Don’t just ask AI to continue. Ask for proof.
AI agents often say:
- “Done.”
- “Fixed.”
- “Should work now.”
But if you are not an engineer, the real question is not whether the AI sounds confident.
The real question is:
- What did it actually change?
- Did it verify the result?
- What evidence is missing?
- Is it safe to continue?
- What should you ask next?
AskProof is an AI acceptance and follow-up Skill for non-engineers.
AI agents summarize what they did.
AskProof helps you decide whether you can trust it.
It does not do the work for the AI, and it does not replace tests, CI, PR review, or verification tools. It helps you understand what the AI did, check whether the claim is believable, and turn vague feedback like “still broken”, “continue”, or “fix it yourself” into a prompt that asks for proof.
AskProof can turn an agent’s latest reply into the next clear, self-contained prompt with current goal, current status, constraints, proof requirements, and final output expectations.
Core path:
Agent reply -> AskProof checks current status -> AskProof generates the next structured prompt -> user sends it back to the agent
The AI says:
Fixed. It should work now.
A normal user might say:
Great, continue.
AskProof reminds you:
Current scenario: Done Check
Completion confidence: Low
Current status: The AI claims the issue is fixed, but no verification evidence is visible.
Missing evidence:
- What changed.
- How the AI verified it.
- What the verification result was.
- What remains unverified.
You should now ask the AI this one sentence:
Please do not continue building new features yet. First explain what you changed, how you verified it, what the result was, and what remains unverified.
- Why You Need AskProof
- What AskProof Helps You Do
- Who It Is For
- V0.2 Core Experience Enhancements
- Try It in 30 Seconds
- Try It Without Installing
- Ask Claude Code / Codex To Install It
- Core Features
- Common Use Cases
- Examples
- How It Differs From Verification Tools
- Installation
- Project Structure
- Roadmap
AI agent summaries are often useful, but they are not the same as acceptance evidence.
An AI can say:
- “I fixed the bug.”
- “I updated the implementation.”
- “The issue should be resolved.”
- “Everything is ready.”
Those sentences may only mean that files were changed. They do not prove that the original problem is gone, that the user-facing flow works, or that the next task is safe to start.
AskProof helps you separate:
- what was changed
- what was actually verified
- what evidence was provided
- what is still unknown
- what you should ask next
You do not need to know Git, terminals, logs, stack traces, or test commands.
AskProof helps non-engineers ask AI agents for the right proof before trusting “done”.
It can help you:
- check whether “done” is actually verified
- turn an AI reply into a structured confirmation prompt for the next agent message
- identify missing evidence before continuing
- rewrite vague feedback into an actionable prompt
- ask for the easiest useful evidence when you do not know what to provide
- understand AI technical reports in plain language
- stop scope drift before the current work is accepted
- create a handoff note before switching chats, models, or platforms
AskProof is not another AI coding tool. It is an AI acceptance assistant for the human using the AI tool.
AskProof is built for people using AI agents without being professional developers:
- founders and solo builders
- product managers
- operators
- designers
- content creators
- education product owners
- anyone using Claude Code, Codex, Cursor, OpenClaw, domestic agent platforms, or other AI tools
You can use AskProof even if you do not know how to read logs, run tests, inspect diffs, or explain technical errors.
AskProof V0.2 strengthens the acceptance workflow without adding a UI, integrations, or an automatic test runner.
- Acceptance Brief: turns a “done” claim into a structured acceptance result with confirmed facts, existing evidence, missing evidence, confidence, status, and next action.
- Evidence Type: labels proof as code evidence, log evidence, visual evidence, user acceptance evidence, or reproducible evidence.
- Same-Agent Verification Mode: helps Claude Code, Codex, or another current agent inspect local project evidence instead of only asking the user for logs.
- Self-contained follow-up prompts: includes the feature name, target, evidence needed, and a stop condition before expanding scope.
- Minimum Acceptance Path: gives the smallest practical path to verify the current result.
- Reply Confirmation Prompt: turns an AI reply into a copy-ready prompt that confirms current goal, current status, constraints, required proof, and final output expectations.
Reply Confirmation Prompt is not background automation. It works when the user invokes AskProof after an agent reply or pastes the reply into the chat and asks what to send next.
You can test the idea in any AI chat by copying this prompt:
Act as AskProof, an AI acceptance and follow-up assistant for non-engineers.
When I paste an AI agent’s “done”, “fixed”, or “should work now” message, do not assume it is true.
Tell me:
1. The current scenario.
2. Whether the claim is believable.
3. What information is confirmed.
4. What information is still missing.
5. The risk of continuing now.
6. The safest next action.
7. One copy-ready prompt I should send back to the AI agent.
Use plain language. If there is no evidence, mark it as unverified.
AI says it is fixed. Is it really done?
If AskProof is working, you should see ideas like:
- completion confidence
- confirmed information
- missing evidence
- risk
- next prompt
It should not simply say “yes, it is done”.
After installation, you can also write naturally:
- “AI says it is fixed. Is it really done?”
- “still broken”
- “I do not understand this error.”
- “Rewrite this into a better prompt.”
- “I am opening a new chat. Create a handoff.”
- “What did this round actually do?”
- “What should I ask the AI now?”
- “Turn this AI reply into the next prompt I should send.”
Advanced users can still use optional commands:
/askproof done-check/askproof prompt-rescue/askproof evidence/askproof explain/askproof drift-guard/askproof handoff/askproof confirm
If you use Claude Code, Codex, Cursor, or another AI agent that can work with local files, you can ask it to install AskProof for you.
Copy this prompt:
Please help me install the AskProof Skill from:
https://github.com/happiness9527/askproof-skill
Goal:
Install the local `askproof/` Skill folder so this AI tool can use it.
Please:
1. Check whether this environment supports local Skills or instruction folders.
2. Download or clone the repository if needed.
3. Copy the complete `askproof/` folder, not only `SKILL.md`.
4. Tell me the exact destination folder you used.
5. Tell me whether I need to restart or reload the app.
6. Verify installation with this test prompt:
"AI says it is fixed. Is it really done?"
Do not add integrations, background services, or automatic test runners.
If your AI tool does not support Skills, ask it to explain the closest supported alternative before copying files.
The two most common entry points are Done Check and Reply Confirmation Prompt.
Checks whether an AI agent’s “done”, “fixed”, or “should work now” message is believable.
It reports:
- completion confidence: low, medium, or high
- current status: changed, verified, unverified, risky, safe to continue, or not recommended
- evidence the AI provided
- missing evidence
- how the user can accept the result
- the next prompt to send to the AI
Turns low-quality feedback into a clear prompt.
Weak feedback includes:
- “still broken”
- “continue”
- “just fix it”
- “fix it yourself”
- “don’t make bugs”
AskProof explains why the feedback is inefficient and rewrites it into a prompt that asks the AI to restate the failure, list what changed, request minimum evidence, avoid blind refactors, and define the next verification step.
AskProof does not assume you know logs.
When information is missing, it asks for the easiest useful evidence first:
- L0: you only know “it does not work”
- L1: you can describe what you see
- L2: you can share a screenshot
- L3: you can paste the AI’s last “done” message
- L4: you can paste the last red error line
- L5: you can provide terminal logs, Git diff, or test results
AskProof explains AI technical reports in plain language:
- what the AI is saying
- what it actually did
- what it means for your project
- what has been confirmed
- what has not been verified
- what you should ask next
When the task keeps expanding before the previous result is verified, AskProof warns you to close the loop first.
It helps you decide whether to continue, stop, verify, or create a handoff note.
When the chat is too long or you want to switch models, tools, or platforms, AskProof creates a handoff note with:
- current goal
- completed work
- verified work
- unverified work
- current risks
- key decisions
- next prompt
- mistakes to avoid repeating
When an AI agent finishes a turn and you are unsure how to respond, AskProof can turn that reply into a structured prompt for the next step.
It keeps the next message clear:
- current goal
- current status
- what the agent claimed
- what remains unverified
- before continuing
- requirements
- constraints
- acceptance requirements
- final output
It only works when you ask AskProof to do this or paste the agent reply. It does not automatically listen to Codex, Claude Code, Cursor, or other agent tools.
AskProof is useful when:
- Cursor says it fixed a UI bug, but you did not see a screenshot or browser check
- Claude Code reports file edits, but skipped tests
- Codex ran a technical check, but you are unsure whether it proves the user-facing issue
- a domestic agent platform says “fixed and redeployed”, but gives no evidence
- a product manager only knows how to say “still wrong”
- a designer sees the UI still looks off but cannot describe it in engineering terms
- an operator wants an automation fixed without blindly approving another rewrite
- a creator or education product owner needs a handoff before opening a new chat
- a user has an agent reply and needs the next structured prompt to send back
See the example gallery for more fictional but realistic cases.
User:
The AI said: “Fixed. It should work now.”
AskProof:
Current scenario: Done Check
Completion confidence: Low
Current status: changed claim, unverified
What is confirmed:
- The AI claims it made a fix.
What is not confirmed:
- Which files changed.
- Whether the app was run.
- Whether the original failure disappeared.
- Whether any test, screenshot, or manual check was done.
Risk:
- If you continue adding features now, later failures may be harder to trace.
Safest next action:
- Ask the AI to provide proof before making more changes.
You should now ask the AI this one sentence:
Please do not continue building yet. First show what you changed, what exact verification you ran, the result of that verification, and what remains unverified.
User:
still broken
AskProof rewrites it:
This feedback is too vague because the AI does not know what failed, where you saw it, or what changed since the last attempt.
Copy this prompt:
Please stop making broad changes. First restate the current failure in plain language, list the files you changed in the previous attempt, explain the most likely failure point, tell me the smallest evidence you need from me, and do not refactor unrelated code before we verify the current fix.
AskProof does not say only “send logs”.
It says:
You do not need to understand logs. Send any one of these:
1. A screenshot of the current page.
2. The AI’s last “fixed/done” message.
3. The last red error line from the page or terminal.
4. A plain-language description of what you clicked and what happened.
User:
The login fix is not verified yet, but please also add a new dashboard and redesign settings.
AskProof:
Current scenario: Drift Guard
Current judgment: Do not expand scope yet.
Reason:
The previous fix has not been verified. If you add new features now, later problems may be hard to trace.
Safest next action:
Ask the AI to prove the current fix works before adding new work.
You should now ask the AI this one sentence:
Please pause new features. First verify the login fix with clear evidence, then summarize what is safe to build next.
AskProof can generate a handoff before switching tools:
Current goal:
Make the fictional “TinyInvoice” app export invoices as PDF.
Completed:
- Added an export button.
- Connected the button to a PDF export function.
Verified:
- None provided yet.
Unverified:
- Whether the PDF downloads.
- Whether invoice totals are correct.
Current risk:
- The AI may have changed UI code without proving the export flow works.
Next prompt:
Please continue from this handoff. Do not add new features until you verify that the PDF export button downloads a PDF with the correct invoice total.
User:
The AI said it changed auth.py and the login flow should work now. I want to ask Codex what to do next, but I do not want it to blindly continue.
AskProof:
Current scenario: Reply Confirmation Prompt
Current goal:
Verify whether the login failure is really fixed.
Current status:
The AI claims it changed auth.py, but no verification result is visible.
What the agent claimed:
- It changed auth.py.
- It updated the login flow.
- It believes the login flow should work now.
What remains unverified:
- Whether the original login failure is gone.
- Whether the login flow was actually run.
- Whether there is a test result, log, screenshot, or manual acceptance result.
Copy this prompt:
Please continue from the current login fix, but do not add new features yet.
Current goal:
Confirm whether the login failure is actually fixed.
Current status:
You said auth.py was changed and the login flow should work, but I have not seen proof yet.
Before continuing, please confirm:
1. What exact login problem you fixed.
2. What files or behavior changed.
3. Whether you ran or manually checked the login flow.
Requirements:
1. If you verified it, show the command, screenshot, log, or exact manual steps and result.
2. If you did not verify it, mark it as unverified.
3. Do not expand scope before this login fix is accepted.
Constraints:
- Do not invent verification results.
- Do not treat “code changed” as “feature verified”.
- Do not refactor unrelated login code before proving the current fix.
Acceptance requirements:
- Provide a passing login check with command output, screenshot, log, or reproducible manual steps.
Final output:
List changed, verified, unverified, risk, and the safest next action.
AskProof does not replace your tests, CI, PR review, or verification tools.
It helps non-engineers ask better follow-up questions when AI agents claim work is done.
Other tools may help an agent submit verification evidence. AskProof helps the human user ask for that evidence, understand it, and avoid being carried along by confident but unproven AI summaries.
For Codex-style local skills, copy the askproof/ directory into your skills directory, then
restart or reload your agent environment.
See docs/install.md for step-by-step instructions.
If you do not use Git, see docs/install-for-non-engineers.md.
Advanced users can start with:
git clone https://github.com/happiness9527/askproof-skill.gitThen copy the complete askproof/ folder into the Skill directory used by your AI tool.
askproof-skill/
├── README.md
├── README.zh-CN.md
├── askproof/
│ ├── SKILL.md
│ ├── references/
│ ├── examples/
│ └── scripts/
├── docs/
└── demo/
Useful docs:
- Quick start for non-engineers
- AI acceptance checklist
- Prompt library
- How it works
- GitHub publish guide
See docs/roadmap.md.
V0.1 focuses on:
- clear README and examples
- the AskProof Skill instructions
- reference templates
- fictional demos
- lightweight local memory helpers
It does not include integrations, a UI, background services, PR bots, or automatic test execution.
V0.2 focuses on structured acceptance:
- AskProof Acceptance Brief
- Evidence Type classification
- Same-Agent Verification Mode
- self-contained follow-up prompts
- Minimum Acceptance Path
- Reply Confirmation Prompt
MIT. See LICENSE.
If AskProof helps you avoid blindly trusting “done”, consider starring the repo so more non-engineers can find it.