Skip to content
← all systems

Claude Workflow Framework

8-phase spec-driven SDLC framework that AI assistants follow on every project - with the rule that AI never commits code.

8 phases · 6 reusable templates

role
Author
stack
Claude Code · workflow framework · templates
status
oss

The Claude Workflow Framework is a ruleset that AI coding assistants read and follow on frontend projects - delivery-agnostic across Claude, GPT, and others. It runs an 8-phase cycle from session initialization (Phase 0) through discovery, requirements, task planning, implementation, testing, documentation, and review (Phase 7), and ships six reusable templates the assistant fills into the project's ./docs/ folder. It enforces a knowledge-base query protocol - check past decisions and a lessons-learned log before implementing, so a failed approach is not repeated - under one governing rule that makes the rest trustworthy: the AI proposes changes but never runs git commit, so history stays under human control.

// 01 - PROBLEM

AI-assisted development without structure produces inconsistent patterns, undocumented decisions, repeated mistakes and incomplete implementations. This framework gives the assistant a disciplined process to follow instead of improvising.

// 02 - APPROACH

  • An 8-phase cycle: session init, discovery, requirements, planning, implementation, testing, documentation, review.
  • 6 reusable templates: ADR, lessons learned, testing strategy, component docs, project overview, manual test checklist.
  • Lessons-learned tracking: failed approaches get documented so they are not repeated.

// 03 - ARCHITECTURE

▸ 8-phase cycle · session-init to review

  1. Init

    Detect project type, onboard, create ./docs, init CLAUDE.md

  2. Discovery

    Analyze structure, stack, and existing patterns

  3. Requirements

    Requirements checklist and clarification protocol

  4. Plan

    Break work into atomic, single-session tasks

  5. Implement

    Minimal changes; verify libraries; follow patterns

  6. Test

    Playwright if configured, else a manual checklist

  7. Document

    Component docs and ADRs for significant decisions

  8. Review

    Self-review and a mandatory task summary; the human commits

AI never commits code
The framework's top rule: the assistant never runs git commit, push, or add-for-commit. If asked, it declines and asks the human to review and commit. History and the review that catches accidental secrets stay under human control - the governance principle that makes delegating the rest safe.
Lessons-learned as a queryable knowledge base
Every failed approach is logged with why it failed and the better alternative, and a knowledge-base query protocol requires checking that log and past ADRs before implementing - so the assistant does not repeat a mistake an earlier session already made.
Testing that adapts to the project
A decision tree picks Playwright end-to-end tests when the project has them configured and falls back to a structured manual-test checklist the human runs when it does not. Testing is mandatory either way, not skipped when automation is absent.

// 04 - PRODUCTION-GRADE

  • Battle-tested across real client and personal projects
  • Delivery-agnostic - the same ruleset works with Claude, GPT, and other assistants
  • Six reusable doc templates: ADR, lessons-learned, testing strategy, component docs, project overview, manual test checklist
  • Documentation is a dedicated phase, not an afterthought
  • Playwright-or-manual testing decision tree built into the workflow
  • Markdown-lint and link-check CI on the framework docs themselves

// 05 - ARTIFACTS