Skip to content

CLAUDE.md

Latest commit

 

History

History
79 lines (60 loc) · 3.5 KB

CLAUDE.md

File metadata and controls

79 lines (60 loc) · 3.5 KB

Agent-PROVE

Agent-PROVE is a Claude Code plugin that applies structured multi-framework thinking to validate approaches before execution, audit outputs during execution, and enforce evidence-based accountability. Extracted from the Mission 1086 methodology.

Architecture

Four layers, bottom to top:

  1. Framework Agents (14 built-in, user-extensible) — Independent thinking lenses across 6 categories. Each applies ONE framework and returns PASS/FAIL with cited evidence. Located in .claude/agents/.
  2. Orchestration Agents — T1 Validator (pre-execution) and Cycle Auditor (during execution). They dispatch framework agents in parallel and synthesize verdicts.
  3. Skills — Workflow logic for validate-approach, think-cycle, evidence-audit. Located in .claude/skills/.
  4. Commands — User-facing entry points: /validate, /think, /audit. Located in .claude/commands/.

Agent Communication Protocol

Every agent prompt MUST include:

  • Objective — Single unambiguous sentence starting with a verb
  • Input files — Exact paths to read
  • Output — Structured result with verdict
  • Schema — Expected output structure
  • Verification — How to confirm success
  • Constraints — What NOT to do

Orchestrator Accountability

Before presenting ANY result:

  1. Verify agent output exists and is valid
  2. Verify counts match expectations
  3. Verify no null/empty required fields
  4. Cross-check agent claims against actual data

Verdict Taxonomy

T1 Validator:

  • VALIDATED — All frameworks PASS, no critical findings
  • VALIDATED_WITH_ADJUSTMENTS — No critical findings, adjustments recommended
  • REJECTED — Critical finding or majority FAIL

Cycle Auditor:

  • CYCLE_APPROVED — All pass, no evidence violations
  • CYCLE_APPROVED_WITH_WARNINGS — Pass with non-critical warnings
  • CYCLE_FAILED — Evidence violation or critical framework failure
  • ESCALATE — 3 consecutive failures, stop and ask user

Framework Agents:

  • PASS — No issues found under this framework
  • FAIL — Issues found, with specifics and evidence

Framework Taxonomy (Phase-Aware)

Category Phase Frameworks Default For
Exploration Brainstorming six-thinking-hats /validate
Deconstruction Problem Understanding first-principles, five-whys /validate
Strategy Approach Design inversion, occams-razor, one-thing /validate
Verification Implementation Validation deep-dive, technical-research, tri-strike /think
Refinement Reduction & Tightening via-negativa, occams-razor manual
Agentic Audit AI Work Audit gap-finder, blind-spot, devils-advocate, drift-detector /think

/validate auto-selects: Exploration + Deconstruction + Strategy (5 frameworks) /think auto-selects: Verification + Agentic Audit (5 frameworks) Override with --framework-names for any combination.

Evidence Protocol

Every claim cites a source. Every number has a derivation. Zero uncited inference.

  • Factual claims → [source: file_path:line or URL]
  • Numbers → [derived: computation or file_path]
  • "Not found" → [searched: list of files/paths checked]

Custom Frameworks

Drop a .md file in .claude/agents/ following this pattern:

  1. Role definition ("You are a [Framework] thinking agent")
  2. Input spec (CONTEXT, SUBJECT, INPUT_FILES)
  3. Core questions (3+)
  4. Output format (structured findings with evidence)
  5. Verdict (PASS/FAIL with specifics)
  6. Evidence citation requirement