SkillPartly tested (needs an account)
Safer backend auth, data and migrations
backendskills
Three backend skills that make agents scope queries to the caller, lock read-then-write paths, add idempotency keys and run expand-contract migrations.
- Works with
- Claude Code, Codex CLI
- Good for
- Build an app, Web developers
What it does
backendskills is a set of three instruction skills for backend work. api-endpoint makes the agent scope every query to the caller, turn bad input into a 4xx answer and rate-limit login and costly routes. data-integrity makes it use locks or conditional updates for read-then-write steps, add UNIQUE constraints and put idempotency keys on anything that charges or sends. db-migration makes it change schemas in an expand-then-contract order, with concurrent indexes, lock timeouts, batched backfills and down migrations.
The repository also ships the same rules as dist/AGENTS.md for Codex and other AGENTS.md agents, and as Cursor rule files. The author publishes the benchmark tasks and raw runs in the bench folder.
Example prompts
No example prompts are published in the README or SKILL.md files. The skills load from their descriptions: api-endpoint when you add or change any endpoint (especially login), data-integrity for code that changes money, stock, balances, counters or statuses, and db-migration for any new table, column, index, constraint, backfill or migration file.
How to install
npx skills add globalmpc/backendskillsHow to use it
Install, then ask for backend work as usual (a login route, a checkout, a schema change). The matching skill loads and the agent applies its defaults; each skill ends with a Make it yours section for your team's error format and API docs.
Check it works
Install with npx skills add globalmpc/backendskills -a claude-code, start a new session in the project and ask for a column rename. The plan should add the new column first, backfill in batches and drop the old column in a later step, instead of a single rename.
Common problems
- Skills not picked up by Claude Code — the default install is project-only and asks which agents to use; rerun with -a claude-code (and -g for every project) — https://github.com/vercel-labs/skills
- Installed with -y and nothing changed in Claude Code — -y without an agent flag puts the skills in ~/.agents/skills only — https://github.com/vercel-labs/skills
- Codex ignores the rules — append dist/AGENTS.md to your project's AGENTS.md — https://github.com/globalmpc/backendskills
Other assistants
- Codex:
npx skills add globalmpc/backendskills -a codex, or append dist/AGENTS.md — https://github.com/globalmpc/backendskills - No Gemini CLI route is documented.
What to watch out for
- npx skills add needs Node.js 22.20 or later. Without flags it is interactive and installs into the current project; add -a claude-code (or -a codex) to pick the agent and -g for all projects. With -y and no agent flag it installs only to ~/.agents/skills
- The skills CLI sends anonymous usage data and security-audit requests; set DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 to turn that off
- Instruction-only skills (three SKILL.md files, no bundled scripts), but they tell the agent to run commands: add and run tests, send 10 to 20 parallel requests, and for db-migration run every migration up, down and up again before calling it done. Make sure the agent points at a local or test database, not production
- Loads by itself: none of the three skills has disable-model-invocation, and the descriptions cover almost any endpoint, database write or schema change
- Benchmark numbers are the author's own runs on Sonnet 5.5 and Haiku 5.5 (5 runs per task, 2026-11-04). On the payment-webhook task Haiku did worse with the skills (1 of 5 to 2 of 5 runs with a critical mistake); the README says out-of-order webhook events are not fixed yet
- Young repo: the main branch has one commit (2026-11-05) and no releases
Licence and author
- Licence
- MIT
- Author
- globalmpc
- Source code
- github.com/globalmpc/backendskills
How we tested it
We installed it on an empty test computer on 8 Nov 2026. Partly tested: installed in a clean test system, but its main features need an account or key we did not use.
We did not run the skills in a live agent session or rerun the benchmark.
Technical details of the test
- Date
- 8 Nov 2026
- Method
- Reviewed by our team against the repository at 6131a31: README, the three SKILL.md files, dist/AGENTS.md and the bench folder. In a clean container, npx skills add globalmpc/backendskills -a claude-code -y (skills 1.7.6, telemetry off) copied the three skills into the project's .claude/skills folder.
Checked again on 8 Nov 2026
Alternatives
- pg-aiguide — Postgres guidance for table design, indexes and queries; it overlaps the schema half, not the endpoint and concurrency rules
FAQ
- Q: Does it run anything? A: The skills ship no scripts, but they tell the agent to run tests, parallel requests and migration up/down round trips with your own tools.
- Q: Which databases? A: The examples are Postgres with TypeScript and pg (error codes such as 23505 are Postgres codes); the skills say the rules apply to any stack.
- Q: Will it change my error format? A: Left as not set, the agent follows what your repo already does.
Sources
- github.com/globalmpc/backendskills
- github.com/globalmpc/backendskills/blob/main/skills/data-integrity/SKILL.md
- github.com/vercel-labs/skills