A task queue skill for agentic coding tools. Capture requests fast, process them later.
npx skills add bladnman/do-workThis skill gives you a two-phase workflow:
-
Capture: Throw ideas, bugs, and feature requests at your assistant as they come up. Each one becomes a structured request file in
do-work/. -
Process: When you're ready, tell the assistant to work. It picks up pending requests one by one, triages complexity, and builds until the queue is empty.
The idea: separate thinking of things from doing things. Capture is instant. Processing is thorough.
Add a request:
do work add dark mode to the settings page
Creates do-work/REQ-001-dark-mode.md
Add multiple at once:
do work the search is slow, also add an export button, and fix the header alignment
Creates three separate request files.
Process the queue:
do work run
Starts the work loop. The assistant triages each request by complexity:
- Simple (config changes, small fixes) → straight to implementation
- Medium (clear goal, unknown location) → explore codebase first
- Complex (new features, architectural) → plan, explore, then build
Each completed request gets archived with its implementation notes and a git commit.
do-work/
├── REQ-018-pending.md # Queue (pending requests)
├── REQ-019-pending.md
├── user-requests/ # Verbatim input + assets per user request
│ └── UR-003/
│ ├── input.md # Original user input (source of truth)
│ └── assets/
├── working/ # Currently being processed
│ └── REQ-020-in-progress.md
└── archive/ # Completed work (self-contained units)
├── UR-001/ # UR folder with its completed REQs inside
│ ├── input.md
│ └── REQ-013-done.md
└── REQ-010-legacy.md # Legacy REQs archive directly
Every do invocation creates a User Request (UR) folder preserving the verbatim input. REQ files in the queue reference their UR. When all REQs from a UR are completed, the UR folder moves to archive as a self-contained unit.
Legacy REQs (created before the UR system) work the same as before — they archive directly without a UR folder.
This skill assumes your tool supports:
- File editing and shell access
- Optional subagent or multi-agent workflows (Plan, Explore, Build)
- Git integration for per-request commits (optional)
It was originally written for Claude Code and should work with other tools that provide similar capabilities. If your tool does not support subagents, run the Plan, Explore, and Implementation phases sequentially in the same session.
Invoked when you provide descriptive content. Optimized for speed:
- Minimal questions - capture what was said, don't interrogate
- Handles simple one-liners and complex multi-feature specs
- Always creates a UR folder preserving the full verbatim input
- Checks for duplicates against existing requests
See actions/do.md for the full capture logic.
Invoked when you say "run", "go", "start", or just confirm the prompt. Runs the build loop:
- Triages each request to determine the right amount of planning
- Spawns specialized agents only when needed
- Archives completed work with implementation notes
- Creates atomic git commits per request
See actions/work.md for the full processing logic.
Invoked when you say "verify", "check", "evaluate", or "review requests". Quality gate:
- Reads the original user input from the UR folder
- Compares against extracted REQ files for completeness
- Scores coverage, UX detail capture, intent signal preservation
- Optionally fixes identified gaps
See actions/verify.md for the full evaluation logic.
Invoked when you say "cleanup", "tidy", or "consolidate". Also runs automatically at the end of every work loop. Keeps the archive organized:
- Closes UR folders in
user-requests/when all their REQs are complete - Moves loose REQ files from
archive/root into their UR folders - Moves legacy REQs (no UR reference) into
archive/legacy/ - Fixes misplaced folders (e.g.,
archive/user-requests/UR-NNN→archive/UR-NNN)
See actions/cleanup.md for the full consolidation logic.
MIT
