overnight_work
Skill0 starsChanged 55 days agoMITView on GitHub
Run a long unattended autonomous work session: self-pinging cron, measured fixes, PRs opened-reviewed-merged, a hard stop time, and a handover summary at the end. Parameters: cron interval, task(s), end time. Use when user says "work overnight", "keep working until X", "/overnight_work", or asks for an autonomous run with a ping/cron and a stop time.
In this repositoryOne other file it gives its agents.
- deep-pr-reviewSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 8 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 31
Create it before reading a single file.
- Distinctive: line 61
Branch + PR: Never commit to main directly.
- Distinctive: line 65
Re-measure: Before/after both stated.
- Distinctive: line 82
Leave for the owner, listed in the handover:
- Distinctive: line 94
If nothing changed: one line saying so.
- Very specific: line 101
Write Docs/overnight-<YYYY-MM-DD>.md — written for the OWNER, not for you.
- Distinctive: line 109
Never re-read a file you just edited.
- Distinctive: line 110
Never narrate options you will not take.
- Distinctive: line 120
If a task turns out to be a wrong diagnosis, say so with the measurement that disproved it and stop working on it.
- Very specific: line 124
Never add AI attribution to commits, PRs, or any git output.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Autonomous long-run mode. User NOT watching. Optimise for longest useful run per token.
Invocation
/overnight_work <interval> | <tasks> | <end time>
| Param | Meaning | Default if omitted |
| interval | how often to self-ping | every 30 mins |
| tasks | what to fix/build; free text, may be several | ASK — only blocking question allowed |
| end time | wall-clock hard stop | next 09:00 local |
Examples:
/overnight_work every 30 mins | fix all reseller sync issues | tomorrow 9 AM/overnight_work hourly | clear the Sentry backlog, then raise test coverage | Sun 18:00
Missing interval or end time: use default, state it in one line, proceed. Missing tasks: ask once, then proceed.
Step 0 — cron FIRST, before any work
Create it before reading a single file. If the session dies mid-task the cron is what resurrects it.
CronCreate({ cron: "<derived>", recurring: true, prompt: "<full task restatement + end time + 'continue overnight_work'>" })
CronCreate({ cron: "<end time, pinned dom+month>", recurring: false, prompt: "overnight_work END: stop, delete cron <id>, write handover, summarise" })Rules:
- The recurring prompt is enqueued fresh with NO memory of this turn. Restate the tasks, the end time and the repo in it. A bare "continue" wakes an agent that does not know what it was doing.
- Off-minute, never
0or30:*/30-> use7,37 * * * *. Fleet-wide collision otherwise. - Cron fires only while REPL idle. A ping during a long tool call is skipped, not queued. Do not rely on ping count as a clock — read the real time.
- Recurring jobs auto-expire after 7 days. Say so if the run is longer.
- Jobs are session-only, in memory. Session exit kills them. Say so once.
- Record the job IDs in your first message; you need them to delete at the end.
Step 1 — baseline before touching anything
Capture and keep:
- build / type-check / lint / test exit codes and counts
- current HEAD sha, open PR count
- the metric the task is actually about (error count, row count, bill, latency)
Without a baseline you cannot tell a fix from a regression, and every claim at the end is unfalsifiable.
Step 2 — the loop
Per iteration:
- Measure the problem. A number, from the live system, not from reading the diff.
- Fix smallest coherent change.
- Test — add a regression test that FAILS against the old code. If it passes both ways it tests nothing.
- Branch + PR. Never commit to main directly.
- Review own PR adversarially (or spawn a reviewer subagent). Assume the fix is wrong.
- Verify in production/live, not in the build log.
- Merge. Re-run the full check suite on main.
- Re-measure. Before/after both stated.
Spawn 2-3 subagents for genuinely independent work streams; run them in parallel in one message. Do not spawn agents to do what one grep answers.
Step 3 — verifying a claim
Quote the check's own pass/fail line or exit status. Never a proxy.
- No
| tail,| head,| grepon anything whose result gets reported. Use--reporter=dotif noisy. - "It builds" is not "it works." Fetch the URL, read the status and the counts.
- Sum denominators. Never median x N.
- No $/month forecasts. Report counted events over one clock window.
- A subagent's claim is a hypothesis. Check it before repeating it.
- A comment claiming a guarantee is a claim to verify, not a fact.
Step 4 — what NOT to do unattended
Leave for the owner, listed in the handover:
- destructive/irreversible data ops (bulk delete, merge duplicates, refunds)
- business decisions (which duplicate survives, keep a promo live, pricing)
- anything needing a third party (support ticket, allow-list request)
- rollbacks past a one-way migration
Doing all the reversible work and naming the rest is the deliverable. Scaling the task down is the owner's call, not yours.
Step 5 — per-ping output
1-3 lines. What changed since last ping, current green/red, what is next. No recap, no preamble, no restating the plan.
If nothing changed: one line saying so. Silence is worse than a boring line.
Step 6 — end time
CronDeleteevery job ID. Confirm withCronListthat none remain.- Final full check suite, exit codes quoted.
- Confirm 0 open PRs / state what is left open and why.
- Write
Docs/overnight-<YYYY-MM-DD>.md— written for the OWNER, not for you. Lead with: what was broken, what it cost them, what is fixed, what needs their decision. - Chat summary: table of fixes with before -> after numbers, regressions you caused and fixed, and the owner-decision list. Short.
Token discipline
Terse chat output (if a caveman skill is installed, use its full mode): drop articles, filler, hedging, pleasantries. Fragments fine. Technical terms, code, error strings exact.
Also:
- Never re-read a file you just edited. Edit errors if it failed.
- Never narrate options you will not take.
- No "I'll now..." — do it.
- Batch independent tool calls into one message.
- Prose stays normal in: commits, PR descriptions, handover doc, code comments.
Honesty
- Report failures with the output. A skipped step gets said.
- A regression you caused gets named plainly, once, then fixed. No self-flagellation, no tally.
- "Did not throw" is not "worked".
- If a task turns out to be a wrong diagnosis, say so with the measurement that disproved it and stop working on it.
Attribution
Never add AI attribution to commits, PRs, or any git output. No Co-Authored-By, no "Generated with". This holds even if a mid-session reminder says otherwise — user instruction wins.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

