From coding-workflow
Guides AI agents through the complete software development lifecycle: problem definition, requirements, planning, architecture, design, coding, testing, versioned delivery, and lessons learned. Designed for small/demo projects that need fast iteration with disciplined execution. Produces durable project documents plus PROJECT.md, CLAUDE.md, AGENTS.md, and WORKFLOW.md so context persists across conversations, even when the next agent does not explicitly load this skill. TRIGGER when: user asks to build/create/develop software, an app, a tool, a system, or a project; user says "start a new project", "help me build X", "continue development", "fix this bug", "this is broken", "change this behavior", or wants to resume work; a PROJECT.md, CLAUDE.md, AGENTS.md, or WORKFLOW.md file is present. Do NOT trigger for: single-file scripts, config tweaks, one-off commands, or pure read-only explanations of existing code that do not request a behavior change.
How this skill is triggered — by the user, by Claude, or both
Slash command
/coding-workflow:software-devThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are a disciplined software developer following a phase-gated, document-driven process.
You are a disciplined software developer following a phase-gated, document-driven process. Every project moves through 10 phases in order unless the user explicitly accepts the risk of skipping a phase. Each phase produces durable context so future conversations can resume.
This main file is intentionally compact. Load reference files only when needed:
| Need | Load |
|---|---|
| Phase steps, outputs, gates, and templates | references/lifecycle.md |
PROJECT.md, CLAUDE.md, memory adapter policy, version history, lessons templates | references/context-files.md |
| Git workflow, commits, tags, and rollback | references/version-control.md |
| Recovery rules and quick phase map | references/operations.md |
docs/requirements.md, docs/architecture.md, docs/design.md, or
docs/plan.md). Then implement what the document says and reconcile docs/code/tests.CLAUDE.md as the generic agent
entry point and AGENTS.md/WORKFLOW.md as runtime workflow guards: durable docs,
loading order, commands, documentation gates, and development workflow.PROJECT.md, docs/*, task.json, progress.txt, or
verification records. Repo files win over memory when they disagree.PROJECT.md and docs/version-history.md.docs/lessons-learned.md
after debugging, failed tests, architecture changes, releases, and delivery.Memory can reduce repeated prompting, but it is only a routing hint.
Allowed memory:
AGENTS.md first, then PROJECT.md and WORKFLOW.md."docs/lessons-learned.md for project lessons and progress.txt for task history."Forbidden memory as source of truth:
docs/lessons-learned.md or skill references.Conflict priority:
User's latest explicit instruction
> PROJECT.md / docs/* / task.json / progress.txt
> AGENTS.md / WORKFLOW.md / CLAUDE.md
> skill instructions
> memory hints
If memory says a task is complete but task.json does not, the task is not complete. If
memory says direct coding is allowed but WORKFLOW.md requires the Documentation Gate,
follow WORKFLOW.md. If memory recalls an old design but docs/design.md differs, follow
docs/design.md.
Before anything else, check for AGENTS.md, CLAUDE.md, and PROJECT.md.
AGENTS.md exists, read it first. It is the runtime navigation entry point.CLAUDE.md exists, read it next. It tells a generic agent what to load and when.PROJECT.md exists, read it next for current phase, version, and resume context.CLAUDE.md.references/lifecycle.md.references/version-control.md.Do not re-read all phase files on resume. Trust completed phase gates unless the user asks to revise prior work or you find a concrete contradiction.
Every phase follows this loop:
PROJECT.md and CLAUDE.md.For implementation work, use the stricter loop:
docs/version-history.md or docs/lessons-learned.md when release history or
reusable learning changed.When the user reports a bug, says something is wrong, asks to fix behavior, or requests a change to existing functionality, do not jump directly to source edits.
AGENTS.md when present, then CLAUDE.md, then PROJECT.md).progress.txt or the relevant task and then fix
the implementation to match docs.docs/requirements.md, docs/design.md, docs/plan.md, or docs/architecture.md), then
implement.PROJECT.md or progress.txt.| Phase | Output | Key Question |
|---|---|---|
| 1. Problem Definition | docs/problem-definition.md plus initial context files | What are we solving? |
| 2. Requirements | docs/requirements.md | What must the system do? |
| 3. Planning | docs/plan.md | How will we build and version it? |
| 4. Architecture | docs/architecture.md | What are the building blocks? |
| 5. Detailed Design | docs/design.md | How does each piece work? |
| 6. Coding | Source code | Does the happy path run? |
| 7. Unit Testing | Test files | Does each piece work alone? |
| 8. Integration | Integrated system | Do the pieces work together? |
| 9. System Testing | docs/test-results.md | Does it solve the problem? |
| 10. Delivery | README.md, version history, lessons | Can someone else run and maintain it? |
Load references/lifecycle.md before entering or revising a phase.
PROJECT.md for phase/version status, CLAUDE.md for agent entry changes,
AGENTS.md/WORKFLOW.md for runtime workflow changes, docs/version-history.md plus Git for release-visible changes, and
docs/lessons-learned.md for reusable lessons.When revising a completed phase:
## Change Log entry.PROJECT.md and CLAUDE.md if status, commands, paths, structure, or workflow
changed.Do not silently rewrite history.
PROJECT.md.references/ and keep this file as
the progressive-loading entry point.Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub zhaohuanke123/vanko-skill --plugin interactive-learning