Make AI Code Run the Gauntlet So You Don't Have To Read It

What it does

old-coder is a plain-markdown skill that instructs AI coding agents to stop dumping code for human review and instead run it through a brutal verification pipeline. Before writing a line of implementation, the agent must draft a SPEC test plan for your approval; after coding, it delivers an EVIDENCE report summarizing test results, coverage, mutation scores, and supply-chain checks. You read two documents instead of hundreds of lines of agent-generated Python.

The interesting bit

The methodology is essentially extreme TDD adapted for an era where code volume is infinite and attention is not. It explicitly borrows from Uncle Bob’s strategy of surrounding agents with constraints—unit tests, Gherkin, mutation testing, property-based fuzzing—until the only rational move is to produce correct code. The demo rate limiter even caught real defects and an unsound mutation runner when rerun in a fresh context, which is either a bug or a surprisingly honest feature.

Key highlights

• Agent-agnostic: works with Claude Code, Cursor, Codex CLI, Aider, or custom loops via SKILL.md and AGENTS.md.
• Risk-calibrated gauntlet: a typo fix gets light checks; anything touching money, auth, or concurrency runs the full suite plus hostile-input probing.
• Rigid honesty rules: agents must label unverified checks as unverified, cannot weaken tests to force passes, and must disclose if the spec lacked human approval.
• Includes a working demo rate limiter with published evidence: 41 tests, 100% statement/branch coverage, and 22 of 22 planted bugs caught by mutation testing.
• Evidence reports bounded confidence rather than absolute proof, explicitly acknowledging that the gauntlet cannot validate its own checkers or prove spec completeness.

Caveats

• The gauntlet is only as good as the spec you approve; it cannot catch missing requirements or authenticate its own tooling.
• The README admits the demo’s earlier “green” evidence contained real behavioral defects that only surfaced during fresh-context reruns, so trust but verify (and rerun).
• It is a process and a prompt template, not an installable framework; adoption requires agent cooperation rather than package installation.

Verdict

Worth adopting if you are drowning in AI-generated code and want to replace line-by-line review with auditable, rerunnable proof. Skip it if you expected a testing library; this is discipline dressed up as markdown.