A plain-text, human-in-the-loop workflow for small-to-medium software projects. One agent implements, another reviews, and the human coordinator stays involved at every review gate and phase boundary. No tools, no frameworks — just a protocol in Markdown that any LLM coding agent can read and follow.

Coding agents are powerful, but without explicit collaboration rules they wander, self-approve, expand scope, and leave inconsistent handoffs. This workflow gives agents a shared contract: roles, review gates, sign-off rules, and a live handoff surface (docs/CURRENT.md) for active tasks.

Bootstrap workflow         ← fresh repo or existing/forked project
       │
       ▼
Planning brainstorm        ← human + agents brainstorm together
       │
       ▼
Phase N spec + taskboard   ← human approves scope before implementation
       │
       ▼
┌──────────────────────┐
│  Main Developer      │   implements task → writes handoff → sets status to "review"
│  (agent session 1)   │
└──────┬───────────────┘
       │  docs/CURRENT.md (live state)
       ▼
┌──────────────────────┐
│  Reviewer            │   reviews code → writes findings → signs off or requests fixes
│  (agent session 2)   │
└──────┬───────────────┘
       │
       ▼
   Human as coordinator    looped in at every review gate and phase boundary

Either install from plugin marketplace:

/plugin marketplace add jameswei/human-in-loop-workflow
/plugin install human-in-loop-workflow@human-in-loop-workflow

or simply fetch single skill file:

mkdir -p ~/.claude/skills/human-in-loop-workflow && \
  curl -o ~/.claude/skills/human-in-loop-workflow/SKILL.md \
  https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/SKILL.md

Then prompt like following:

Use the human-in-loop-workflow skill to bootstrap this repository.
Only install or adapt workflow docs. Do not change product code.

$skill-installer install https://github.com/jameswei/human-in-loop-workflow/tree/{tag}

Then prompt like following:

Use the human-in-loop-workflow skill to bootstrap this repository.
Only install or adapt workflow docs. Do not change product code.

Copy the protocol docs into your project, then any agent that reads project files will pick up AGENTS.md and follow the protocol:

curl -o AGENTS.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/AGENTS.md
mkdir -p docs/phases
curl -o docs/agent-guidelines.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/docs/agent-guidelines.md
curl -o docs/current-task-template.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/docs/current-task-template.md
curl -o docs/phase-spec-template.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/docs/phase-spec-template.md
curl -o docs/taskboard-template.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/docs/taskboard-template.md
curl -o docs/phases/README.md https://raw.githubusercontent.com/jameswei/human-in-loop-workflow/refs/tags/{tag}/docs/phases/README.md

Do I need two different LLM models? No. The two agents can be the same model running in separate sessions. What matters is separate sessions — the Reviewer must not share context with the Developer, otherwise the review gate is meaningless.

Can I use more than two agents? Yes. The protocol defines four roles (Main Developer, Architecture Reviewer, Code Reviewer, Test Verifier). Two agents splitting these roles is typical for small projects; larger projects can assign one per role.

Is this overkill for a solo project? For a one-file script, yes. For anything with multiple phases and the risk of agents going off the rails — the overhead is about 30 minutes of initial setup and pays back in fewer reverted commits and clearer handoffs.