Skip to content
Back to skills

Audit

ASecurity

Deep 12-dimension audit of an EXISTING codebase via cortex-x existing-project-audit prompt (P0 detect → P1 repo-map → P2 4-agent audit → P3 human gate → P4 research → P5 synthesis → P6 ADR backfill opt-in). Saves AUDIT.md, recommendations.md, repo-map.md, and optional retroactive ADRs to cortex/ in the project. Auto-primed by SessionStart hook when .cortex-bootstrap-pending marker is mode=existing. Triggers (CZ+EN): "/audit", "audit my project", "audituj projekt", "co je v tomhle projektu", "...

  • 2 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added June 5, 2026
ai-agentsgodocumentation

Security analysis

A100/100

Scanned June 5, 2026

npx -y skills add Rejnyx/cortex-x --skill audit --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Audit?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Audit
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/rejnyx-audit/badge)](https://www.skillsdirectory.com/skills/rejnyx-audit)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: audit
description: Deep 12-dimension audit of an EXISTING codebase via cortex-x existing-project-audit prompt (P0 detect → P1 repo-map → P2 4-agent audit → P3 human gate → P4 research → P5 synthesis → P6 ADR backfill opt-in). Saves AUDIT.md, recommendations.md, repo-map.md, and optional retroactive ADRs to cortex/ in the project. Auto-primed by SessionStart hook when .cortex-bootstrap-pending marker is mode=existing. Triggers (CZ+EN): "/audit", "audit my project", "audituj projekt", "co je v tomhle projektu", "prozkoumej kódbázi", "review this repo", "what does this codebase do".
disable-model-invocation: false
---

# /audit — Deep existing-project audit (cortex-x)

**Voice charter:** see [`standards/voice.md`](../../../standards/voice.md). No greetings, no emoji, no emotion words, counts-not-praise. Phase 4 research follows `standards/web-research.md` citation discipline; AUDIT.md uses `[cortex/recall]` + `[^cN]` footnotes when grounded in cortex memory.

You are running the cortex-x existing-project deep audit. Read and execute the
prompt at `~/.claude/shared/prompts/existing-project-audit.md` from start to
finish.

## Phase summary

| Phase | Goal | Output |
|---|---|---|
| P0 — Detect | What kind of project is this? | `cortex/audit-context.md` |
| P1 — Repo map | What's the symbol-level shape? | `cortex/MEMORY/repo-map.md` |
| P2 — Audit (4 parallel agents, 12 dims) | Where are the bones, hot spots, gaps? | `cortex/AUDIT.md` |
| P3 — Human gate | What CAN'T be derived from code | 5 questions, folded into AUDIT.md |
| P4 — Auto-research | What does 2026 say about this stack? | `$CORTEX_DATA_HOME/research/<slug>-audit-<date>.md` |
| P5 — Synthesis | What should we DO? | `cortex/recommendations.md` + CLAUDE.md patches |
| P6 — ADR backfill (opt-in) | What past decisions deserve documentation? | `cortex/decisions/ADR-*.md` |

## Flags

- `--backfill-adrs` → run Phase 6 (otherwise skipped, audit summary surfaces detected count)
- `--token-budget=NNN` → repo-map renderer token budget (default 1500)
- `--no-research` → skip Phase 4

## On completion

- Delete `.cortex-bootstrap-pending` marker (if present)
- Print Phase 7 closing instructions per the prompt
- Suggest `/retrofit` as the natural next step (apply cortex-x patterns informed by the audit)

## Don't confuse with

- `~/.claude/shared/prompts/project-scan.md` — quick 5-section institutional summary, populates `$CORTEX_DATA_HOME/projects/<slug>.md`. Different scope.
- `~/.claude/shared/prompts/retrofit.md` — APPLIES cortex-x patterns. Best run AFTER `/audit` so it has `cortex/AUDIT.md` to ground decisions.

## Reference

Full prompt: `~/.claude/shared/prompts/existing-project-audit.md`
Sprint 1.5 design rationale: `$CORTEX_HOME/docs/sprint-1.5-design.md` §2.3

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.