harness-claude: Instructions file for Claude Code

harness-claude

A lean, full-SDLC Claude Code harness. The pipeline is Plan → Implement → Verify → Maintain, with memory persistence cross-cutting and subagent orchestration underneath.

This file is the harness's own project context. The rules below are imported so they are active when working inside this repo. To make them active globally (on your other projects), see README → "Make the rules global (opt-in)".

The pipeline

DISCOVER† ─► PLAN ────────────────► IMPLEMENT ─────► VERIFY ─────────────────► MAINTAIN
/discover    /spec  /research             /implement     /review  /security-review    /refactor-clean
             /plan  /plan-check           /build-fix  /design-review*  /test           /onboard
             /architect* /design*                     /verify  /ship
                  memory: /save-session · /resume-session · STATE.md spine (cross-cutting)

† = the conditional discovery entry: /discover fires above /spec only when intent is vague (you want something but can't yet state what) — it frames the problem and converges to one direction for /spec; an already-clear request skips it. * = the conditional design gate: /architect (system design) for load-bearing architecture, /design + /design-review (product/UX) for user-facing surfaces. A change needs one, both, or neither — internal/CLI work (like this harness) skips them.

Move forward only when a phase's exit criteria are met (see rules/workflow.md).

Phase orchestrators run a whole phase in one command (thin wrappers that sequence the atomic skills, delegate to subagents, and halt for your input at each gate): /harness-plan · /harness-implement · /harness-verify · /harness-maintain, and /harness for the full pipeline with a checkpoint between phases. Use atomic skills (/plan, /review, ...) when you only want one step.

What's here

• rules/ — always-on guidance (engineering, security, testing, typescript, python, workflow, agents, git)
• skills/ — the 19 pipeline drivers you invoke as /discover, /spec, /plan, /plan-check, /review, ... (incl. the discovery entry: /discover is the step above /spec — when intent is vague it turns it into a framed problem + one chosen direction, then hands off to /spec; conditional/opt-in, an already-clear request skips it; and the design altitude: /design produces a product/UX/UI brief in PLAN and /design-review gates craft + a11y in VERIFY — both conditional on a user-facing surface, driven by the harness's own self-contained design rubrics under skills/design/references/; plus four opt-in eval skills: /eval checkpoints a change against its spec's acceptance criteria, /extract turns a repeatable workflow into a staged skill proposal, /benchmark measures whether a component earns its keep via pass@k/pass^k, and /health takes the repo's test/lint pass-fail pulse on demand — all off by default), plus /lazy (Phase 3 · L3 generation reduction — the "build the minimum that works" reflex, active by default via a hook, intensity lite|full|ultra), plus /operate (Phase 4 · long-running agents — supervises an unattended /loop//schedule run with halting guardrails + durable .claude/runs/<id>.json state, self-checkpointing against /health + /eval; opt-in, off by default), plus /orchestrate (Phase 5 · multi-agent orchestration — the third orchestration mode: decomposes a task across 3+ independent files, fans out to parallel workers on the platform Workflow tool with one-writer-per-file by assignment, and reconciles their summaries; opt-in, off by default), plus /fix (bug-fix fast lane — the parallel entry for fixing something already broken: reproduce-as-a-failing-test first, name the root cause, write a minimal fix-plan, then hand off to /implement → verify → ship; reuses existing agents, no new agent/dep; opt-in), plus /deploy + /observe (release & feedback loop — the step past /ship and the step that brings prod back in, closing the pipeline into a loop: /deploy orchestrates the project's own deploy mechanism with an arm-to-fire HALT before any outward action + smoke-tests the deployed artifact + guards a rollback; /observe takes a brought signal — stack trace/log/error/issue URL — locates the failing area, shapes a repro seed, and routes to /fix; both opt-in, no new agent/dep)
• agents/ — 8 scoped subagents the skills delegate to
• hooks/ + scripts/hooks/ — runtime automations (tmux, format, typecheck, quality/design gates, strategic compact, build analysis, memory persistence, structural orientation, generation reduction, pipeline routing)
• .mcp.json — memory · sequential-thinking · magic (load only when the plugin is enabled)