Recursive-Self-Audit
SamsaraMethod/Recursive-Self-Audit
Claude Skill that will help prevent unseen mistakes, smoothing, omissions, and lies. This skill will force the AI to review and analyze its own outputs against four specific failure modes: Mismanagement, Malfeasance, Rent Seeking and Incompetence.
Instalação
npx convez-ai SamsaraMethod/Recursive-Self-Audit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Recursive Self-Audit — v1.2 Complete Deliverable
This zip contains the v1.2 build of the Recursive Self-Audit skill plus the workspace artifacts from its three-iteration build (v1.0 → v1.1 → v1.2).
Contents
v1.2-complete/
├── README.md (this file)
├── recursive-self-audit/ (the skill — install this)
│ ├── SKILL.md (main skill entry point, 255 lines)
│ └── references/
│ ├── four-modes-detailed.md (each mode treated in depth + non-Samsara worked findings)
│ ├── audit-protocol.md (step-by-step protocol for running an audit)
│ ├── anti-patterns.md (eight ways audits fake themselves)
│ └── templates.md (six ready-to-paste audit templates)
└── v1.2-workspace/ (build artifacts, not part of the skill)
├── adversarial-test-cases.md (v1.2 test design)
├── test-A-keto.md (F-stress test: user-wants-specific-conclusion)
├── test-B-bridge.md (I-stress test: analyst-lacks-domain-knowledge)
├── test-C-coffee.md (R-stress test: framework-doing-all-the-work)
├── tests-D-E-compressed.md (strategic plan + peer review, compressed)
└── audit-on-v1.2.md (the v1.2 build's audit on itself)
What to install
The recursive-self-audit/ directory is the skill. Install it where your skill loader expects skills (typically ~/.claude/skills/ or equivalent). The skill's entry point is SKILL.md; the four reference files are loaded as needed via progressive disclosure.
What the workspace is for
The v1.2-workspace/ directory documents the iterative build. It is the audit trail for how the skill was tested and refined across three iterations. It is not part of the skill's runtime. Read it if you want to understand:
- How the skill was tested (
adversarial-test-cases.md+ the five test runs) - What the skill's most honest limitations are (
audit-on-v1.2.md) - The lineage of findings from v1.0 through v1.2 (audit-on-v1.2.md engagement of carry-forward findings)
Key finding from v1.2 (worth knowing before use)
The skill formalizes and extends analytical-honesty dispositions; it does not create them. On a model with baseline honesty / refusal / calibrated-uncertainty training, the skill's marginal value is in:
- Vocabulary for naming failure modes
- Structured audit sections in deliverables
- Second-order findings (audit-on-audit) that primary disposition misses
- Calibration for borderline cases
The skill is scaffolding for an existing discipline, not a substitute for it. Use it to extend and document the discipline, expecting that it amplifies what is there rather than supplying what is absent.
See recursive-self-audit/SKILL.md section "An honest note on what this skill adds, and what it relies on" for the canonical statement.
Versioning
- v1.0: initial build from Samsara corpus (Samsara Analytical Discipline v1/v2, OWR2 v3.0 audit registers, Tier VI self-audit, Meaning Gap v2.0→v3.0 closure map)
- v1.1: added non-Samsara worked findings (DB migration, forecasting, code review) after five-test run; resolved consolidation question on templates.md / anti-patterns.md
- v1.2: adversarial test pass (keto F-stress, bridge I-stress, coffee R-stress) plus two additional contexts (strategic plan, peer review); SKILL.md updated with honest marginal-value-vs-baseline disclosure
Authorship
Quantel Bolden in collaboration with Claude AI. Built within the Samsara Method project.