do-work (Capture + Build)
A two-skill Claude Code system that separates idea capture from execution — one Claude captures ideas into a queue in real time, another works through that queue autonomously.
Before & after
Same prompt, same agent — the only difference is the skill.
That's an interesting workflow concept. Setting up autonomous builds with idea capture would involve a few pieces—you'd need a way to queue ideas, then have a separate process that works through that queue and builds things automatically. There are different ways to approach this depending on how structured you want it. You could use a folder-based system, a database, or a message queue, each with different trade-offs. And the build agent would need to know when to plan versus execute, how to handle errors, and when to commit. It's definitely possible to set up, but there are several design decisions to make first. What's your technical comfort level, and how often do you generate ideas?
DO-WORK TWO-TERMINAL SYSTEM: Ready to deploy
ARCHITECTURE:
Terminal 1 (Capture): You type ideas in real time
→ Skill captures to shared queue folder (JSON files with timestamp)
→ Format: title, description, complexity (simple/medium/hard), priority (1-5)
Terminal 2 (Build): Autonomous loop running continuously
→ Detects new files in queue folder
→ Runs planning (for hard tasks only)
→ Spins up git worktree per task
→ Commits after completion
→ Moves completed file to archive folder
QUEUE FILE EXAMPLE:
{"title": "Fix navbar hover bug", "description": "Links don't show on mobile", "complexity": "simple", "priority": 3}
BUILD AGENT DECISION TREE:
→ Complexity=simple? Skip planning, execute immediately
→ Complexity=medium? Run 2min planning loop, then execute
→ Complexity=hard? Full planning + sub-agent delegation + testing before commit
→ Error detected? Pause task, log to errors.txt, notify user
COMMIT PATTERN:
Each task creates one clean commit with title from queue file. Message format: "Fix navbar hover bug" (Cherry-pick most relevant work from build loop output)
SETUP TIME: 30 minutes
QUEUE LOCATION: ./work-queue/
ARCHIVE LOCATION: ./work-complete/
RECOMMENDED: Test with 5 small tasks first before production ideasAbout this skill
name: do-work description: Use when A two-skill Claude Code system that separates idea capture from execution — one Claude captures ideas into a queue in real time, another works through that queue autonomously.
do-work (Capture + Build)
do-work is a two-terminal workflow built entirely through conversation with Claude Code, no hand-written code. One skill (capture) takes ideas the moment you have them and drops them into a shared folder-based queue. The other skill (build) runs continuously, pulling from that queue, planning when needed, spinning up sub-agents to build and test, and committing automatically — looping until the queue is empty.
What you get
- Two markdown skill files (capture + build) plus the folder-based queue pattern connecting them, from the open-source repo.
Customize your output
- What counts as a task in the queue
- when the build agent should plan vs. just execute
- commit/test conventions.
Example output
A live two-terminal setup where ideas typed in one window are automatically built, tested, and committed by the other.
Best for
Developers and engineers using Claude Code, Cursor, or similar AI coding assistants who want an autonomous build loop.
SKILL.md preview
---
name: do-work
description: Use this skill when idea capture needs to be separated from execution, with one workflow queuing ideas in real time and another working through that queue autonomously.
version: 1.0.0
category: Developer Tools / Workflow
author: AgentVolt
license: proprietary
tags:
- developer-tools
- workflow
- standard
---
# do-work (Capture + Build)
Splits work into two cooperating halves: a capture flow queuing ideas the instant they occur, and a build flow working the queue continuously.
## When to use
… (sign up to view the full skill)