Agent Skills: Confidence Levels

>-

UncategorizedID: jagreehal/jagreehal-claude-skills/confidence-levels

Install this agent skill to your local

pnpm dlx add-skill https://github.com/jagreehal/jagreehal-claude-skills/tree/HEAD/skills/confidence-levels

Skill Files

Browse the full folder contents for confidence-levels.

Download Skill

Loading file tree…

skills/confidence-levels/SKILL.md

Skill Metadata

Name
confidence-levels
Description
>-

Confidence Levels

Critical rules

  • Express confidence as a percentage, not vague certainty ("70% confident").
  • Below 95%: mandatory "Why not 100%?" with gaps.
  • Show the math: evidence adds, gaps subtract. Thorough presentation ≠ correctness.
  • If confidence < 80% and you can gather more evidence yourself, gather first (research-first).
  • Trigger words: "root cause is", "definitely", "certainly", "clearly", "obviously", "complete clarity".
  • Before scoring or formatting, read references/scoring.md.

Workflow

  1. Inventory evidence (direct vs assumed). Strongest piece?
  2. Falsifiability: what would invalidate this? Did you look?
  3. Audit assumptions: mark each [VERIFIED] or [ASSUMED].
  4. List alternatives and why yours is more likely.
  5. Score from ~50% using adjustments in the scoring reference; present in the standard format.
  6. If more evidence is reachable (search, read, run), gather and reassess before asking the user.

Resources

  • references/scoring.md — scale, calibration, adjustments, format, examples, rationalizations. Read before stating a conclusion.

Validation

  • [ ] Confidence expressed as a percentage
  • [ ] Gaps explained when below 100%
  • [ ] Evidence shown for the claimed %
  • [ ] Self-gatherable evidence already gathered
  • [ ] Falsifying evidence considered

Constraints

  • Do not theatre-wrap trivially checkable facts (file exists, test passes) in percentages.
  • Related: research-first, critical-peer, concise-output, debugging-methodology.