Project Workflow

project-workflow v3

**Spec-driven feature development blueprint for AI-assisted coding** — methodology docs + starter template + runtime adapters.

v3 is docs-first and adds dual Claude/Codex plugin packaging plus spec lifecycle management.

---

Why v3 exists

AI-assisted coding has three Tier 1 engineering problems. Every part of v3 — the 5-phase workflow, 4 pillars, skills, template, and spec lifecycle — exists to address these three:

• Verification — AI generates code 10x faster than humans can validate. Boris Cherny (Claude Code lead): "The most important thing is to give Claude a way to verify."
• Context-as-RAM — Context windows behave like RAM, not storage. Long sessions drift, token costs scale, attention degrades past 200-line files.
• Drift — Without long-term memory, AI conventions drift across time (same code reviewed twice gives different verdicts), modules (A vs B style mismatch), and project lifecycle (month 1 vs month 6 codebase).

Cross-layer inconsistency (frontend/backend/DB contracts) is treated as a stack-specific tactic (`docs/workflow.md §8.1`), not a top-level proposition — community evidence frames it as a general architecture problem, not AI-specific.

Full framing with sources: [`docs/workflow.md §0.1`](docs/workflow.md).

---

Project goal

project-workflow helps real projects move from **chat-driven AI coding** to **spec-driven, reviewable, maintainable AI-assisted development**.

Personal and team development use the same per-change workflow. A team does not need a project-workflow collaboration layer: each contributor applies the same DIRECT/LIGHT/FULL classification, verification, and archive rules to their own change before submitting it.

v3 separates **methodology core** from **runtime adapters**:

• Core: AGENTS.md, docs/specs/(域现状), docs/specs/changes/(tracked changes), conditional ADR, compact delivery receipt, L1/L2/L3 review model, canonical workflow actions in docs/actions/, and canonical re