[!NOTE]
clean-backend is a Claude Code skill. It supplies the operational habits an AI assistant does not reach for on its own when it designs, writes, or reviews your backend code.
Most "best practices" lists tell your AI things it already does perfectly well. We measured which ones it actually misses — then deleted everything else.
The measurement
We put neutral production-code tasks — a payment endpoint, a product listing, a delete endpoint — in front of a fresh model with no skill loaded, then repeated them with the skill, and diffed the output. Thirteen runs.
Applied reliably with zero prompting, so we cut them from the skill: field-limited responses, validation at the boundary, graceful degradation, intention-revealing naming, and complete idempotency on money-moving endpoints in 6 of 6 runs.
Never appeared in a single neutral run, so they became the skill: API versioning, feature flags, circuit breakers.
The full protocol, every prompt, the scorecard, and the two places our own methodology was wrong: evals/.
See it in action
Seven neutral baselines. A charge endpoint, a product listing, a delete endpoint. Not one of them versioned the route:
// What a fresh model writes, every time:
POST /charges
// What it does not write unless something tells it to:
POST /api/v1/charges
Mobile clients outlive your refactors. An unversioned route turns the first breaking change into a coordinated migration — and nothing inside the endpoint you're writing ever reminds you.
The seven
Part 1 — absent from every baseline. Lifecycle decisions that are invisible from inside a single endpoint.
Part 2 — present only when the task cues them. The model does these when the task makes them obvious, and skips them when it doesn't.
Read the skill →
Install
As a Claude Code plugin (recommended)
/plugin marketplace add AllanOps/Clean-Backend
/plugin install clean-backend@allanops
Then either invoke it explicitly with /clean-backend:clean-backend, or just ask for backend code — it triggers on its own.
Manual install
git clone https://github.com/AllanOps/Clean-Backend.git
cp -r Clean-Backend/skills/clean-backend ~/.claude/skills/clean-backend
Installed this way it's invoked as plain /clean-backend.
Why you can trust this repo
A skill is instructions injected into your agent, so every change is treated as a supply-chain change:
- Schema-validated — the skill and plugin manifests are checked against the Agent Skills spec on every PR.
- Scanned for prompt injection — CI hard-fails on invisible or bidirectional Unicode, homoglyph/mixed-script words, encoded blobs, raw-IP links, un-allowlisted URLs, and instructions that would steer an agent to exfiltrate or execute anything. The scanner has its own test suite.
- Locked-down automation — Actions pinned by commit SHA, read-only tokens, and a zizmor audit of the workflows themselves.
See SECURITY.md for the threat model and private reporting.
Contributing
The bar here is unusual: a new habit has to be one the model doesn't already apply on its own. Bring evidence, or bring a scenario we can run. Start with a new-habit issue; the format and the evidence bar are in CONTRIBUTING.md.
By contributing you agree your work is licensed under the same MIT license as the project.
License
MIT © AllanOps
