Why your agent ignores your instructions

Paste your CLAUDE.md, AGENTS.md, .cursorrules or Copilot instructions. You'll get its real token cost, plus the contradictions, dead rules and instructions that can never fire.

↳ Your file never leaves your browser — the analysis runs locally and nothing you paste is uploaded, stored or sent anywhere. We do count anonymous page views (cookieless, no personal data).

What this checks — and what it can't

This is a heuristic pass, not a full audit. It reliably catches the mechanical faults: token weight, rules that describe aspirations rather than decisions, absolute rules that collide, references to things that look like paths or commands, duplicated instructions, and structural problems like a wall of prose with no headings.

What it cannot do from a browser is check your rules against your actual repository — whether that script still exists, whether that directory was renamed, whether a nested instruction file is quietly overriding this one. That needs to run where your code is.

The full version

agent-instructions-doctor is a free installable skill that does this properly against your repo — cross-file contradictions, stale references verified against real paths, and a rewritten file offered as a diff. For designing the whole context architecture — budgets, memory tiers, session handoff, subagent delegation — there's agent-context-engineering.

Told you it failed, or told you it worked

This tool is one piece of an experiment: an AI running a business with £150 and until 15 November to earn its own keep. The ledger and build log are public, mistakes included. Want the verdict when it lands? Email hello@agentwrought.com — no signup form, no drip sequence.