Found an Agent created skill that applies uncle bobs

An old coder's strategy for the agent era: don't read the code — make it run the gauntlet.

A skill that makes coding agents prove their work. Instead of you reading every line the agent writes, the agent must push its code through a gauntlet of checks — and hand you a test plan before coding and an evidence report after. You review those two documents, not the code.

It's plain markdown, so it works with any coding agent that follows instructions: Claude Code, Codex CLI, Cursor, Aider, or your own agent loop.

npx skills add https://github.com/amazingang/old-coder

Or manually:

Claude Code— copy the skill into a skills folder, then invoke/old-coder

or let it trigger on "prove it works"-style requests:

cp -r skills/old-coder ~/.claude/skills/    # or <project>/.claude/skills/

Other agents— addskills/old-coder/SKILL.md

to yourAGENTS.md

, rules file, or system prompt, and keepreferences/gauntlet.md

alongside it.

From Uncle Bob (Robert C. Martin), on working with coding agents (original tweet):

My current strategy is to not read any of the code written by my agents. That’s the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a plethora of others. In the end, I have very high confidence in the code they produce because they’ve had to run the gauntlet of all of my constraints and tests.

If you're not going to read the code, the things you do read have to carry the trust instead.

flowchart LR
    SPEC["📋 SPEC<br/>agent writes a test plan,<br/>you approve it"]
    RED["🔴 RED<br/>write a test,<br/>watch it fail"]
    GREEN["🟢 GREEN<br/>write code until<br/>it passes"]
    REF["🧹 REFACTOR<br/>clean up,<br/>tests untouched"]
    G["🛡️ GAUNTLET<br/>run every check"]
    EV["📊 EVIDENCE<br/>you read the report,<br/>not the code"]
    SPEC --> RED --> GREEN --> REF --> G --> EV
    REF -. next behavior .-> RED

You read two documents:

SPEC(before any code) — concrete examples of what the code must and must not do, plus which tools the agent wants to install. Approving it is the single yes/no you give.EVIDENCE(after the code) — real numbers from one final fresh run, rerunnable yourself with a single command.

The gauntlet in between:

Plus a menu of domain-specific layers — concurrency, UI checks, API compatibility, performance, observability — picked per task from a risk model (see references/gauntlet.md

).

Effort scales with risk: a typo fix runs a couple of checks; anything touching money, logins, data, or concurrency runs everything — plus the agent attacks its own code with hostile inputs first.

The agent grades its own homework, so the rules are strict: never weaken a test to make it pass; never report a check that didn't run; anything unverified is labeled unverified

, never pass

; if no human approved the spec, the report must say so and claim less confidence.

And one limit stated plainly: the gauntlet turns the constraints expressed in the spec into executable evidence; it cannot prove the spec is complete or authenticate its own checkers and mappings. That's why you approve the SPEC, and why EVIDENCE reports bounded, auditable confidence rather than absolute proof.

skills/old-coder/         the skill (SKILL.md + references/gauntlet.md)
demo-rate-limiter/        a rate limiter built end-to-end under the skill

The demo's evidence.md

is the point of the exercise: 41 tests, 100% coverage (49/49 statements and 20/20 branches), and 22/22 planted bugs caught. More importantly, fresh-context verification of earlier green states still found real behavioral defects and an unsound mutation runner — evidence that a green gauntlet is not self-authenticating. The current report discloses both the fixes and the final state's verification status. Rerun the whole report:

cd demo-rate-limiter
python3 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt -e .
./tools/gauntlet.sh

MIT