dev-workflow
AI coding agents forget everything between sessions. This fixes that.
Context fills up. You restart. Twenty minutes re-explaining what you were building, what you decided, where you left off. Or:
/dev-resume
Three commands. That's it. /dev-plan once to create a plan. /dev-checkpoint when you stop. /dev-resume when you come back. Everything comes back — decisions, progress, blockers, the exact next step. Try it once.
The Problem
LLM performance degrades as context fills up. This isn't a theoretical concern — after ~200K tokens of accumulated conversation, tool outputs, and debugging tangents, responses get slower, details get missed, and earlier decisions get contradicted. Even with 1M context windows, more context means worse output.
Every developer using AI coding agents hits the same wall:
- Session 1: Great output. Claude is sharp, follows the plan, remembers everything.
- Session 1 (continued): Context filling up. Responses slower. Starts forgetting things you discussed 30 minutes ago.
- Session 2: Fresh start. But now you're the one who has to remember everything and re-explain it. Poorly.
The common workaround? Manually copy plans into files, paste fragments back into new sessions, hope you didn't forget anything important. It works. It's also tedious, error-prone, and doesn't scale past one feature.
dev-workflow automates this. It saves structured progress to disk — not conversation dumps, but the specific state needed to resume: what's done, what's next, what was decided, and what to watch out for. Each new session starts clean with minimal context and full continuity.
How It Works
Plan Build Checkpoint New Session Resume
───── ───── ────────── ─────────── ──────
/dev-plan → implement → /dev-checkpoint → restart → /dev-resume
│
▼
build again...
| Step | What you do | What happens |
|---|
| 1. Plan | /dev-plan | Generates a structured PRD in .dev/ with phases and gates |
| 2. Build | Implement | Work until context gets heavy |
| 3. Checkpoint | /dev-checkpoint | Saves progress, git state, decisions, next steps |
| 4. Restart | Close and reopen Claude | Fresh context window, clean slate |
| 5. Resume | /dev-resume | Loads ~2KB checkpoint, rebuilds context, picks up where you left off |
Repeat steps 2–5 until the feature is complete. Each session starts fresh with high-quality context.
Installation
Claude Code (Plugin)
/plugin marketplace add andreaserradev-gbj/dev-workflow
/plugin install dev-workflow
Updating & Troubleshooting
/plugin marketplace update dev-workflow
If the plugin doesn't load after updating, clear the cache and reinstall:
rm -rf ~/.claude/plugins/cache/dev-workflow
rm -rf ~/.claude/plugins/marketplaces/dev-workflow
Then re-run the install commands above.
Codex
Tell Codex:
Clone https://github.com/andreaserradev-gbj/dev-workflow.git and follow .codex/INSTALL.md from the local checkout.
Or see docs/README.codex.md for manual setup.
Gemini CLI
gemini skills install https://github.com/andreaserradev-gbj/dev-workflow.git --path plugins/dev-workflow/skills
Or see docs/README.gemini.md for alternatives.
Skills
/dev-plan — Plan a feature
Creates a structured PRD with phases, status markers, and gates. Three phases:
- Understand — gather requirements (or infer from inline arguments)
- Research — explore the codebase using parallel agents
- Write — produce
.dev/<feature-name>/00-master-plan.md
/dev-plan add OAuth login with Google and GitHub providers
/dev-plan refactor the database layer to use connection pooling
/dev-checkpoint — Save progress
