Agent Architecture Audit Kit
A strong model masks architecture debt. Docs that describe a system you retired months ago, the same rule pasted into four files bloating every context window, a format drift that silently disabled an automation with zero errors: my own agent had all of it, and I never felt a thing because the model kept routing around the mess. This kit packages the exact review process that found it. You point your own agent (Claude Code, Cursor, Codex CLI, anything that reads files) at one file and it audits itself across 7 phases: discovers its own architecture without assuming any layout, fact-checks every claim in its docs against the live system, hunts ghosts of retired tools, kills duplication, catches parser drift, then installs a weekly staleness scanner so the rot stays gone. You approve every phase. Nothing leaves your machine. You finish with a styled local HTML report plus a one-minute summary of what changed and what still needs your call.
Paid Digital Thoughts subscriber?
Yearly subscribers: all products free ($1179.94+ value)
Monthly subscribers: 1 free product per month — subscribe from $29.99/mo
Claim your free copy →Best for
- +Anyone running a coding agent for 3+ months who has retired or replaced tools along the way
- +Agent setups that feel fine on a strong model (that feeling is exactly what hides the debt)
- +People who want the audit to stay done: the kit leaves guards behind, not just a clean snapshot
Not for
- -Brand-new agent setups with nothing to drift yet
- -Anyone expecting a one-click tool: your agent does the work and you approve each phase
What you get
Package includes
- •START-HERE.md: the one instruction that matters, plus honest limits and FAQ
- •AUDIT.md: the master file your agent reads and follows, phase by phase
- •SAFETY.md: eight hard rules your agent must obey during the audit
- •phases/00-discovery.md through 06-report.md: the seven-phase protocol
- •adapters/claude-code.md, cursor-windsurf.md, generic.md: harness-specific discovery hints
- •templates/report-template.html: self-contained styled report skeleton
- •templates/staleness-scanner.md: the scanner recipe with full reference implementation
- •templates/inventory-table.md: the architecture inventory format
- •CHANGELOG.md: update history
FAQ
Which agents does it work with?
Any agent that can read files: Claude Code, Cursor, Windsurf, Codex CLI, aider, or a custom loop. Harness adapters cover the common ones; a generic adapter plus a degraded no-shell mode covers the rest.
Does any of my data leave my machine?
No. Your own agent runs everything locally. The report is a local HTML file. The kit makes no network calls.
How long does the audit take?
Typically 60 to 120 minutes of agent time across the 7 phases, with short check-ins where you approve fixes. You can stop after any phase and resume later; the kit writes its state to disk.
What if my agent architecture is unusual?
Phase 00 assumes nothing: the agent discovers whatever exists (instruction files, memory surfaces, scheduled jobs, generated context) and builds the audit around that inventory. The kit was QA tested by an agent running all seven phases against a deliberately broken synthetic setup.
Is this just a prompt?
It is a 17-file protocol: phased instructions with stop rules, safety constraints, harness adapters, a working scanner implementation, and report templates. The value is the method and the guards it leaves behind, not a single clever prompt.
You might also like:
Secure checkout by Stripe. Instant download + guided Claude Code setup.