AgentSkillsCN

fpf-evidence

规划并记录证据（测试/基准/实验），以验证或确认某项主张。在宣称“这有效”“这是正确的”，或在增强信心之前，必须先调用此功能。

SKILL.md
--- frontmatter
name: fpf-evidence
description: Plan and record evidence (tests/benchmarks/experiments) to validate or verify a claim. MUST be invoked before stating "this works", "this is correct", or when promoting confidence.
argument-hint: "[claim-or-short-title]"

Goal

Turn a claim into an evidence-backed statement.

This skill produces an Evidence Record under .fpf/evidence/ and (when possible) runs the smallest viable test/measurement.

Output

Create a new file:

  • •.fpf/evidence/EVID-${CLAUDE_SESSION_ID}--<slug>.md

Use the template:

  • •.fpf/templates/evidence-record.md

Procedure

  1. •Define the claim
  • •Write the claim under test in one sentence.
  • •Link it to:
    • •a hypothesis in an anomaly record, or
    • •a decision in a DRR.
  1. •Deduce predictions
  • •What should be true if the claim holds?
  • •What result would falsify it?
  1. •Design the harness Pick the smallest credible check, e.g.:
  • •unit/integration tests
  • •type checks / linters
  • •reproducible benchmarks
  • •runtime traces/logs
  • •minimal simulation
  1. •Run the check (run-time)
  • •Run commands needed to generate evidence (ask permission when required).
  • •Record:
    • •exact commands,
    • •environment details,
    • •commit/revision,
    • •raw outputs (or links to them).
  1. •Interpret conservatively
  • •Label result: corroborated / refuted / inconclusive.
  • •Update confidence conservatively.
  • •Add valid_until if evidence can go stale.

Quality bar

  • •Evidence must correspond to a deduced prediction, not a vague "feels good".
  • •Keep raw output or provide reproduction steps.
  • •Don't hide failures -- refutations are valuable.