Stats
Links
Categories
harness-engineering
Lifecycle orchestrator for AI-native software development.
A top-level Claude Code skill that coordinates the entire software development lifecycle — from requirements capture through release — by applying harness theory (compact machine state, structured artifacts, deterministic transitions) at the phase level instead of the task level.
English · 中文
What this is
harness-engineering is not a replacement for harness-plan. It is one layer up:
harness-engineering ← lifecycle orchestrator (this project)
│
└── invokes harness-plan as the implementation-phase executor
│
└── harness-plan drives feature-level coding
The 7 phases (locked)
discovery → design → architecture → implementation → test → release → ops
│ ▲
│ │
└────────────────┘
(parallel where upstream allows)
Each phase owns a schema-validated JSON artifact. Transitions are enforced by scripts, not ceremony. Failed validation blocks advance. Stale upstream propagates downstream.
Dependency: REQUIRES harness-plan skill
This skill delegates implementation phase to the
harness-plan
skill. Without harness-plan, phases 1-3 and 5-7 work but phase 4 (implementation)
will block.
Why not merge the two skills?
- harness-plan is useful on its own for pure coding tasks
- merging would force all harness-plan users to load engineering's 3500-token lifecycle body on every invocation
- "Compose, do not embed" is principle #8
Where harness-plan is expected:
~/.claude/skills/harness-plan/SKILL.md, or- any installed plugin providing
harness-plan/skills/harness-plan/SKILL.md
Install
Option A: manual (repo clone)
git clone https://github.com/suntao2yl/claude-skill-engineering.git && cd claude-skill-engineering
# install engineering only (checks for harness-plan, warns if missing):
./install.sh
# install engineering + harness-plan from a local source:
./install.sh --with-harness-plan /path/to/harness-plan-skill
# custom skills directory:
./install.sh --prefix /custom/skills/path
Option B: Claude Code skill system
If you install engineering via Claude's skill manager/marketplace,
install harness-plan first via the same mechanism. Engineering's
SKILL.md includes a preflight check that verifies harness-plan is present
before running init.
Quick start
# after installing both skills:
# in Claude Code:
/harness-engineering init "Build a Python CLI that counts lines in a file, emit JSON, ship v0.1"
# Claude will:
# 1. create .engineering/ with 7 phase subdirs + seed REQ-001
# 2. auto-drive: fill discovery → advance → pause at risk gate → await --confirm
# 3. fill design + architecture in parallel → advance each → pause at arch gate
# 4. enter implementation → delegate to harness-plan → drive features to done
# 5. advance implementation (live-executes verification commands)
# 6. fill test-report → advance (live-executes test commands)
# 7. create git tag → advance release → pause at release gate
# 8. fill ops metrics → terminal
The user only confirms at risk gates (discovery / architecture / release approval). Everything else auto-advances.
The 7 phases
| # | Phase | Unit | Artifact | Validation |
|---|---|---|---|---|
| 1 | discovery | problem statement | requirements.json | users ≥3 chars, metrics ≥5 chars, statement ≥20 chars |
| 2 | design | design spec | design-spec.json | each flow has ≥2 non-empty steps, components specs ≥10 chars |
| 3 | architecture | ADR + stack | stack.json + adrs/ADR-NNN.json | ≥1 ADR file, declarations match files |
| 4 | implementation | harness-plan campaign | campaign-ref.json | harness-plan done==total, live-execute each verification.command |
| 5 | test | test plan | test-report.json | ≥1 pass, 0 fail, live-execute plan commands |
| 6 | release | release candidate | release-checklist.json | git tag is valid ref, no pending checklist items |
| 7 | ops | metrics/incidents | metrics.json + incidents/ + postmortems/ | minimal |
Minimal mode (--mode minimal) skips phases 1/2/3/6/7, keeping only
implementation + test — for bug fixes and small features.