Overnight Task Queue

Use this skill when work should be broken into a local queue and processed safely over time by Claude Code. The goal is not "let the agent do anything"; the goal is a bounded queue where each item is small, checked, and stopped if risk or ambiguity appears.

Trigger phrases

Russian:

• очередь задач
• ночная очередь
• пусть Claude Code работает пока я сплю
• запусти задачи на ночь
• автономная очередь
• обработай backlog по одному
• поставь очередь для Claude Code
• продолжай задачи без меня

English:

• overnight task queue
• run while I sleep
• autonomous task queue
• process backlog one by one
• unattended Claude Code run
• queue runner

Core rule

One queue item = one Claude Code invocation = one proof packet.

Never combine multiple queue items into one vague prompt. Never continue after an item fails unless the queue policy explicitly marks the next item safe and independent. Default behavior is to stop on failure or uncertainty.

Workflow

1. Check whether the current folder is a local project and whether claude is available in PATH.
2. Create or update:
   • .claude/queue/tasks.md
   • .claude/queue/logs/
   • .claude/queue/evidence/
   • skills/overnight-task-queue/scripts/claude_queue_runner.mjs
3. Turn the user's backlog into small checklist items in .claude/queue/tasks.md.
4. Add constraints to the top of the queue:
   • no purchases;
   • no production deploys without explicit permission;
   • no broad deletes;
   • no secret exposure;
   • stop if credentials, payment, destructive migration, or owner judgment is required.
5. Run a dry check first:

   node skills/overnight-task-queue/scripts/claude_queue_runner.mjs --dry-run --once

6. For real work, run with a time budget:

   node skills/overnight-task-queue/scripts/claude_queue_runner.mjs --hours 8 --max-tasks 20

7. In the morning, inspect:
   • .claude/queue/status.json
   • .claude/queue/logs/
   • .claude/queue/evidence/
   • git status

Queue format

Use markdown checkboxes. Keep each item concrete and verifiable.

# Claude Code task queue

## Policy

- Work one item at a time.
- Stop on secrets, payments, prod deploys, destructive actions, or unclear owner decisions.
- After each completed item, write evidence to `.claude/queue/evidence/`.
- Mark an item done only after checks pass.

## Tasks

- [ ] Update the README section about local setup and run the markdown linter.
- [ ] Add missing alt text to the landing page images and verify with a grep check.
- [ ] Review the docs folder for duplicate setup instructions and write a short cleanup report.

Good queue items

Good:

• "Update README.md install steps so they match package.json, then run the documented check command."
• "Find broken internal markdown links under docs/ and fix only the links that can be verified locally."
• "Create a first draft of docs/faq.md from existing project docs; do not invent product facts."

Bad:

• "Improve the whole project."
• "Make the site better."
• "Refactor everything."
• "Deploy when done."
• "Buy the domain and configure billing."

Stop conditions

Stop the queue and write a report when any of these happen:

• a task requires a secret, password, API key, payment, or account recovery;
• a task requires deleting many files or running a destructive migration;
• a task touches production deploy, DNS, billing, customer data, or legal wording;
• checks fail twice for the same reason;
• the project has unexpected unrelated dirty git changes;
• Claude cannot prove the acceptance criteria;
• the next task depends on an unresolved failed task.

Runner

Use the bundled script:

node skills/overnight-task-queue/scripts/claude_queue_runner.mjs --help

The script intentionally processes the first open checklist item only, waits for it to be marked done by Claude, then moves to the next item. If Claude exits successfully but the item is still open, the runner treats that as uncertainty and stops.

Proof packet

Each completed item should leave a small evidence file in .claude/queue/evidence/ with:

• task text;
• files changed;
• checks run;
• result;
• remaining risks;
• git commit hash if a commit was created.

Morning report

Summarize:

• completed items;
• stopped item, if any;
• reason for stop;
• files changed;
• checks that passed or failed;
• next recommended human decision.