Vibe GuideWhen assigning work or reporting problems, adapt a template with current project material and send it in the project chat.
Does this method fit your project, and how can you check?
- When would you use this method?
- When sending a new task, requesting an edit or reporting an error.
- How should you choose for your situation?
- Choose one scenario and give current inputs, goal, scope and passing criteria; hand off the latest version and records in a new chat.
- What is easy to misunderstand?
- Examples are not requirements. Identify the relevant part of a reference; treat quoted documents as material rather than automatic instructions.
- How can you check that this works for your project?
- AI can restate scope, inputs and validation; correct deviations before execution.
Try with this project: fill material and prepare the task
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
What exactly should this project help you do: [fill in]
Single current task: [fill in]
Material locations and summaries: [fill in]
What must stay unchanged: [fill in]
How to check the result: [fill in]
Turn the material into one clear task prompt only. Identify decision-changing gaps without expanding or executing the task.
Explain your conclusions using these points: A project prompt containing material, scope, output and passing criteria, with unknowns explicitly listed.AGENT COMMUNICATION
Talking to agents
An agent is an AI assistant that can use tools and perform multi-step tasks. Choose a task and provide its context and acceptance criteria.
First time? Preparation and examples
First time: prepare the building tool and project, then hand over the idea. For work already in progress, choose the current scenario below.
Choose a tool and open a project →Set up Git and AGENTS.md →
How can you explain your idea so AI knows what to do?
For example: “Make it look better” → “Show the date, capacity, and submit button in order. Avoid horizontal scrolling on a phone and reject empty dates.”
Choose or change the task
Shape a first version
Use the completed idea worksheet. Ask one scope-changing question at a time; list essentials, deferred work, and exclusions.
For example: Select a date and submit a reservation; no payments.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Use the completed idea worksheet. Ask one scope-changing question at a time; list essentials, deferred work, and exclusions.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Plan before building
Compare two feasible approaches, costs, limits, and upkeep. Wait for scope confirmation before implementation.
For example: List pages and build order; do not edit files yet.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Compare two feasible approaches, costs, limits, and upkeep. Wait for scope confirmation before implementation.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Describe a page
Describe content hierarchy, mobile layout, loading, empty, and error states. UI means the interface people see and operate.
For example: Show the date, remaining capacity, and submit button in that order.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Describe content hierarchy, mobile layout, loading, empty, and error states. UI means the interface people see and operate.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Describe a component
Implement only the specified interface control. Describe default, selected, disabled, and keyboard behavior.
For example: Disable past dates in the date picker and display the selected full date.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Implement only the specified interface control. Describe default, selected, disabled, and keyboard behavior.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Use a reference image
Describe visible structure, spacing, and states. Ask about interactions the image cannot prove; do not copy personal data from it.
For example: Use the attached layout while retaining current brand colors and copy.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Describe visible structure, spacing, and states. Ask about interactions the image cannot prove; do not copy personal data from it.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Change one thing
Locate the target and affected behavior. Make only this change and preserve other behavior and existing uncommitted work.
For example: Rename the reservation button to “Confirm date” without changing reservation logic.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Locate the target and affected behavior. Make only this change and preserve other behavior and existing uncommitted work.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Report an issue
Use reproduction steps, expected behavior, and actual results. Confirm reproducibility before proposing the smallest fix.
For example: Submitting a selected date gives no feedback; a reservation number is expected.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Use reproduction steps, expected behavior, and actual results. Confirm reproducibility before proposing the smallest fix.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Investigate first
Gather evidence and propose hypotheses without presenting guesses as conclusions. Give result branches and stop conditions for each check.
For example: Reservations disappear after refresh; inspect save and retrieval records first.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Gather evidence and propose hypotheses without presenting guesses as conclusions. Give result branches and stop conditions for each check.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Recover a stable version
Summarize attempted changes and evidence, identify a recoverable version, explain restoration impact, preserve current material, and plan one fix.
For example: The date picker stopped working after the second change; review the changes first.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Summarize attempted changes and evidence, identify a recoverable version, explain restoration impact, preserve current material, and plan one fix.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Check completion
Check agreed acceptance criteria and record actions, observed results, and unverified items. A successful build does not prove the user flow works.
For example: Complete a reservation and find it after refresh; reject an empty date.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Check agreed acceptance criteria and record actions, observed results, and unverified items. A successful build does not prove the user flow works.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Hand off to a new conversation
Summarize goals, scope, file locations, start commands, status, checks, and next steps. The next conversation must verify original files before continuing.
For example: The reservation page is built; persistence is unverified. Test it next.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Summarize goals, scope, file locations, start commands, status, checks, and next steps. The next conversation must verify original files before continuing.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Review before release
Identify the release version, public entry, costs, data handling, acceptance checks, and recovery steps. List blockers and wait for a release decision.
For example: Complete a reservation at the test address without real notifications; do not publish yet.
Why this helps: AI can separate the current goal from existing work and provide results that can be checked.
View prompt and checks
Edit the template with actual material, confirm, then copy into the current project conversation. Mark unknowns as undecided and replace examples as needed.
Confirmed · personalized content ready to copy
[What stage has the project reached] [Enter the purpose, current progress, and existing page.]
[Who will use the project] [Enter the intended users.]
[What is wrong now, and what should happen instead] [Enter the current problem and desired result; refer to the separately labeled example above.]
[What files, screenshots or steps are available] [Attach agreed scope, relevant screenshots, or reproduction steps; request missing material first.]
[What are the limits, and what should stay unchanged] [Limit work to this goal; budget and release date are undecided.]
Task: Identify the release version, public entry, costs, data handling, acceptance checks, and recovery steps. List blockers and wait for a release decision.
Working instructions:
1. Restate the goal, scope, and acceptance criteria; separate facts, inferences, and unknowns.
2. Ask one essential question when missing material could change the result. Do not invent details.
3. Provide actionable steps, deliverables, and checks. Explain technical words in plain language.
4. Preserve a recoverable version before changes. Await confirmation before deleting data, spending money, expanding access, or publishing.
5. Report changes, actual checks, observed results, unverified items, and the next step.
[Acceptance] Check the stated goal item by item, including normal and failure cases. If execution is unavailable, label the output as advice or unverified work rather than complete.What should you check after AI responds?
- Does it address this goal without adding unrelated features?
- Does it include evidence and clearly mark unexecuted work?
- If a criterion fails, identify the deviation and request a focused correction and recheck.
Troubleshoot by observed result
AI loses the requirements
Ask AI to restate the scope. Incorrect → provide the latest agreed material and recheck. Correct but implementation differs → use “Change one thing.” Repeated drift → prepare “Hand off to a new conversation.”
AI advises but does not execute
Check whether the tool can edit project files and run commands. If yes, request actual files and execution results. If not, move the handoff material to a capable tool. Advice is not completed work.
Repeated changes do not solve it
Stop broadening changes and record attempts and results. Reproducible → investigate evidence, then make one fix. Not reproducible → add device, steps, and time. Preserve the current state before overwriting anything.
Need an AI tool?
Six checks before choosing an AI tool
Compare tool types here; unverified brands and prices are not recommendations. The assistant helps build; choose product technologies in the stack module.
Would you rather work in a browser or install a tool?
Browser-based tools can reduce local setup, but verify export and regional availability. Local assistants suit existing files but require installation and environment checks.
Check this with your own project: No installation → exclude tools requiring local installation. Local-only processing → exclude tools requiring project upload.
Will it only display content, or also save user input?
An introduction page needs content and UI. A data app also needs service connections, identity rules, and real write/read checks. Generating a screen does not connect these.
Check this with your own project: Submit and reload a test record, then check another identity. Tools unable to test real storage need a separate execution environment.
How can the tool continue an existing project?
Provide files, startup steps, approved requirements, and recent checks. Ask AI to read and restate them, then change one text label before rebuilding anything.
Check this with your own project: Expand the task only after opening the original project, preserving behavior, and explaining changed files. Resolve import problems first.
What do the tool, AI usage and project running each cost?
Tool subscriptions pay for assistant access; AI usage follows the provider’s billing rules; product operation covers domains, hosting, and data services. They may be billed separately.
Check this with your own project: Inspect current billing, allowances, and overage rules. Mark unverified amounts unknown and set a budget before buying or releasing.
Can you keep the files and continue after switching tools?
Check regional availability, file export, independent execution, and data export/migration before building the full product.
Check this with your own project: Export a sample and run it in a separate directory. Record restrictions and decide whether the dependency is acceptable.
Try one small task before committing
Give candidates the same small requirement: a date input and result page. Record cost, export ability, and whether a reported error can be fixed.
Check this with your own project: Compare the same task and criteria. An attractive screenshot does not prove persistence or maintainability.
Prepare tools and material
Chat tools can support early discussions. Development requires an assistant able to access project files and run checks. For an existing project, include file locations, start commands, and recent check results. Explain which part of a reference image matters and remove personal data and keys. Confirm capabilities and costs in the current tool and official documentation.
A prompt is the task instruction sent to AI. Context is the background material provided. Acceptance criteria are observable conditions for completion.
Next: turn the agreed scope into requirements →