and current official guidance, or postpone the workflow.

Plan deletion rather than confusing opt-out with deletion

A regular chat and an uploaded file may have separate lifecycles. OpenAI’s retention guidance, read on 5 October 2026, says regular chats remain until deleted or removed under an applicable workspace policy. Deleting a chat removes it from the account view and schedules permanent deletion within 30 days, subject to OpenAI’s stated exceptions. It also says that deleting the chat does not delete a file that remains saved in Library; that Library copy must be removed separately.

Procedure: before capture, add a simple post-session task: delete the regular chat when it is no longer needed, then use Library on the web to check for and delete the uploaded scan separately. Do not make Library management the purpose of the study session; this is merely the necessary clean-up distinction. Deletion is scheduled rather than guaranteed as immediate backend erasure, and OpenAI documents exceptions.

Example: deleting “Quadratics revision” from chat history may remove the conversation from view while “quadratics-pack.pdf” remains in Library. The intended clean-up therefore has two checks: the chat is gone from history, and the file is gone from the main Library view after separate deletion.

If you intend to retain neither the interaction nor the source file, perform both deletion steps. If a school-managed workspace imposes a different retention policy, follow that policy and ask its administrator where necessary. Never promise a classmate or teacher that pressing one delete control immediately erases every copy.

After the study session, inspect chat history and Library separately. For any consequential privacy or records decision, obtain confirmation from the account or workspace administrator rather than treating the model’s explanation as evidence of deletion.

3. Capture a usable revision pack on iPhone (3 minutes)

The capture stage has one purpose: produce a finite, ordered set of pages that can survive a literal legibility check. It is not yet the point at which ChatGPT should summarise, explain or generate questions. Scan may not yet appear on every iPhone (see section 1); if it is absent, use the photo fallback below.

Minute 0–1: define and order the pack

Choose only the pages needed for this one session. Apply the permission and content filter from section 2 to every page. The practical trade-off is simple: a smaller, coherent pack is easier to inspect page by page, whereas an oversized pack consumes capture time and makes omissions harder to notice.

1. Set a narrow topic. For example, select “electromagnetic induction definitions and two worked examples”, rather than an entire physics folder.
2. Remove irrelevant sheets. Exclude blank pages, duplicate photocopies and unrelated homework unless their presence is necessary to understand a page.
3. Arrange the pages in the intended reading order. Put the first page on top and maintain that order throughout capture.
4. Add visible page identifiers if permitted. A small handwritten “1”, “2”, “3” in an empty margin can make later verification more reliable. Do not write over source text or alter assessment material.
5. Count the physical pages yourself. Keep that count for the gate in section 4; ChatGPT’s reported count must be compared with it.

Include a page only if it contributes evidence for the topic and you are permitted to upload it. If the meaning of page 4 depends on a formula defined on page 2, retain both. If a page merely covers a different topic, leave it out. Human review begins here: before opening the camera, look through the stack and confirm its physical count, order and permission status.

Minute 1–2: prepare each page for legibility

Good framing does not guarantee recognition, but poor framing creates avoidable ambiguity. OpenAI says unclear images can produce less accurate results, rotated text may be misinterpreted, and images are resized before analysis. Its image guidance says: “Big text: Enlarge text within the image to improve readability, but avoid cropping important details.” These are reasons to make the handwriting prominent while retaining all margins, not claims that a particular page will be read correctly. The current guidance is in the ChatGPT Image Inputs FAQ, accessed 5 October 2026.

• Use one physical page per frame. Two side-by-side pages reduce the space available to each and make page order less obvious.
• Keep the page upright. Position the top of the writing at the top of the captured image. Rotate the paper physically rather than expecting later interpretation to correct it.
• Retain every margin. Include question numbers, diagram labels, footnotes and formula subscripts near the edges. A neatly cropped photograph is unusable if it removes the left-hand question number or the end of an equation.
• Use even, sufficient light. Move away from glare, hard shadows and reflections. Ensure pale pencil contrasts with the page. This is preparation advice, not a promise about camera processing.
• Make small writing large enough to inspect. Move closer while keeping the complete page in frame. If a full page makes a dense derivation visibly tiny, plan a full-page photograph plus a separate permitted close-up during repair.
• Flatten folds and curved bindings. Do not let a hand, clip or book spine obscure symbols. Pay particular attention to minus signs, powers, vector arrows and diagram annotations.
• Check focus before moving on. Zoom into the captured page on the phone and inspect the smallest writing. If you cannot distinguish similar characters yourself, recapture it.

For example, suppose a chemistry worksheet contains the expression “2H₂ + O₂ → 2H₂O” beside a labelled energy-profile diagram. The frame must preserve subscripts, coefficients, the reaction arrow, both axes and every diagram label. A photograph showing the equation clearly but cutting off the vertical-axis label has not captured the complete source. Recapture rather than assuming that the missing label can be inferred.

If glare affects only one corner, do not accept the image merely because the main paragraph is readable. Reposition the light or page and capture it again. If pencil remains faint, place the page against a plain contrasting surface and improve the ambient lighting; do not digitally rewrite or “clean up” the content in a way that could change it. If a fold permanently obscures a word, preserve the full-page image and mark that location as potentially unreadable for the literal check.

Minute 2–3: use Scan, or attach source photographs

In a regular ChatGPT conversation on iPhone, open the ChatGPT camera, open its three-dot menu and select Scan if it is present. Capture the first page, inspect the preview, then continue through the stack in the established order. The documented Scan workflow combines consecutive captures into one PDF ready to upload. Before submitting it, compare the preview sequence with the physical stack: page 1 must precede page 2, no page should occur twice, and the last captured page should match the last physical page.

If Scan is not visible, do not spend the session searching menus or assume that another platform has an equivalent. Use + > Add photos & files and attach clear, upright page images instead. The documented still-image formats include JPEG, PNG and non-animated Graphics Interchange Format (GIF)A palette-based image file format that can store still images or short animations. A successful file upload does not itself prove that writing in an image was read accurately. Open glossary entry files. For this handwritten workflow, JPEG or PNG is the practical choice. Attach images in reading order and, where the interface permits, submit a manageable group whose order you can verify.

Choose Scan when it appears and the initial goal is to create an ordered PDF for the legibility check in section 4; choose Add photos & files immediately when Scan is absent. Neither route earns automatic trust. A PDF must pass the next section’s page-level inventory, and individual photographs must still be checked for literal readability.

For example, with six pages labelled 1–6, capture exactly those six in order. If the preview reads 1, 2, 4, 3, 5, 6, correct the sequence or start again rather than asking ChatGPT to reconstruct it. If page 5 is blurred, recapture page 5 before upload. The human verification is the physical-stack comparison: count six source pages, count six previews and inspect the smallest visible notation on every preview.

Uploads remain subject to the limits and settings shown by the reader’s account. OpenAI documents plan-dependent file, image and rate limits, and says some limits can change or be reduced; follow the current messages displayed in your account rather than designing the pack around a presumed allowance. Study mode does not add upload or message capacity. If a limit interrupts capture, reduce the pack to the pages essential for this session rather than merging several pages into one unreadably small image.

After checking the handwritten pages against their source images, a separate long-document evidence workflow using OpenAI’s application programming interface, retrieval provenance and claim verification is described in Build a Human-Reviewed Long-Document Evidence Packet; it is not a fix for missing PDF page visuals in a consumer upload.

4. Apply the required legibility gate: PDF or individual photographs? (4 minutes)

Do not request revision questions immediately after upload. First ask for a literal, page-by-page inventory and compare it with the originals. The operational sequence is mandatory for this tutorial: upload the scanned PDF → request a literal transcription and legibility check → replace every missed or uncertain page with a clear JPEG or PNG source photograph → repeat the check. This distinguishes evidence that ChatGPT demonstrably identified from content it might merely infer.

Why the PDF must be tested rather than trusted

A Scan-created PDF is a convenient container, but the way ChatGPT retrieves its contents differs by plan. OpenAI’s File Uploads frequently asked questions (FAQ)A collection of recurring questions and concise answers about a subject. Open glossary entry states: “All other plans and document files only support text-based retrieval. This means that ChatGPT will extract digital text from the file and discard any images.” Its separate visual-retrieval documentation says that reading text and visual elements embedded in PDFs is available only to ChatGPT Enterprise customers and is not supported for Free, Pro, Team or Edu accounts. These statements were current when accessed on 5 October 2026; consult the File Uploads FAQ and Visual Retrieval with PDFs FAQ.

This limitation matters particularly for handwriting. A photographed page placed inside a PDF may contain little or no digital text for text-based retrieval to extract. The visible words, diagrams and page layout can exist as image content rather than as a digital text layer. Consequently, successful upload is not evidence of successful reading. Enterprise visual retrieval changes the documented PDF capability, but it still does not guarantee perfect recognition of handwriting, notation or diagrams; Enterprise users must run the same literal check.

Meaningful trade-off: the PDF preserves a convenient multi-page sequence, while individual page photographs use the still-image input route and make each source page directly available as an image. The photograph fallback costs additional attachments and may encounter plan limits, but it is materially different from repeatedly uploading the same image-based PDF on a non-Enterprise plan. If literal evidence is missing, convenience loses to verifiability.

Run the first inventory without inviting inference

Attach the scanned PDF and send a request that forbids reconstruction. Use the following as an example instruction, not as a guarantee of transcription quality:

Inspect this revision pack before teaching from it. Produce a literal page-by-page inventory only. For each page, report: (1) the page number or “no visible number”; (2) every visible heading; (3) every visible question number and question; (4) every formula exactly as you read it; (5) every diagram title and visible label; and (6) each word, line, symbol or region that is unclear or not readable. Do not correct spelling, complete sentences, infer cropped content or reconstruct a likely formula. If you cannot access a page’s visual content, say so. Finish by stating the total number of pages you could inspect.

This prompt asks for an inventory, not a polished transcription. That distinction is useful. A polished summary can conceal omissions by producing plausible subject knowledge; a literal inventory creates items that can be checked against visible marks. “This page discusses photosynthesis” is insufficient. “Page 2, heading ‘Limiting factors’; questions 3(a), 3(b) and 4; graph labels ‘rate’, ‘light intensity’ and ‘X’” is checkable.

Compare the response with the physical originals in order. Use a finger or ruler to move down each page and verify four categories:

1. Identity: Does the reported page number and heading match the source?
2. Coverage: Are all questions, formulae and diagram labels represented?
3. Literal accuracy: Are signs, subscripts, powers, units and qualifiers preserved?
4. Uncertainty: Has ambiguous content been marked “unclear/not readable”, rather than silently completed?

The PDF route passes only if every page needed for practice is present and its relevant heading, questions, formulae and diagram labels survive the comparison. A single consequential omission—such as a missing negative sign, an unread question condition or an absent axis label—fails that page. Do not average accuracy across the pack. Pages can be routed separately: keep verified pages as they are and replace only failed pages with source photographs.

Worked gate example: distinguish a pass, a repair and a stop

Assume the physical pack has four pages. Page 1 is headed “Newton’s laws”; page 2 contains questions 1–5; page 3 has the formula F = ma and a force diagram labelled “normal reaction”, “weight” and “friction”; page 4 contains a worked calculation ending in “3.2 m/s²”. The following are example outcomes and decisions, not reported product test results.

• Pass: the inventory reports four pages, reproduces the heading, lists questions 1–5, identifies F = ma, names all three diagram labels and transcribes the final value with its unit. The student checks each item against the originals and finds no discrepancy relevant to practice. The PDF may remain the source for those pages.
• Repair: the inventory reports four pages but calls the diagram’s “normal reaction” label “unclear” and reads the final value as “3.2 m/s”. Attach clear photographs of pages 3 and 4 and repeat the literal check for those pages. Do not ask a mechanics question that depends on either disputed item yet.
• Stop: the response gives a confident explanation of Newton’s laws but cannot list the pages or visible items. That is not a pass, even if the explanation is academically correct. Treat the source as unread and move to individual photographs.

A correct answer from ChatGPT is not proof that the notes were transcribed accurately. A model may know the subject independently, infer a conventional formula or supply a plausible label. The human test concerns correspondence with this specific page. If the original says “resultant force” while the response says “force”, inspect whether that difference affects the intended definition. For a consequential educational judgement, disputed marking decision or interpretation of restricted material, stop and ask the instructor or another authorised human reviewer rather than relying on the model.

Repair failed pages through the individual-photo route

For each failed page, place the original flat, upright and in even light. Capture one complete page as a JPEG or PNG with all margins visible. If tiny writing remains difficult to inspect, add a close-up of the affected region after the complete-page photograph; the full page supplies context, while the close-up enlarges the text. Do not submit only a crop that removes the question number, heading or diagram context.

Attach the repair images through + > Add photos & files, then use an example repair instruction such as:

These are replacement source photographs for pages 3 and 4. Inspect each image directly. Transcribe page 3’s formula and every force-diagram label exactly; then transcribe page 4’s final calculation line, including sign and unit. Mark any uncertain character as “[unclear]”. Do not use general subject knowledge to fill a gap.

Check the returned text character by character where notation matters. For an equation, verify coefficients, brackets, fraction bars, powers, subscripts, vector marks, signs and units. For a diagram, verify each label and which line, arrow or region it belongs to. For prose, verify negations and limiting words such as “not”, “only”, “always” and “except”. These small elements can reverse meaning even when the surrounding sentence appears correct.

If the individual photograph is still reported as unclear, do not keep asking for guesses. Recapture it with larger readable writing in frame, improved light and corrected orientation. If the original itself is faint or damaged, type only the specific line yourself after checking it, label it explicitly as your manual transcription, and retain the visible original for comparison. If you cannot establish the source wording, remove that item from the session and record it for review with a teacher or authorised peer.

Freeze a verified source map before practice

Once all required pages have passed, ask for a compact source map rather than another summary. This is an example:

Create a source map from verified material only. For each usable page, give its page number, exact heading and the question numbers, formulae or diagram names confirmed there. For anything absent or still uncertain, write “not found/unclear”. Do not add outside facts.

Save the response only after comparing it with the originals. Correct any mismatch in the chat with a narrowly scoped statement, for example: “Correction after checking the original: page 4 ends in 3.2 m/s², not 3.2 m/s. Repeat the page 4 source entry and mark it as user-verified.” This records the correction but does not eliminate the need to consult the original later. The source page remains authoritative for this exercise.

Final gate rule: proceed to one-at-a-time Study-mode practice only when the source map names every page that will support questions and each disputed detail has either been verified, replaced or excluded. Require later answers to cite an exact page and heading and to say “not found/unclear” when the pack does not support a claim. If the response cites no page, cites the wrong heading or introduces outside material, pause and verify before answering the next question.

Continue in the regular chat, and follow the deletion plan described in section 2.

For a separate overview of Privacy Center topics and its signed-in rollout scope, review ChatGPT Privacy Center Guide: Memory, Personalized Ads, Temporary Chat, whose title includes Memory, Temporary Chat, Connected Apps, Model Improvement, Data Export and Account Security; check the account interface and official policy for actual setting behaviour and deletion rules.

5. Start Study mode and lock the source boundary (2 minutes)

Begin only after the source map from the legibility gate identifies which pages are readable, which have been replaced by individual photographs, and which remain excluded. This is a different decision from merely confirming that an upload exists: the chat may accept a file without having dependable access to its handwritten contents. Keep the physical pages or their original photographs open for later verification.

Minute 0–1: open Study in a regular conversation

1. On the iPhone, start or continue the regular conversation containing the confirmed pages.
2. Tap +, then choose Study. If it is not immediately visible, use the menu search described in OpenAI’s Study-mode guidance.
3. Check that the confirmed source attachments and source map are present in this conversation before proceeding.
4. If the scanned PDF file was unreadable but replacement JPEG or PNG images passed the legibility gate, tell Study mode to use those images rather than the failed PDF pages.

As documented by OpenAI and accessed on 5 October 2026, Study mode can be asked to quiz the learner one question at a time, wait for an answer and explain what to review next. It can still make mistakes or occasionally give a direct answer, so selecting Study changes the interaction style rather than proving that the source was read accurately. It also does not add messages, uploads or capacity beyond the limits shown by the learner’s account. Follow the current account display if a file, message or rate limit interrupts the session. See OpenAI’s Study-mode instructions.

Open Study in the regular conversation containing the verified pages, or reattach those pages to a new regular conversation. Do not move to a Temporary Chat; Study mode is unavailable there (see section 2).

If Scan is absent (see section 1), this does not prevent the session. Use + > Add photos & files and attach the already verified individual page images. Do not infer an Android Scan route from this fallback. The exercise depends on confirmed source images, not on use of the Scan command itself.

Minute 1–2: give one source-bound instruction

Send one compact instruction that controls evidence, turn-taking and failure handling. Replace the bracketed fields, but do not add private information or secrets. The following is a suggested example, not a guarantee that the model will follow every condition perfectly:

I am a [level] student revising [topic] for [permitted assessment type/date]. Use only the supplied pages that passed the legibility check: [list the confirmed PDF pages and replacement photo labels]. Exclude [list failed or unreadable pages].

Before asking questions, show a source map containing each usable page number or photo label, its exact heading, and any remaining unreadable word, line, formula or diagram label. Do not infer missing handwriting.

Then identify the three highest-priority concepts supported by those pages. Run the session one question at a time and wait for my reply. Use a mixture of recall and application. If I am wrong or uncertain, offer a hint before showing an answer. For every correction, cite the exact page or photo label and heading. If the support is absent, say “not in the supplied pages”. If the wording is unreadable, say “unclear in the supplied pages”. Do not fill either gap with outside knowledge.

Near the end of the session, give me two uncertain claims to check against the visible original pages, without resolving them for me. Stop if a required answer cannot be supported by a confirmed page. Finish with three items to reread, each with an exact page and heading reference.

This instruction makes three distinctions that are easy to blur. “Not in the supplied pages” means that no confirmed source contains the claim. “Unclear in the supplied pages” means that a potentially relevant mark exists but cannot be read reliably. “Incorrect according to page 4, Enzyme activity” means that legible source evidence directly conflicts with the learner’s answer. Only the third condition supports a source-based correction; the first two require stopping or repairing the evidence.

Worked example: suppose the source map includes “Photo 2 — Rates of reaction — collision frequency”, while a handwritten activation-energy formula on the same page remains unclear. A question about collision frequency can proceed if its supporting lines are legible. A question requiring the uncertain formula cannot proceed merely because ChatGPT knows a standard formula from elsewhere. The required response is “unclear in the supplied pages”, followed by a request for a clearer image or a learner-led check of the original page.

Compare the returned source map with the physical pack before answering the first question. Verify the number or photo label, heading, formula symbols and any qualifiers such as “increases”, “decreases”, “except” or “only”. A plausible summary or correct subject answer is not evidence that the handwriting was transcribed faithfully. If any map entry differs, send a repair instruction such as: Pause. Transcribe Photo 2 literally, preserving symbols and marking every unreadable character as [unclear]. Do not begin the quiz.

The strict boundary matters particularly for a Scan-created handwritten PDF; as section 4 explains, non-Enterprise plans may not read its handwriting, so an individual photograph that passes the legibility test is a material replacement for a failed PDF page.

Failure rule: if Study mode starts questioning before confirming the map, interrupt it. If it cites a page that was excluded, asks from general knowledge, or supplies an unsupported correction, do not count that turn as practice. Reply: Stop. That claim is not yet tied to a confirmed page and heading. Mark it “not in the supplied pages” or identify the exact visible support. Continue only after personally checking the cited location.

6. Run the 30-minute guided session

Sections 2 to 5 are separately timed preparation steps and do not count towards the 30-minute session. Start a fresh 30-minute timer here at minute 0; it runs through this section (minutes 0–27) and the close in section 7 (minutes 27–30). Its purpose is diagnostic: retrieve one core idea, answer a small mixed set, retry weaknesses and verify uncertain claims. It is not a substitute for teaching or a full test, and six questions cannot cover the whole pack. Use your own timer for this, because the chat should not be relied upon to measure elapsed time precisely.

Minutes 0–3: confirm the source map and choose priorities

Ask Study mode to return the promised map before it teaches or quizzes. A usable map has four fields: source identifier, exact heading, readable scope and unresolved content. For example, an acceptable entry format would be:

• Page 1 — “Cell membrane structure”: paragraph and labelled bilayer diagram readable; one small side note unclear.
• Photo 2 — “Diffusion and concentration gradient”: definition and worked example readable; bottom margin absent, so no claim should rely on it.
• Page 3 — excluded: formula symbols not confirmed; do not question from this page.

These are sample formats rather than claims about what ChatGPT will detect. Compare each entry with the original. A missing margin, invented heading or silently normalised symbol fails the check even if the surrounding explanation sounds sensible.

Next, request three priorities supported by the usable pages. Rank them by their value for this short session: a core definition, a relationship or mechanism, and an application that the pages demonstrate. Do not let the model rank material by assumed examination frequency unless the supplied pages themselves say so. A suitable instruction is: Rank three concepts only by their centrality and connections within these confirmed pages. Give one sentence of source-based justification and an exact page-heading reference for each.

Accept a priority only if its justification points to visible content. Replace any vague choice such as “important for exams” with the next source-supported concept. If fewer than three concepts are legible, use one or two rather than manufacturing a complete list.

Open the cited page for the first-ranked concept. Confirm that its heading and supporting lines are visible. Keep this page face down or move the image off-screen for the recall attempt, but leave it available for checking afterwards.

Minutes 3–8: explain one core concept from memory first

Choose the first priority and explain it without rereading the page. This is recall-first practice: the learner produces an account before seeing correction. It differs from asking ChatGPT to summarise the material, which tests recognition of a generated explanation rather than retrieval from memory.

Use a narrow prompt such as:

Ask me to explain the first-priority concept from memory. Wait until I finish. Then separate your response into: supported points, omitted points, and claims that conflict with the confirmed page. Cite the exact page and heading for every correction. Give no outside additions.

A concise learner response is enough. For example: Diffusion is the net movement of particles from an area of higher concentration to lower concentration because particles move randomly. Study mode should compare that wording only with the confirmed source. If the notes use a more qualified definition, the correction must point to the relevant line rather than relying on a standard textbook formulation.

Meaningful trade-off: a broad explanation reveals conceptual gaps but makes source checking slower; a one-sentence definition is easier to verify but may conceal weak understanding of mechanism or application. For a five-minute slot, begin with a definition and add one “why” sentence. Do not spend the entire session polishing prose.

If Study mode introduces a technically correct detail absent from the pack, label it outside scope rather than learning it as if sourced. Reply: That may be generally relevant, but is it present in a confirmed page? If not, mark it “not in the supplied pages” and remove it from this session.

Uncover the page and inspect every correction. Mark your own explanation as secure, uncertain or missed. Do not let ChatGPT assign the final status without this comparison. A correction unsupported by the visible original remains unresolved.

Minutes 8–18: answer six mixed questions, one at a time

Ask for six questions across recall and application, but insist that only one appears in each turn. “Mixed” should mean different cognitive operations within the supplied material, not the introduction of unrelated topics. A practical sequence is two direct recall questions, two relationship or explanation questions, and two short applications using situations already represented by the pages.

Send this control:

Prepare six source-supported questions, but show only Question 1. Wait for my answer. If I am wrong or say I am uncertain, give one hint linked to the page before revealing the answer. After feedback, cite the exact source identifier and heading, then ask whether I am ready for the next question.

For example, a recall question might request the handwritten definition under “Diffusion and concentration gradient”. An application question might vary quantities in a worked example already visible on that page. It must not introduce a new scenario whose solution depends on facts absent from the notes.

For each turn, use the same learner procedure:

1. Answer without opening the source.
2. Add confident, uncertain or guess after the answer.
3. If incorrect or uncertain, request one hint rather than the complete answer.
4. Attempt the answer again after the hint.
5. Read the correction and its page-heading citation.
6. Open the original only if the citation, transcription or judgement is disputed.
7. Record the item as secure, missed or unresolved.

The confidence label distinguishes a lucky answer from dependable recall. A correct guess belongs in the uncertain set because correctness alone does not show that the learner can reproduce the reasoning. Conversely, wording that differs from ChatGPT’s preferred phrasing may still be correct if the original page supports it.

Pause rule: stop the question sequence immediately if the answer key cannot be located on a confirmed page, if a formula symbol is unclear, or if the model cites a failed source. Do not accept general knowledge as a patch. Repair the relevant image and repeat the literal check, or discard that question and reduce the set. The time limit favours five verifiable questions over six unsupported ones.

Spot-check at least the first correction and every disputed correction against the original. Check the complete sentence, not only a matching keyword. Words such as “net”, “directly”, “approximately” and “under these conditions” can alter whether an answer is supported.

Minutes 18–24: reattempt only missed and uncertain items

Do not request six fresh questions. A fresh set increases coverage, whereas this stage tests whether feedback repaired the gaps already exposed. Ask Study mode to restate only items marked missed, uncertain or guess, changing the wording without changing the source requirement.

A suitable instruction is: Reask only the missed or uncertain items. Change the surface wording, but keep the same source-supported concept. Ask one at a time, give no hint unless I request it, and cite the page and heading after my answer.

For example, an original question might ask for the definition of diffusion. Its reattempt could ask the learner to identify what makes a movement “net” rather than merely asking for the same sentence again. This checks the same recorded concept from another angle. It should not expand into osmosis unless osmosis is both in the confirmed pages and part of the missed concept.

Promote an item to secure only when the second response is supported by the original and does not depend on seeing the previous answer. Leave it uncertain if the learner needed another substantive hint, if the page citation is disputed, or if the replacement wording added an unsupported assumption.

If no missed items remain, use the time to explain the least confident correct answer once more from memory. If several remain, prioritise the core concept and one application rather than rushing through all of them. This sacrifices breadth for an evidence-based check of whether feedback changed performance.

Compare the revised answer with the cited source yourself. Record the final status; do not ask the model to infer mastery from fluency or confidence alone.

Minutes 24–26: verify two uncertain claims against the original pages

Ask for two claims whose support was uncertain during the session. The model should provide the claim, the exact source identifier and heading, and a short description of where to look. It should not reveal a rewritten “correct” version before you inspect the page.

Use: Give me two uncertain claims to verify manually. For each, state the claim as used in this session, the exact page or photo label and heading, and the line, formula or diagram area I should inspect. Do not resolve it for me.

Open the visible original, not merely ChatGPT’s transcription. Classify each claim as:

• supported: the original clearly states or demonstrates it;
• contradicted: the original says something materially different;
• not found: no supporting content appears on the cited page;
• unclear: the handwriting, crop, symbol or diagram label cannot be read confidently.

If supported, quote only enough of your own permitted notes to identify the evidence and continue. If contradicted, provide the literal wording from the page and ask ChatGPT to correct the session record. If not found, remove the claim. If unclear, take a better upright photograph with one complete page, adequate light and contrast, readable writing and all margins present; then repeat the literal page-heading-formula check before accepting it.

Stop rule: when neither the original nor a clearer permitted image can support the claim, end that line of questioning. “Not in the supplied pages” is a successful boundary check, not an invitation to improvise. For consequential academic decisions—including marking, assessment submission or deciding whether disputed material is authoritative—seek human review from the relevant teacher or instructor.

Minutes 26–27: prepare the final reread list

After checking the two uncertain claims, ask for up to three reread items before minute 27. These should not be a new lesson or a flashcard system. Each item must contain the concept, exact source identifier, heading and reason for inclusion based on this session: missed, uncertain, contradicted or unclear.

A sample output structure is: 1. Meaning of “net movement” — Photo 2, “Diffusion and concentration gradient” — uncertain on first attempt; 2. Variable relationship — Page 4, “Rate graph” — application missed; 3. Formula symbol — replacement Photo 3, “Worked calculation” — verify symbol before reuse. This is an example format, not a product guarantee.

Reject generic entries such as “revise the whole topic” because they cannot be checked efficiently. If only one item caused difficulty, keep one item rather than padding the list. Save no sensitive course detail to a prompt or memory merely to preserve the result; copy the three references into your permitted personal notes if needed. Carry these items into the “I need to reread” list in section 7 rather than keeping a separate list.

Complete the two-step deletion planned in section 2 if you intend to remove both the chat and a separately saved Library copy.

Before deleting anything after the session, check whether captured pages remain as reusable Library files; then see ChatGPT Library File Lifecycle News for details on file reuse, separate chat and file deletion, and Temporary Chat limits.

Final human check: before treating the session record as usable, confirm that every retained correction has an exact pointer to a visible original page. Delete unsupported claims from your notes, flag unclear ones for a teacher, and follow your school or instructor’s artificial-intelligence rules.

7. Close cleanly in three minutes

The closing step records evidence from this session; it does not declare the topic mastered. Keep the original pages beside you, because a plausible answer from ChatGPT does not prove that the handwriting, formula or diagram was read correctly. Any consequential judgement about readiness for an assessment should be reviewed by you and, where appropriate, a teacher or tutor.

Minute 27–28: classify what happened without inflating confidence

Ask for three short lists: I know, I need to reread and source unclear. Every entry must identify the exact page and heading or section. The distinction matters: “I need to reread” means the source was legible but your answer was missing, incorrect or dependent on a hint; “source unclear” means the evidence itself was not verified, so it cannot fairly be treated as a knowledge gap.

Use this final instruction as an example rather than a guarantee of the product’s output:

Close this session with three lists: “I know”, “I need to reread” and “source unclear”. Include only claims tested in this conversation. For every item, give the exact source page and heading. Put an item under “I know” only if I answered it without a revealing hint and the answer matches a page we verified as legible. Put an item under “I need to reread” if the page was legible but my answer was wrong, incomplete or hint-dependent. Put anything with missing handwriting, an uncertain formula, a disputed diagram or no exact source pointer under “source unclear”. Do not claim mastery or add material from outside my pages.

If ChatGPT cannot supply a page-and-heading pointer, move the item to “source unclear”, even when the explanation sounds correct. If the pointer exists, open the visible original and confirm that the cited words, symbols or diagram labels are actually there. A correct general-knowledge answer may have come from the model rather than the revision pack.

For example, an appropriately scoped closing record might say:

• I know: “Define osmosis — page 2, ‘Transport across membranes’.” Keep this classification only after confirming that page 2 contains the definition used in the feedback.
• I need to reread: “Explain why surface area affects diffusion — page 3, ‘Factors affecting rate’.” Use this when the source was readable but the answer needed a substantial hint.
• Source unclear: “Final symbol in the rate formula — page 4, formula beneath ‘Worked example’.” Do not guess the symbol or convert the item into a practice question until the original is checked or replaced with a clearer photograph.

This sample describes the required format, not an actual assessment result. Avoid percentages, grades or readiness labels unless a teacher has supplied a valid marking method and a human checks its use. A six-question mobile session is too narrow to establish comprehensive coverage of a topic.

Minute 28–29: resolve only quick source disputes

Choose at most one “source unclear” item that can be settled in under a minute. Enlarge the original page on the iPhone and compare the disputed line character by character. If the page image in the conversation is blurry, rotated, shadowed or missing a margin, attach a new upright JPEG or PNG photograph containing one complete page, with sufficient light and contrast and writing large enough to inspect.

A suitable repair request is: Transcribe only the formula under “Worked example” on source page 4. Preserve every symbol literally, mark any unreadable character as [unclear], and do not complete it from subject knowledge. Then compare the response with the physical page yourself. If either version remains ambiguous, leave it under “source unclear” and ask a teacher or consult an authorised clean copy later.

Stop rule: do not spend the closing period repeatedly uploading the same poor image. One improved capture and one literal check are enough for this short workflow. Continuing to prompt around missing evidence encourages inference and consumes upload or message allowance without repairing the source. It is better to preserve an honest unresolved item than manufacture a confident answer.

Minute 29–30: save only what you intend to retain

Copy the three-list record into your own permitted notes if you want a local action list. Do not copy a long model-generated explanation merely because it is available; retain the page pointer and the precise rereading task. For example, save “Reread page 3, second paragraph; explain the surface-area link without a hint”, rather than an unverified replacement set of notes.

If you intend to remove the session, complete the chat and any separately saved Library-file deletion planned in section 2. A model-improvement preference is not a deletion control; follow the account or workspace policy when ending this session.

Schedule another session only if you want one. A useful diary entry is specific and conditional, such as “Thursday: reread pages 3–4, recapture page 4 if the formula is still unclear, then test three questions”. Do not ask ChatGPT to infer a revision timetable from sensitive personal information, and do not place credentials, private classmates’ work or other secrets in a prompt. If no return is needed, close the chat without creating an artificial study obligation.

Once you have checked each handwritten revision page against the image and its heading, a separate flashcard workflow becomes an option; ChatGPT Native Flashcards Prompting Playbook explains how to turn verified notes into interactive flashcard practice. Its prompts cannot substitute for checking the photographed page or for your own answer.

8. Troubleshooting and honest limits

Troubleshoot in this order: confirm the feature route, inspect the source capture, test what ChatGPT can literally read, and only then alter the study instruction. Interface availability, account settings and limits may differ. The facts below reflect OpenAI documentation accessed on 5 October 2026; follow what your own account currently displays.

The Scan option is missing

As noted in section 1, Scan is still rolling out on iOS; the documentation does not specify a minimum app version, completion date or Android equivalent. Do not delay the session searching through undocumented menus.

1. In a regular iPhone conversation, check the camera route and its three-dot menu once for Scan.
2. If Scan is absent, use + > Add photos & files and attach clear individual page photographs.
3. Keep one upright page per image, include every margin, use adequate light and contrast, and make the writing large enough to inspect.
4. Ask for a literal page number, heading and formula inventory before requesting any questions.

The fallback changes the evidence route, not the learning objective. Individual photographs can be preferable for handwriting because they preserve each page as a still-image input. OpenAI warns that images are resized, ambiguous or unclear inputs are less accurate, and rotated text may be misinterpreted, so image input still does not guarantee accurate transcription.

If the page-level inventory matches the original, continue. If it misses a heading, line, label or formula, recapture only the affected page and repeat the literal check. Stop if a permitted, readable source cannot be produced.

Study is not visible or cannot be selected

OpenAI documented Study mode across ChatGPT plans, models and supported web and mobile platforms as of 5 October 2026, but the normal account, workspace and interface conditions still apply. In a regular conversation, open +, look for Study, or search the menu if the interface offers that route. Do not assume Study adds upload capacity, messages, a different model or relief from rate limits.

Study remains unavailable in Temporary Chats, custom GPTs and Projects. Use a regular conversation for this guided session; section 2 explains the different history, data-use and deletion choices.

If Study remains unavailable after checking the correct regular-chat surface, do not pretend that an ordinary answer is the same feature. You may postpone the session or run a manually controlled exchange by asking for one question at a time and checking every response against the original. Treat that as ordinary conversational practice, not as Study mode. OpenAI also says Study can sometimes provide a direct answer, so its presence does not guarantee a consistently question-led exchange.

The handwriting is blurry, rotated or partly missing

Repair the capture before repairing the prompt. Prompt wording cannot restore strokes that the camera did not record. Place the page flat, orient the writing upright, remove shadows and glare, include all edges, and photograph one page at a time. Move close enough to make the handwriting large while retaining the margins and page identifier.

Then request a constrained inspection:

For this image only, report the visible page number and heading, transcribe the second paragraph literally, and reproduce the formula exactly. Mark uncertain characters as [unclear]. Do not infer words from the subject.

Compare that output with the original. Pay particular attention to minus signs, superscripts, subscripts, decimal points, arrows, units and similar letterforms. A fluent paraphrase is a failed legibility test because it can conceal an incorrect symbol. If the literal transcription is wrong twice despite a clearer capture, omit the page from the session and seek human review.

The page count or order is wrong

Distinguish a capture problem from a reading problem. First count the physical sheets and inspect the attachment thumbnails. If six pages were intended but only five thumbnails or scan pages exist, add the missing source before prompting. If all six are present but ChatGPT reports five, ask for a fresh inventory containing only the visible page sequence, first heading and final line on each page.

For example: I attached six source pages labelled 1–6. Make a six-row inventory. If you cannot locate a label, write “label not found”; do not renumber pages by topic. Compare every row with the originals. Handwritten numbering may be duplicated or absent, so use neutral identifiers such as “attachment 3, handwritten page unnumbered” rather than silently correcting it.

Practice begins only when the source map accounts for every intended page. A duplicated, omitted or reordered page can produce misleading coverage even if each individual answer sounds reasonable.

The Scan PDF is attached but the handwriting is not being read visually

This is a plan-dependent document-processing issue, not merely a weak prompt; see section 4 for why non-Enterprise plans may not read handwriting inside a PDF.

Apply the mandatory route test: ask for a page-by-page literal transcription and legibility report; compare it with the original; then attach every missed or uncertain page as a clear JPEG or PNG and repeat the check. Do not rely on Projects, on knowledge files attached to custom GPTs (user-built ChatGPT assistants) or on repeated PDF uploads to work around failed visual reading. The individual-photo fallback is material because it changes the content from an embedded PDF visual into a supported still-image input.

For example, if the PDF response identifies a topic but cannot reproduce the handwritten formula on page 3, do not accept topic recognition as a pass. Attach the original page 3 photograph and ask for the formula literally. Continue only after human comparison confirms the characters. Enterprise visual retrieval also should not be treated as perfect handwriting recognition; uncertain material still requires inspection.

An upload, message or rate limit interrupts the session

Study mode supplies no additional allowance. OpenAI’s published limits, accessed on 5 October 2026, included a general ceiling of 512 megabytes per file, 20 megabytes per image, and three file uploads per day for Free users; it also said limits can be reduced during peak periods. Limits and counters are subject to change, so the message shown by the account should govern the immediate decision.

Do not evade a limit by combining unreadable pages into a dense collage or by placing secrets in an external prompt. Instead, stop and preserve the compact closing list locally. On a later permitted session, upload only the pages marked “I need to reread” or “source unclear”. The trade-off is narrower coverage in exchange for legible evidence and compliance with the displayed allowance.

If a large file fails, verify its type and size, but do not assume size is the only cause. Account settings, plan allowances, workspace controls and temporary rate conditions may also matter. A human should decide whether the remaining material is important enough to defer rather than continuing from incomplete sources.

Study gives a direct answer instead of waiting

Study mode can guide learning with questions, but OpenAI says it may make mistakes and can sometimes provide a direct answer. Reassert the interaction boundary once: Do not reveal the answer yet. Ask one question, wait for my attempt, then offer one non-revealing hint. After feedback, cite the exact page and heading.

If it reveals the answer anyway, mark that item “exposed” rather than “known” and switch to a different concept. Later, test the exposed concept using a newly worded question without showing the answer. This avoids treating recognition immediately after disclosure as independent recall.

If feedback cites no precise source, ask: Which exact supplied page and heading supports that correction? If none does, say “not in the supplied pages”. Check the visible original yourself. For graded, medical, safety-related or otherwise consequential conclusions, obtain qualified human review rather than relying on the model’s judgement.

Memory or personalisation introduces material outside the pack

OpenAI says Memory features and controls vary by plan, region, platform and workspace settings. When available and enabled, Study may use saved memories or past chats for personalisation, but Study works with Memory off. Memory is therefore optional, not a prerequisite for this 30-minute source-bounded session.

If the response mentions a previous course, preference or fact not present in the pack, restate the boundary: Ignore prior-chat context for factual content. Use only the verified supplied pages, and label anything else “not in the supplied pages”. Review available controls under Settings > Personalization > Memory if you do not want memory used. Do not store sensitive course details merely to improve personalisation.

Deleting a chat does not necessarily remove a separately saved memory. Review and remove any relevant saved memory separately where the controls are available. Conversely, turning Memory off does not delete the uploaded chat or Library file. These controls solve different problems: Memory affects personalisation; Data Controls affect use of new conversations for model improvement; chat and Library deletion govern retained items.

Permission or source status becomes uncertain after upload

Stop using the material if you discover that it includes private classmates’ notes, live or restricted assessment questions, answer keys, instructor-only content, or copyrighted or proprietary pages you were not permitted to upload. Follow school and instructor rules even if the interface accepts the file. Technical acceptance is not permission or academic-integrity approval.

Remove the chat and any separately saved Library copy when that is the intended response, then tell the relevant teacher or administrator if their rules require it. Do not paste the disputed material into another service or prompt while seeking help. For consequential decisions about permitted use, obtain human guidance from the institution rather than asking ChatGPT to approve its own use.

Know when to abandon the session

End the workflow when any of these conditions applies: the intended pages cannot be accounted for; a key page remains unreadable after one proper recapture; the PDF’s handwriting cannot be verified and individual photographs are unavailable; the source is not permitted; or the account limit prevents the necessary evidence from being attached. The honest output is then a “source unclear” record, not improvised practice.

The operational distinction is simple: a failed answer can become a rereading task, but a failed source cannot support a fair question. Preserve that boundary even when ChatGPT produces a polished explanation. The objective is a short, traceable study session grounded in visible pages—not a claim that the model has reconstructed material it could not reliably inspect.

Useful Links

• OpenAI ChatGPT release notes, including the iOS Scan announcement
• OpenAI guide to using Study mode in ChatGPT
• OpenAI ChatGPT image inputs FAQ
• OpenAI ChatGPT file uploads FAQ
• OpenAI visual retrieval with PDFs FAQ
• OpenAI guide to Data Controls in ChatGPT
• OpenAI explanation of how content may be used to improve model performance
• OpenAI chat and file retention policies
• OpenAI Temporary Chat FAQ
• OpenAI guide to managing files with Library
• OpenAI ChatGPT Memory FAQ