# The AI Quality-Control Clipboard Six review prompts for checking sources, missing steps, beginner clarity, files, and focused repairs. **For:** People reviewing an AI-generated guide, document, image, or web page before sharing it. **You will make:** A prioritized defect list and a truthful record of what was checked. ## Why use it A polished answer can hide a missing input, an invented detail, or a beautiful button that goes nowhere. The clipboard is where enthusiasm meets the actual file. ## What you need - The actual candidate output. - Original source material and acceptance criteria. - Tools capable of inspecting the relevant format; otherwise record the limitation. ## First time? Start here 1. **Open and start:** Open an AI chat. Put the thing you want reviewed and its original notes in the same conversation: paste text into the message box, or use the attachment button to upload actual files. A filename alone does not supply the file. 2. **Next step:** Choose the review that matches your question. Start with “1. Audit the claims” for factual accuracy, or “3. Read it like a stranger” for clarity. Use the discovery prompt first if you do not know what evidence or checks to supply. Send one review at a time and read its findings before asking for repairs. 3. **Save your work:** Copy the findings into a document named review-notes next to your draft. Save a separate revised version after fixes. Mark a check “not run” when no supported tool or real trial was available; do not turn a proposed test into a passed test. **Done when:** You have specific problems linked to actual text or files, and each important problem is fixed or clearly marked unresolved. A content review alone does not establish that a file opens or a real person can finish the process. Choose the review that matches the risk. A source audit answers a different question from opening a PDF or trying a customer workflow. Use the final repair prompt only after identifying specific defects. Asking the same model to praise its work six times is not six independent tests. ## The steps ### 1. Collect the evidence Keep the candidate, source material, and intended result together. Mark which files the assistant can inspect. **Expected output:** A review packet. **Check:** The reviewer has more than a summary of the missing file. ### 2. Review facts and sequence Run the evidence audit and missing-steps prompts. Check the highest-impact claims yourself. **Expected output:** A source-linked defect list. **Check:** Unknown facts remain unknown instead of being filled in. ### 3. Review usability and the actual artifact Run the stranger review and the relevant artifact check. If the reviewer cannot render a file or use a browser, preserve that limitation. **Expected output:** Separate editorial and technical findings. **Check:** A content pass is not recorded as a visual pass. ### 4. Repair and recheck Apply the repair prompt to the named failures. Recheck only the affected areas and reasonable regressions. **Expected output:** A revised version and dated QA note. **Check:** Every critical failure has been fixed or remains clearly blocked. ## Help with the details For the prompt you need, copy its discovery prompt into your AI chat without editing it. Answer one question at a time. Review the filled prompt it gives you, then send that reviewed prompt to start the task. Examples are illustrative; use your own facts for your task. ### Prepare: 1. Audit the claims #### Discovery prompt ```text Help me get ready to use “1. Audit the claims”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Candidate and the actual sources. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. What exact draft or page should I audit? Paste or upload it. 2. What original notes, documents, or official references support its claims? 3. Which sources can I actually access here, and which are missing? 4. Which claims matter most to the reader’s ability to finish the task? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Audit [MATERIAL] against [SUPPLIED SOURCES]. Extract meaningful factual claims, exact tool instructions, settings, quotes, prices, and claimed outcomes. For each, identify its actual supporting source or mark it unsupported. Use SOURCE for directly observed original material; FIRST-PARTY DOCUMENTED for identifiable official documentation you actually consulted; RECOMMENDED ADDITION for an explicitly proposed step; and UNRESOLVED for missing evidence. Do not browse or cite a document you cannot access, and do not treat your general knowledge as inspected evidence. Return claim, source location, evidence label, confidence limitation, and proposed correction. Preserve exact original quotations. Highlight claims that could prevent a beginner from completing the task. Do not silently rewrite an unsupported claim into another invented fact. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) MATERIAL: this sample workshop guide copy: “Build your first workshop checklist in five minutes. Paste your notes into any AI app and it will save a print-ready PDF automatically. Include materials, rehearsal, and a recap.” SUPPLIED SOURCES: these inline author notes: “I want a checklist covering materials, rehearsal, and a recap. The intended output is editable text. I have not timed the task or tested PDF creation. No tool documentation is supplied.” ```text Audit this sample workshop guide copy: “Build your first workshop checklist in five minutes. Paste your notes into any AI app and it will save a print-ready PDF automatically. Include materials, rehearsal, and a recap.” against these inline author notes: “I want a checklist covering materials, rehearsal, and a recap. The intended output is editable text. I have not timed the task or tested PDF creation. No tool documentation is supplied.”. Extract meaningful factual claims, exact tool instructions, settings, quotes, prices, and claimed outcomes. For each, identify its actual supporting source or mark it unsupported. Use SOURCE for directly observed original material; FIRST-PARTY DOCUMENTED for identifiable official documentation you actually consulted; RECOMMENDED ADDITION for an explicitly proposed step; and UNRESOLVED for missing evidence. Do not browse or cite a document you cannot access, and do not treat your general knowledge as inspected evidence. Return claim, source location, evidence label, confidence limitation, and proposed correction. Preserve exact original quotations. Highlight claims that could prevent a beginner from completing the task. Do not silently rewrite an unsupported claim into another invented fact. ``` ### Prepare: 2. Find the missing steps #### Discovery prompt ```text Help me get ready to use “2. Find the missing steps”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Full instructions and expected result. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. What are the complete instructions as the reader currently sees them? 2. What final result should the reader produce? 3. What accounts, files, tools, or skills does the reader already have? 4. Which actions have actually been tried, and which button names or steps are still unknown? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Read [INSTRUCTIONS] as a first-time user trying to create [RESULT]. Map each step's inputs, action, decisions, output, and dependency on earlier steps. Find missing accounts or files, unexplained terms, actions that happen off-screen, outputs used before creation, and choices with no decision rule. Quote the relevant passage for every finding. Separate a required missing step from an optional improvement. Do not invent the omitted button name, price, setting, or prompt. Propose a beginner-friendly clarification only when supported; otherwise write the question that needs answering. Return the shortest corrected sequence plus an unresolved list. Explain how the reader knows each major stage is complete. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) Illustrative reader setup: a beginner can open an AI chat and a text document. The supplied source notes say choose a craft, list materials, rehearse, explain the first step, leave time for questions, and send a recap. No account-specific export or sharing controls have been inspected. INSTRUCTIONS: this sample draft: “1. Put your notes into AI. 2. Use the worksheet to check the draft. 3. Export it and share the link.” No worksheet, export format, or sharing destination is supplied RESULT: an editable preparation checklist for a first-time workshop host, saved privately before sharing ```text Illustrative reader setup: a beginner can open an AI chat and a text document. The supplied source notes say choose a craft, list materials, rehearse, explain the first step, leave time for questions, and send a recap. No account-specific export or sharing controls have been inspected. Read this sample draft: “1. Put your notes into AI. 2. Use the worksheet to check the draft. 3. Export it and share the link.” No worksheet, export format, or sharing destination is supplied as a first-time user trying to create an editable preparation checklist for a first-time workshop host, saved privately before sharing. Map each step's inputs, action, decisions, output, and dependency on earlier steps. Find missing accounts or files, unexplained terms, actions that happen off-screen, outputs used before creation, and choices with no decision rule. Quote the relevant passage for every finding. Separate a required missing step from an optional improvement. Do not invent the omitted button name, price, setting, or prompt. Propose a beginner-friendly clarification only when supported; otherwise write the question that needs answering. Return the shortest corrected sequence plus an unresolved list. Explain how the reader knows each major stage is complete. ``` ### Prepare: 3. Read it like a stranger #### Discovery prompt ```text Help me get ready to use “3. Read it like a stranger”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Actual page or complete copy and the intended reader. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. What complete page copy or actual page should I read? 2. Who is the intended reader and what do they already know? 3. What should they make or learn, and what is the main next action? 4. Can I inspect the layout, or should this review be limited to supplied text? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Review [PAGE OR COPY] as [READER], who has no previous context. Answer from the material itself: What is it? Who is it for? Why use it? What will I make or learn? What do I need? Where do I begin? Where are the prompts? How do I check success? What happens next? For each answer cite the relevant text or mark it missing. Identify jargon, vague benefits, buried prerequisites, unearned claims, and unclear links. Distinguish problems in the supplied copy from layout issues you could not inspect. Suggest precise replacement wording for the three most important gaps. Preserve the brand's practical voice. Do not invent testimonials, guarantees, prices, or tool capabilities to make the page sound more persuasive. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) Illustrative intended offer: a free guide for making one editable workshop-preparation checklist from the host’s own notes. Main action: begin with the copyable brief prompt. Only page copy is supplied; no layout or live links are available to inspect. PAGE OR COPY: this sample page copy: “Workshop magic. Get unstuck with our AI-powered workflow. Everything you need to make it happen. Start now.” READER: a first-time online workshop host who has never used an AI prompt template ```text Illustrative intended offer: a free guide for making one editable workshop-preparation checklist from the host’s own notes. Main action: begin with the copyable brief prompt. Only page copy is supplied; no layout or live links are available to inspect. Review this sample page copy: “Workshop magic. Get unstuck with our AI-powered workflow. Everything you need to make it happen. Start now.” as a first-time online workshop host who has never used an AI prompt template, who has no previous context. Answer from the material itself: What is it? Who is it for? Why use it? What will I make or learn? What do I need? Where do I begin? Where are the prompts? How do I check success? What happens next? For each answer cite the relevant text or mark it missing. Identify jargon, vague benefits, buried prerequisites, unearned claims, and unclear links. Distinguish problems in the supplied copy from layout issues you could not inspect. Suggest precise replacement wording for the three most important gaps. Preserve the brand's practical voice. Do not invent testimonials, guarantees, prices, or tool capabilities to make the page sound more persuasive. ``` ### Prepare: 4. Inspect the actual file #### Discovery prompt ```text Help me get ready to use “4. Inspect the actual file”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Actual files, required format, intended use. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. Which actual files need inspection? Upload them rather than only naming them. 2. What should be included, in what format, and how will people use it? 3. What pages, links, images, or dimensions matter for this use? 4. If file rendering is unavailable, should I provide a limited content review and list the checks you must do manually? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Inspect [ACTUAL FILES] for [INTENDED USE]. First state what your current tools can open, render, read, and test. Check only what those tools support: expected files present, readable text, page order, image references, links, dimensions, and obvious missing sections. For each check report PASS, FAIL, UNCERTAIN, or NOT RUN with evidence. If you render pages, identify which pages were viewed. If you cannot inspect visual layout, say so; extracting text is not a visual review. Do not infer print quality from a DPI label or image quality from file size. Return critical defects first, then smaller improvements. Separate opening the artifact from executing the process it teaches. Do not claim a successful workflow test unless it actually ran. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) ACTUAL FILES: the workshop-checklist.pdf file I will upload; first ask me for the actual upload if it is not accessible INTENDED USE: a privately downloaded, one-page checklist for a first-time workshop host. It should contain Before, During, and After sections covering craft choice, materials, rehearsal, explaining the first step, questions, and a recap. No print-quality or live-workshop test has been supplied ```text Inspect the workshop-checklist.pdf file I will upload; first ask me for the actual upload if it is not accessible for a privately downloaded, one-page checklist for a first-time workshop host. It should contain Before, During, and After sections covering craft choice, materials, rehearsal, explaining the first step, questions, and a recap. No print-quality or live-workshop test has been supplied. First state what your current tools can open, render, read, and test. Check only what those tools support: expected files present, readable text, page order, image references, links, dimensions, and obvious missing sections. For each check report PASS, FAIL, UNCERTAIN, or NOT RUN with evidence. If you render pages, identify which pages were viewed. If you cannot inspect visual layout, say so; extracting text is not a visual review. Do not infer print quality from a DPI label or image quality from file size. Return critical defects first, then smaller improvements. Separate opening the artifact from executing the process it teaches. Do not claim a successful workflow test unless it actually ran. ``` ### Prepare: 5. Design a small real-world check #### Discovery prompt ```text Help me get ready to use “5. Design a small real-world check”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Workflow, access limits, definition of done. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. What complete task should the person be able to finish? 2. What starting materials, accounts, and tool access do you have? 3. What actions are permitted, and what must remain a local or private trial? 4. What observable result would count as success, and how can we preserve evidence without private information? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Design the smallest meaningful test of [WORKFLOW] that can run within [AVAILABLE ACCESS AND AUTHORIZATION]. State the starting inputs, exact actions, expected intermediate observations, final acceptance check, and how to preserve evidence. Separate checks we can run now from steps requiring a human, external account, purchase, publication, or unavailable tool. Prefer a local or sandbox test when appropriate. Do not perform consequential external actions unless authorized. Do not label this plan a passed test. Return a short test sheet with empty actual-result fields. Include a failure path and what would remain unverified even after a successful small test. Avoid tests that only repeat the implementation without checking the user's goal. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) WORKFLOW: use six inline planning notes to create, copy, and reopen one editable workshop-preparation checklist AVAILABLE ACCESS AND AUTHORIZATION: an AI chat and my own document editor; private text drafting and saving only; no real messages, publication, purchases, or account changes. Starting notes: choose a craft, list materials, rehearse, explain the first step, leave question time, send a recap. Success means all six tasks appear in a saved document that I can reopen ```text Design the smallest meaningful test of use six inline planning notes to create, copy, and reopen one editable workshop-preparation checklist that can run within an AI chat and my own document editor; private text drafting and saving only; no real messages, publication, purchases, or account changes. Starting notes: choose a craft, list materials, rehearse, explain the first step, leave question time, send a recap. Success means all six tasks appear in a saved document that I can reopen. State the starting inputs, exact actions, expected intermediate observations, final acceptance check, and how to preserve evidence. Separate checks we can run now from steps requiring a human, external account, purchase, publication, or unavailable tool. Prefer a local or sandbox test when appropriate. Do not perform consequential external actions unless authorized. Do not label this plan a passed test. Return a short test sheet with empty actual-result fields. Include a failure path and what would remain unverified even after a successful small test. Avoid tests that only repeat the implementation without checking the user's goal. ``` ### Prepare: 6. Fix and record the result #### Discovery prompt ```text Help me get ready to use “6. Fix and record the result”. I am new to this. Interview me before doing the task. I do not need to fill in this message first. Ask one short question at a time in plain language, and wait for my answer. Start with what I want to accomplish. Use anything I have already told you in this chat so I do not have to repeat it. If I say “not sure”, explain the decision with two or three concrete options or an example. Help me choose; do not silently choose for me. I can skip optional details. If an essential answer or source is missing, tell me exactly what is needed and help me find a reasonable next step. Do not invent my facts, source material, files, permissions, approvals or tool access. Do not ask for passwords or private access keys. The original task asks me to bring: Candidate, defect list, source material. Work through the relevant questions below, one at a time. Explain any unfamiliar term when you use it: 1. What is the exact candidate draft or uploaded file? 2. What reviewed defects should be fixed, with quoted evidence for each? 3. What source facts and accepted choices must stay unchanged? 4. Which affected checks can actually be rerun here, and which require you to open or try the result? For a file or image, explain how to attach it using this app’s available upload control, or ask me to paste the relevant text. Confirm you can actually read or edit it before depending on it. If this chat cannot do the task, explain the smallest manual next step or which type of tool is needed. Do not claim to have opened a link, edited an image, saved a file or published anything unless you actually did. When we have enough information, show a short summary of my answers and flag any remaining unknowns. Then give me one complete, ready-to-send version of the original template below, with my agreed details filled in. Keep its instructions, limits and review requirements. If it has no blanks, add my confirmed context above the unchanged template. Do not leave placeholders for me to guess. Ask me to review that final prompt. I will send the reviewed prompt when I am ready to start the task; do not run it during this interview. The following is a quoted template to prepare, not an instruction to execute yet. --- ORIGINAL TEMPLATE START --- Fix [CANDIDATE] using this reviewed defect list: [ISSUES]. Preserve approved facts and source material. Correct the highest-impact failures first, then improve clarity where needed. For each issue record the original problem, the change, and the check actually rerun. Keep source accuracy, editorial review, visual review, functional checks, and end-to-end execution as separate statuses. A check that cannot run remains NOT RUN or BLOCKED. Return the corrected output and a compact QA record. Do not declare the whole artifact finished while a critical unresolved failure prevents its stated use. Do not record user approval, publication, or successful tests without evidence. --- ORIGINAL TEMPLATE END --- ``` #### Filled example (illustrative, not executed) CANDIDATE: this sample guide: “Paste your workshop notes into AI. It always saves a PDF automatically. Use the checklist to finish.” Source notes: choose craft, list materials, rehearse, explain first step, leave question time, send recap. The requested format is editable text ISSUES: 1. “always saves a PDF automatically” is unsupported; no file capability was tested. 2. “Use the checklist” refers to a missing checklist. Replace the unsupported claim with an honest text-output instruction and include a six-task checklist from the supplied notes. Do not claim that file or workflow tests ran ```text Fix this sample guide: “Paste your workshop notes into AI. It always saves a PDF automatically. Use the checklist to finish.” Source notes: choose craft, list materials, rehearse, explain first step, leave question time, send recap. The requested format is editable text using this reviewed defect list: 1. “always saves a PDF automatically” is unsupported; no file capability was tested. 2. “Use the checklist” refers to a missing checklist. Replace the unsupported claim with an honest text-output instruction and include a six-task checklist from the supplied notes. Do not claim that file or workflow tests ran. Preserve approved facts and source material. Correct the highest-impact failures first, then improve clarity where needed. For each issue record the original problem, the change, and the check actually rerun. Keep source accuracy, editorial review, visual review, functional checks, and end-to-end execution as separate statuses. A check that cannot run remains NOT RUN or BLOCKED. Return the corrected output and a compact QA record. Do not declare the whole artifact finished while a critical unresolved failure prevents its stated use. Do not record user approval, publication, or successful tests without evidence. ``` ## Prompt toolbox ### 1. Audit the claims Find unsupported factual claims. Inputs: Candidate and the actual sources. Example: [MATERIAL] = a tutorial draft; [SOURCES] = original slides and caption. Expected output: A claim-to-source table. ```text Audit [MATERIAL] against [SUPPLIED SOURCES]. Extract meaningful factual claims, exact tool instructions, settings, quotes, prices, and claimed outcomes. For each, identify its actual supporting source or mark it unsupported. Use SOURCE for directly observed original material; FIRST-PARTY DOCUMENTED for identifiable official documentation you actually consulted; RECOMMENDED ADDITION for an explicitly proposed step; and UNRESOLVED for missing evidence. Do not browse or cite a document you cannot access, and do not treat your general knowledge as inspected evidence. Return claim, source location, evidence label, confidence limitation, and proposed correction. Preserve exact original quotations. Highlight claims that could prevent a beginner from completing the task. Do not silently rewrite an unsupported claim into another invented fact. ``` ### 2. Find the missing steps Check the workflow’s dependencies. Inputs: Full instructions and expected result. Example: [RESULT] = a downloadable worksheet created from an attached draft. Expected output: A dependency map and concrete missing instructions. ```text Read [INSTRUCTIONS] as a first-time user trying to create [RESULT]. Map each step's inputs, action, decisions, output, and dependency on earlier steps. Find missing accounts or files, unexplained terms, actions that happen off-screen, outputs used before creation, and choices with no decision rule. Quote the relevant passage for every finding. Separate a required missing step from an optional improvement. Do not invent the omitted button name, price, setting, or prompt. Propose a beginner-friendly clarification only when supported; otherwise write the question that needs answering. Return the shortest corrected sequence plus an unresolved list. Explain how the reader knows each major stage is complete. ``` ### 3. Read it like a stranger Check whether a new reader understands the offer and next action. Inputs: Actual page or complete copy and the intended reader. Example: [READER] = a first-time digital-product creator using ChatGPT. Expected output: Specific clarity findings with replacement copy. ```text Review [PAGE OR COPY] as [READER], who has no previous context. Answer from the material itself: What is it? Who is it for? Why use it? What will I make or learn? What do I need? Where do I begin? Where are the prompts? How do I check success? What happens next? For each answer cite the relevant text or mark it missing. Identify jargon, vague benefits, buried prerequisites, unearned claims, and unclear links. Distinguish problems in the supplied copy from layout issues you could not inspect. Suggest precise replacement wording for the three most important gaps. Preserve the brand's practical voice. Do not invent testimonials, guarantees, prices, or tool capabilities to make the page sound more persuasive. ``` ### 4. Inspect the actual file Check a deliverable beyond its written content. Inputs: Actual files, required format, intended use. Example: [FILE] = checklist.pdf; [USE] = phone reading and home printing. Expected output: Checks with evidence and explicit not-run items. ```text Inspect [ACTUAL FILES] for [INTENDED USE]. First state what your current tools can open, render, read, and test. Check only what those tools support: expected files present, readable text, page order, image references, links, dimensions, and obvious missing sections. For each check report PASS, FAIL, UNCERTAIN, or NOT RUN with evidence. If you render pages, identify which pages were viewed. If you cannot inspect visual layout, say so; extracting text is not a visual review. Do not infer print quality from a DPI label or image quality from file size. Return critical defects first, then smaller improvements. Separate opening the artifact from executing the process it teaches. Do not claim a successful workflow test unless it actually ran. ``` ### 5. Design a small real-world check Plan an honest test before calling a workflow proven. Inputs: Workflow, access limits, definition of done. Example: [WORKFLOW] = creating and opening one worksheet download. Expected output: A minimal test plan with expected observations. ```text Design the smallest meaningful test of [WORKFLOW] that can run within [AVAILABLE ACCESS AND AUTHORIZATION]. State the starting inputs, exact actions, expected intermediate observations, final acceptance check, and how to preserve evidence. Separate checks we can run now from steps requiring a human, external account, purchase, publication, or unavailable tool. Prefer a local or sandbox test when appropriate. Do not perform consequential external actions unless authorized. Do not label this plan a passed test. Return a short test sheet with empty actual-result fields. Include a failure path and what would remain unverified even after a successful small test. Avoid tests that only repeat the implementation without checking the user's goal. ``` ### 6. Fix and record the result Make evidence-driven revisions. Inputs: Candidate, defect list, source material. Example: [ISSUES] = missing input file; unclear final check; wrong link. Expected output: Corrected output and a factual QA summary. ```text Fix [CANDIDATE] using this reviewed defect list: [ISSUES]. Preserve approved facts and source material. Correct the highest-impact failures first, then improve clarity where needed. For each issue record the original problem, the change, and the check actually rerun. Keep source accuracy, editorial review, visual review, functional checks, and end-to-end execution as separate statuses. A check that cannot run remains NOT RUN or BLOCKED. Return the corrected output and a compact QA record. Do not declare the whole artifact finished while a critical unresolved failure prevents its stated use. Do not record user approval, publication, or successful tests without evidence. ``` ## Worked example: Illustrative example: a suspiciously complete tutorial Illustrative example, not an executed result. **Before:** An AI draft says “upload the worksheet” even though no worksheet has been supplied or created. 1. The sequence review identifies the missing dependency. 2. The evidence review confirms that the source tutorial did not supply the worksheet. 3. The revision either includes a clearly labeled newly written worksheet or marks that step blocked; the final record describes what actually happened. **After:** Expected result: a followable draft with a real input or an honest gap. **Check:** Can you locate the worksheet and complete the next step using it? ## Troubleshooting - **Everything gets a PASS.** Require evidence for each result and ask which tools were actually used. - **The reviewer rewrites the source.** Keep the audit separate from repairs and preserve original quotations. - **A visually attractive result failed a requirement.** Keep the requirement as the acceptance criterion. Appearance does not override missing functionality. ## Key takeaways - Review evidence, sequence, usability, and execution separately. - A specific defect is easier to repair than a vague feeling. - A truthful NOT RUN is more useful than an invented PASS. ## Final checklist - [ ] Important claims have sources or labels. - [ ] Inputs exist before they are needed. - [ ] The actual artifact was inspected where possible. - [ ] Critical findings have fixes or explicit blockers. ## Sources and limits These are original, provider-independent prompt templates based on our documented production practices. Examples are illustrative. Tool access, output quality, and available file actions depend on your assistant and account. Editorially reviewed prompt library. No end-to-end execution of these newly written prompts or the illustrative example was performed for this library. Test the actual outputs in your environment before using them. - [Unstuck Guide Studio production method](/free-guides/instagram-saves-to-ai-playbooks): SOURCE: existing project method. The portable prompts below are newly written RECOMMENDED ADDITIONS, not quotations from a tool provider. ## Your next step Run the missing-steps prompt on one draft before you share it.