backendskills

agent
Guvenlik Denetimi
Uyari
Health Gecti
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 60 GitHub stars
Code Uyari
  • process.env — Environment variable access in bench/app/src/db.ts
  • process.env — Environment variable access in bench/app/src/payments.ts
  • process.env — Environment variable access in bench/app/src/server.ts
  • network request — Outbound network request in bench/app/test/app.test.ts
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Skills that stop AI agents from writing backend code that passes the demo and breaks in production. Critical mistakes: 25/150 runs without, 2/150 with. Claude Code, Codex, Cursor.

README.md

backendskills

Agent skills that make Claude Code, Codex and Cursor write backend code like a senior engineer:
safe on auth, data, money and schema changes by default.

Same prompt without and with backendskills: password guessing is blocked, a rename no longer breaks a rolling deploy, critical mistakes 25 → 2

Results

We gave agents 15 backend tasks with and without the skills, then ran a hidden check the agent never
sees: 20 buyers at once for 5 items in stock, a user reading someone else's order, the old app version
running against the new schema mid-deploy.

Critical mistakes: 25/150 runs without the skills, 2/150 with them. (Sonnet 5.5 and Haiku 5.5, 5 runs per task.)

Task Haiku 5.5 Sonnet 5.5
checkout 1/5 → 0/5 2/5 → 0/5
get-order 0/5 → 0/5 0/5 → 0/5
login-rate-limit 5/5 → 0/5 5/5 → 0/5
orders-check-constraint 0/5 → 0/5 1/5 → 0/5
orders-index 0/5 → 0/5 0/5 → 0/5
orders-pagination 0/5 → 0/5 0/5 → 0/5
orders-with-items 0/5 → 0/5 0/5 → 0/5
payment-webhook 1/5 → 2/5 0/5 → 0/5
product-discount 0/5 → 0/5 0/5 → 0/5
product-reviews 0/5 → 0/5 0/5 → 0/5
refund 0/5 → 0/5 0/5 → 0/5
rename-column 5/5 → 0/5 5/5 → 0/5
transfer 0/5 → 0/5 0/5 → 0/5
update-profile 0/5 → 0/5 0/5 → 0/5
wallet-adjustments 0/5 → 0/5 0/5 → 0/5
All 12/75 → 2/75 13/75 → 0/75

Runs with a critical mistake, without → with the skills. Most of the gap is two tasks: login-rate-limit
and rename-column account for 20 of the 25, and 10 of the 15 tasks had no critical mistake either way.
Not fixed yet: Haiku on out-of-order webhook events.

  • Team conventions (down migrations, API docs): 30 misses → 4, with the sample app's conventions
    written into each skill's "Make it yours" section (bench/skills-overlay). The
    whole with-skills arm uses that version. As shipped, those lines say "not set".
  • Cost: with the skills the agent does more: $0.11 vs $0.06 per run, 23 vs 14 turns
    (Sonnet $0.20 vs $0.12, Haiku $0.019 vs $0.012).
  • Left out: 25 runs stopped because the bench account hit its usage limit (usage_limit_reached):
    20 Sonnet with-skills runs on rename-column, 1 on orders-index, 4 Opus. Each run's reason is its
    error_reason in runs.jsonl. The 21 Opus runs there are a partial run, not part of the results.
    Every raw run, the tasks and how to rerun it are in bench/.

Install

npx skills add globalmpc/backendskills

Codex and other AGENTS.md agents: append dist/AGENTS.md. Cursor: copy
dist/cursor/ into .cursor/rules/.

Skills

Skill What the agent does by default
api-endpoint Scopes every query to the caller, turns bad input into a 4xx, rate-limits login and costly routes
data-integrity Locks or conditional updates for read-then-write, UNIQUE constraints, idempotency keys on anything that charges or sends
db-migration Expand then contract, concurrent indexes, lock timeouts, batched backfills, down migrations

Each skill ends with a Make it yours section for your team's choices (error format, API docs).
Left as "not set", the agent follows what your repo already does.

Contributing

Agent still made a mistake with the skills loaded? Open an issue with the prompt and the diff.
Edit skills in skills/, then node scripts/build-dist.mjs.

MIT


Maintained by Global MPC (Mineral Project Coin). Official site: https://globalmpc.tech

Yorumlar (0)

Sonuc bulunamadi