build
A structured build workflow for Claude Code (and four portable skills for OpenCode and Codex). Plan before you build, review before you ship, verify before you claim.
Skills
| Skill | What it does |
|---|
/build | Orchestrates the full workflow: plan, review, implement, verify, architect review |
/build:impl-plan | Creates a detailed implementation plan by reading the codebase first |
/build:review-plan | Reviews a plan against its own evidence with severity-tagged findings |
/build:architect-review | Architecture review of completed work across 10 lenses with structured verdict |
/build:verify | Runs tests, build, type checks and reports actual evidence |
/build:eval | Runs test cases against build skills, grades outputs against assertions |
Every skill works standalone. Run /build:impl-plan add user authentication without the full pipeline. Or run /build add user authentication to get the complete workflow.
Install
Claude Code
claude plugin add jameshemson/build
OpenCode — copy the .opencode/ directory (preserving the leading dot) into your project so the final layout is <your-project>/.opencode/skills/<skill-name>/SKILL.md and <your-project>/.opencode/commands/<command-name>.md. OpenCode discovers skills from those paths. Once copied, the four portable skills are invocable as flat slash commands: /impl-plan, /review-plan, /verify, /architect-review. Each command thin-wraps the matching bundled skill.
Codex (two paths, either works):
Via Plugins UI / CLI:
codex plugin marketplace add jameshemson/build
codex plugin install build/build
Or via repo-local discovery: copy the .agents/ directory into your project so the final layout is <your-project>/.agents/skills/<skill-name>/SKILL.md. Codex picks it up automatically.
Four skills are available in OpenCode and Codex: impl-plan, review-plan, verify, and architect-review. The build orchestrator and eval runner are Claude Code only (they depend on sub-agent spawning and Task tools not available elsewhere).
Compatibility
| Skill | Claude Code | OpenCode | Codex |
|---|
build (orchestrator) | ✓ | — | — |
impl-plan | ✓ | ✓ | ✓ |
review-plan | ✓ | ✓ | ✓ |
verify | ✓ | ✓ | ✓ |
architect-review | ✓ | ✓ | ✓ |
eval | ✓ | — | — |
See HARNESSES.md for the full capability matrix and install story.
How it works (Claude Code)
/build drives a 5-phase cycle:
- Plan (Opus) - Read the codebase, choose a discovery level, create
REQ-*/D-* inventories, define Wave 0 validation, and emit an execution_manifest with wave, depends_on, files_modified, must_haves, verify, and done
- Review (Sonnet, forked context) - Adversarial senior engineer review: placeholder scan, workstream independence check, requirement/decision coverage, wave graph validation, test coverage mapping
- Implement - Parallel agents in isolated worktrees, routed by the manifest when available, with mid-reviews for complex changes. Agents report SCOPE_CHANGE to stop work against broken plans. Circuit breakers prevent runaway retries.
- Verify - Run tests, build, type checks, and plan-declared verification commands. Requirement coverage and missing
must_haves evidence are reported explicitly. No claims without fresh output.
- Architect Review (Opus, forked context) - 10-lens review: correctness, trade-offs, anti-patterns, consistency, non-functional, edge cases, overengineering, plan fidelity, weak-test audit, dependency audit
The orchestrator manages state, auto-continues between phases, and deploys model-appropriate agents. It resolves merge conflicts through a structured protocol. Give it a feature description; it builds it.
/build is file-backed. Each workflow writes durable artifacts under .build/plans/ so later phases and fresh agents do not depend on chat history alone:
{slug}-state.md - current phase, base_ref, task IDs, completed tasks, blockers, history{slug}-context.md - repo conventions, user constraints, discovered patterns, assumptions, out-of-scope notes{slug}-requirements.md - canonical requirements, decisions, assumptions, acceptance criteria, must_haves{slug}-plan.md - full implementation plan plus execution manifest{slug}-review.md - plan review findings and verdict{slug}-implementation-summary.md - completed waves, files changed, deviations, blockers{slug}-verify.md - command evidence, requirement coverage, verification verdict{slug}-architect-review.md - final architecture review findings and verdict
Skill prompts are intentionally kept compact. Detailed planning rules live in reference files, and npm test enforces hard line ceilings for the main skill prompts.
