MPC gave Claude Code three backend skills and cut critical mistakes on 15 hidden-check backend tasks from 25 in 150 runs to 2
AI-written backends often work in the demo and break in production on concurrency, permissions and schema changes. backendskills is three open-source skills from MPC (globalmpc): api-endpoint (scope queries to the caller, 4xx on bad input, rate-limit login and costly routes), data-integrity (locks or conditional updates for read-then-write, UNIQUE constraints, idempotency keys on anything that charges or sends) and db-migration (expand then contract, concurrent indexes, batched backfills, down migrations). The authors built 15 backend tasks, each with a hidden check the agent never sees (20 buyers at once for 5 items, reading someone else's order, old and new app versions mid-deploy), and ran Sonnet 5.5 and Haiku 5.5 five times each with and without the skills. Critical mistakes fell from 25/150 to 2/150, mostly thanks to the login-rate-limit and rename-column tasks; each run cost roughly twice as much. We recomputed the published raw runs and ran the benchmark's 22 self-tests against a local PostgreSQL.
Visit site backendskills + Claude Code ↗- Pricing
- Free
- Region
- Global
- Where it's from
- Overseas
- Added
- Nov 14, 2026
- Tested on
- Nov 14, 2026
- Version we used
- globalmpc/backendskills 5a1779f; author bench used Claude Code 2.1.318, skills hash eff11b3dd7ca
- Categories
- Coding
- Original author
- MPC (globalmpc) ↗
- Our test
- Partly tested
- Cost / time
- The three skills are free under MIT. The agent does more with them; the authors measured about twice the cost per task ($0.11 vs $0.06). Rerunning the benchmark costs real API money.
The task
Have Claude Code, Codex or Cursor handle caller scoping, concurrency safety and zero-downtime migrations by default when writing endpoints, orders and schema changes, without reminding it in every prompt.
Steps
- In your project run npx skills add globalmpc/backendskills. For Codex and other AGENTS.md tools, append dist/AGENTS.md to your AGENTS.md; for Cursor, copy dist/cursor/ into .cursor/rules/.
- Fill in the "Make it yours" section at the end of each skill with your team's conventions (error format, where API docs live, whether down migrations are required). Left as "not set", the agent follows your repo. The benchmark's team-convention misses fell from 30 to 4 with conventions filled in.
- Ask for work as usual, for example a checkout endpoint or renaming a column on orders. The skills load when the task touches auth, money or schema.
- To check it yourself: in bench/, start PostgreSQL with docker compose up -d (port 5433), install the app's dependencies, put an API key in .env, run ./run.sh <task> <with|without> <runs>, then node summarize.mjs for the table.
Prompt
Rename the total column on the orders table to amount_cents. We deploy with rolling updates, so the old and new app versions share this database for a while; the old version must keep working.
What the author says
backendskills README (MPC, repo created 5 Nov 2026): 15 tasks × with/without skills × 5 runs on Sonnet 5.5 and Haiku 5.5; critical mistakes 25/150 → 2/150 (Haiku 12/75 → 2/75, Sonnet 13/75 → 0/75). 20 of the 25 come from login-rate-limit and rename-column, and 10 of 15 tasks had no critical mistake either way; Haiku on out-of-order payment webhooks is not fixed (1/5 → 2/5). Team-convention misses 30 → 4. Cost: $0.11 vs $0.06 and 23 vs 14 turns per run. 25 runs stopped at the bench account's usage limit and are left out; 21 Opus runs are a partial run and not part of the results. A full run of all 15 tasks costs about $95.
Our test
On 14 Nov 2026 we cloned globalmpc/backendskills (5a1779f). The repo's node bench/summarize.mjs over bench/results/2026-11-04/runs.jsonl (346 raw records) gives Haiku 12/75 → 2/75 and Sonnet 13/75 → 0/75, matching the README; it also lists the partial Opus runs (1/11 → 0/10), giving 26/161 → 2/160 with Opus, with 25 errored runs excluded. Without Docker, we started PostgreSQL 18.6 locally via embedded-postgres on port 5433, installed the sample app's dependencies and ran node --test bench/test/*.test.mjs: all 22 tests passed in about 54 s. Fifteen confirm each hidden check fails the naive reference on the expected sub-checks and passes the safe one; seven use fake agents to test the runner (skills copied into .claude/skills and kept out of the diff, stopping at the usage limit and resuming). We did not run real Claude tasks.
What we found when we tried it:The published raw runs reproduce the README's numbers, and the hidden checks do separate the typical mistake from the safe version. How many mistakes your own agent avoids needs a paid run of your own.
How to get in
The skills and benchmark are at https://github.com/globalmpc/backendskills . A Chinese write-up by @lovethelife2333 brought it to Chinese readers.
Who it suits
Developers using Claude Code, Codex or Cursor on database-backed web backends, especially around login, payments, stock and schema changes.
Pitfalls
- The gains sit in a few tasks: 10 of 15 tasks had no critical mistake either way, so don't expect every change to improve.
- Haiku got one more payment-webhook mistake with the skills (1/5 → 2/5); the authors list it as not fixed.
- The benchmark uses PostgreSQL and one Express sample app; check results on your own stack.
- Cost and turns roughly double, and small changes get extra checks too.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas
- Scan a skill before installing, then read the flagged linesCodingSkillSpectoruvFree; static scan needs no model● Tested by usChina