handoff-iteration-loop
Installation
SKILL.md
Handoff iteration loop
Multi-agent pattern inspired by Harness design for long-running application development (Anthropic): separate generation from evaluation, structured artefacts across iterations, sceptical external QA, contracts before heavy work, simplify the harness when the task no longer needs every role — and real dev-team shape: discovery / Q&A → goals → sprints → loop, with Git commits as the implementor goes.
Progressive disclosure — load only what you need
| If you are… | Read first | Then if needed |
|---|---|---|
| Spinning up the workflow | This file + workflow.md | discovery.md if the ask is vague; then role files |
| Discovery / customer goals | discovery.md | workflow.md §Phases |
| Redesigning the harness (adding/removing roles, changing contracts) | harness-principles.md + workflow.md | Role files only if prompts change |
Coordinator (main agent orchestrating Task / sub-agents) |
coordinator.md | workflow.md §Contracts, any single planner.md / reviewer.md for prompt patterns |
| Planner sub-agent | planner.md | workflow.md §Handoff filenames + §Sprint contract |
| Implementor sub-agent | implementor.md | workflow.md §Contracts, planner.md §Anti-patterns (mirror only outcomes) |
| Reviewer sub-agent | reviewer.md | workflow.md §Sprint contract + §Evidence |
Do not load all role files into one context unless the coordinator is authoring a full runbook in one shot.