agentleFS
motola10motola10/claude-code-base/.claude/skills-disabled/agent-to-confirm/SKILL.md

agent-to-confirm

Skill1 starsChanged 5 months agoMITView on GitHub

Validation and hallucination-prevention step that forces an honest self-audit before committing to an answer, solution, or action. Use this skill whenever the user wants to double-check, confirm, or validate an agent's output before acting on it. Trigger immediately on phrases like "confirm this", "are you sure?", "double check", "validate this solution", "verify your answer", "don't hallucinate", "be sure before you proceed", "only proceed if confident", "check your work", "sanity check", "are you certain?", "be brutally honest", or any request for a confidence assessment. Also trigger proactively when about to recommend an irreversible action (delete, deploy, force-push, drop table) or when the answer depends on a chain of assumptions that haven't been directly verified. This skill is the go-to whenever trust in accuracy matters — invoke it rather than proceeding with a quietly hedged guess.

  • Deletes or force-pushes

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do not inflate confidence to be reassuring.

    line 74
  2. Distinctive:

    Do not proceed with the full answer.

    line 88
  3. Distinctive:

    Ask the minimum number of focused questions needed to resolve the uncertainty (1–2 max)

    line 91
  4. Distinctive:

    Don't assign a confidence number just to look thorough — only compute it when you've genuinely gone through the decomposition

    line 124
  5. Distinctive:

    Don't hedge everything as "uncertain" to appear safe — that's also dishonest and unhelpful

    line 126
  6. Distinctive:

    Don't ask for clarification on things you can verify right now with available tools

    line 127
  7. Distinctive:

    Don't restate the full solution after declaring uncertainty — wait for the user's response first

    line 128
  8. Distinctive:

    Don't round up confidence to clear the 95% bar.

    line 129

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

What this skill does

Forces a structured self-audit of any opinion, solution, or planned action before presenting it as reliable. The goal is to catch hallucinations, flawed assumptions, and overconfidence before they cause real harm — not to add ceremony for its own sake.

The threshold is 95% confidence. Below that, surface the uncertainty explicitly instead of hedging quietly or guessing.


When to apply this

  • The user explicitly asks to confirm, double-check, or validate
  • You're about to recommend a specific command, code change, or configuration
  • The answer depends on earlier reasoning steps, and one wrong assumption collapses the whole thing
  • You're asserting facts about versions, APIs, external behavior, or state you can't directly observe
  • The action is hard to reverse (deleting files, pushing code, modifying live systems)

Step 1 — Decompose what you're claiming

Break the thing being confirmed into its independently verifiable components.

  • For a factual claim: list each sub-claim ("this function exists", "this flag does X", "version Y supports Z")
  • For a code solution: list each assumption the solution rests on
  • For a plan: identify which steps are certain and which have unknowns

This decomposition reveals where confidence actually lives — and where it doesn't.


Step 2 — Label each component honestly

For each component, determine how you know it:

LabelMeaning
VerifiedDirectly observed in this session (read the file, ran the command, saw the output)
InferredLogically follows from verified facts, but not directly confirmed
AssumedBased on pattern or convention, not confirmed for this specific context
UncertainI don't actually know this; I would be guessing

The point is not to be maximally pessimistic — it's to be honest. Verified and Inferred are legitimate bases for confidence. Assumed and Uncertain need to be surfaced.


Step 3 — Compute overall confidence

Overall confidence is bounded by the weakest critical component. Ask: "if the assumed or uncertain parts are wrong, does the whole answer fall apart?"

  • Any critical component labeled Uncertain → overall confidence cannot be 95%
  • Multiple Assumed components without verification → be skeptical of reaching 95%
  • A long chain of inferences where each step seems right → reduce confidence accordingly

Do not inflate confidence to be reassuring. A quietly wrong answer is worse than an honest "I'm not sure about X."


Step 4 — Respond based on confidence

If confidence ≥ 95%

Present the confirmed answer clearly. Briefly note what was verified vs. inferred so the user understands the basis. Don't pad with caveats — state it with appropriate confidence.

If confidence < 95%

Do not proceed with the full answer. Instead:

  1. State your actual confidence level (e.g., "I'm about 70% confident here")
  2. Name the specific things you're uncertain about — concrete gaps, not vague hedging
  3. Ask the minimum number of focused questions needed to resolve the uncertainty (1–2 max)
  4. Be clear about what you do know vs. what you're missing

Never guess quietly. The user is invoking this skill because they want to catch hallucinations. Surfacing uncertainty is the correct outcome, not a failure.


Response format

When confirming (≥ 95%):

**Confirmed** — [answer or solution]

Basis: [1–2 sentences on what was directly verified vs. inferred]

When uncertain (< 95%):

**Confidence: [X]%** — not sufficient to proceed

I'm confident about: [what you know]
Uncertain about: [specific gap(s)]

To proceed, I need: [1–2 focused questions]

What not to do

  • Don't assign a confidence number just to look thorough — only compute it when you've genuinely gone through the decomposition
  • Don't hedge everything as "uncertain" to appear safe — that's also dishonest and unhelpful
  • Don't ask for clarification on things you can verify right now with available tools
  • Don't restate the full solution after declaring uncertainty — wait for the user's response first
  • Don't round up confidence to clear the 95% bar; the threshold only has value if you respect it

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.