skill
nightshift-skills
Skills for AI coding agents that work while you sleep. Each one benchmarked with and without the skill. Zero dependencies.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/Andrea07072000/nightshift-skills itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 26, 2026nightshift-skills
Skills for AI coding agents that work while you sleep — each one benchmarked with and without the skill.
Drop a folder into Claude Code (or any agent that reads SKILL.md) and it gets better at the boring, high-stakes
parts of autonomous work: checking it is safe to run unattended, and reporting what it did without inflating it.
No installs, no API keys, no network: every script uses only the Python standard library.
The skills
| Skill | What it does | Measured effect (see evals/BENCHMARK.md) |
|---|---|---|
| 3am-safety | Preflight before you leave an agent running overnight: secrets about to be committed, a stuck git lock, a disk about to fill, uncommitted work, detached HEAD, orphan processes, pushes that would publish. Clear GO / GO WITH WARNINGS / NO-GO in ~1 second. | Found a stale git index.lock the agent without the skill missed (it would have made every overnight commit fail): 11/11 vs 10/11 on two planted-problem repos. |
| honest-numbers | Checks a status report, PR description or investor update for claims it does not back up — 85% of what?, "fully tested" with no test run, "10x faster" with no benchmark — and rewrites them so every number can be checked. | Corrected report left with 0 unbacked claims vs 4 without the skill; the script finds 10/10 planted claims and flags none of the honest lines. |
More are on the way. Every skill here ships with its benchmark, or it does not ship.
Install (30 seconds)
git clone https://github.com/Andrea07072000/nightshift-skills.git
# for every project:
cp -r nightshift-skills/skills/3am-safety ~/.claude/skills/
# or only for one project:
cp -r nightshift-skills/skills/honest-numbers your-project/.claude/skills/
On Windows, copy the folders to %USERPROFILE%\.claude\skills\. Restart Claude Code and the skills are picked up
automatically: you do not need to call them by name — the description tells the agent when to use each one.
You can also run the scripts directly, no agent needed:
python skills/3am-safety/scripts/preflight.py path/to/repo
python skills/honest-numbers/scripts/check_report.py STATUS.md
Why "benchmarked with and without"?
Anyone can write a SKILL.md that looks useful. The question that matters is: does the agent do better with
it than without it? For every skill here, two independent agents get the same realistic task — one with the
skill, one without — on fixtures with known, planted problems, and a script grades both. The numbers, including
where the skill did not help, are in each skill's evals/BENCHMARK.md.
That also keeps the skills honest about their limits: a strong model without a skill is already good at a lot of this. The skills earn their place on consistency, on the checks models tend to skip, and on output you can verify.
Repository layout
skills/<name>/SKILL.md instructions + when to use (what the agent reads)
skills/<name>/scripts/ deterministic helpers (standard library only)
skills/<name>/evals/BENCHMARK.md with/without comparison, raw numbers, limits
tests/test_<name>.py tests for the scripts (pytest)
tools/check_skill.py the publication gate every skill must pass
Contributing
Ideas, bug reports and new skills are welcome — see CONTRIBUTING.md. The one hard rule: a new skill comes with a benchmark against the same task without it.
Also from this author
S.T.A.R. — reproducible, tested building blocks for space and flight software (time scales, leap seconds, verification).
License
Apache-2.0. See LICENSE.