A Claude plugin that turns AI-assisted coding from a free-for-all into a structured, repeatable workflow. Each session has a clear start, a plan, and a clean save — so you always know where you are and what happens next.
Built for vibe coders and non-programmers who work with Claude but want more control over the process.
[optional: dev-intake] → dev-0-start → dev-1-research → dev-2-plan → dev-3-plan-review → dev-4-implement → dev-5-review → dev-6-end
↑ |
└──────────────────────────────────────── next session ────────────────────────────────────────────────────────────────┘
| Step | Command | What it does |
|---|---|---|
| pre (optional) | /dev-intake |
Turn a vague idea or bug report into a proper spec |
| 0 | /dev-0-start |
Load game — brief yourself on where you left off |
| 1 | /dev-1-research |
Analyze the codebase before writing any code |
| 2 | /dev-2-plan |
Create a phased implementation plan |
| 3 (optional) | /dev-3-plan-review |
Critical review of the plan to catch problems early |
| 4 | /dev-4-implement |
Execute the plan phase by phase, with checkpoints |
| 5 | /dev-5-review |
Code review before merge or deploy |
| 6 | /dev-6-end |
Save game — commit, push to GitHub, update tracker |
Minimum session flow: /dev-0-start → [work] → /dev-6-end
Recommended flow for vibe coders: /dev-0-start → /dev-intake → /dev-1-research → /dev-2-plan → /dev-4-implement → /dev-6-end
When you ask an AI to just "fix this" or "add that feature", several things go wrong:
- The AI jumps straight into code without understanding what's connected
- Mid-implementation it starts changing things you didn't ask it to change
- When the context window fills up, you lose track of where you are
- Next session you have no idea what was done and what wasn't
This plugin fixes all of that. Each step has a clear purpose, a stop point, and saves its output to a .agent/ folder in your project — so the next session picks up exactly where this one ended.
Trigger: "something's broken", "I want the app to do X", "this is slow"
You describe the problem in plain language. The skill looks at your code to find the relevant files, asks you a few focused questions, and produces a structured specification. No technical knowledge needed.
Output: .agent/specs/bug-YYYY-MM-DD-slug.md
Trigger: start of every session
Reads your project context, checks .agent/ for previous plans and notes, looks at recent git commits, and checks your task tracker (Linear). Gives you a 2-minute brief: what was done, what's in progress, and three options for what to tackle today.
Output: interactive briefing + your choice of direction
Trigger: "analyze this", "where is the logic for X?", "what's connected to Y?"
Runs parallel sub-agents across your codebase: one maps the direct area of the problem, another traces dependencies and data flow, a third checks existing tests. All results are combined into a research report.
Why: jumping straight to coding without a map causes bugs in unexpected places.
Output: .agent/research/problem-name-YYYY-MM-DD.md
Trigger: "plan this", "how should I implement X?"
Creates a detailed implementation plan: which files to touch, what exactly to change in each, phased tasks with verification steps, and a risk assessment. Also produces a 10-line brief (the TL;DR version).
Why: changing a plan costs zero. Changing code after implementation costs hours.
Output: .agent/plans/name-plan.md + name-brief.md
Trigger: "check the plan", "is this approach right?"
A dedicated devil's advocate pass over the plan: checks for missing files, wrong assumptions, hidden complexity, poor test strategy, and violations of your project's conventions. Gives a green / yellow / red verdict.
Output: inline review report with actionable suggestions
Trigger: "implement according to plan", "start coding"
Follows the plan strictly, one phase at a time. After each phase: announces what was changed, runs tests and type checks, reports the result, and waits for your confirmation before moving on. If it finds something broken outside the current scope — it tells you and leaves it alone.
Also monitors context window size and tells you when to clear it — the plan files ensure continuity.
Output: code changes + .agent/implementation/name-state.md
Trigger: "review the code", "check before PR"
Three parallel sub-agents review the changes: one checks conformance with the original plan, one checks code quality and security, one checks test coverage. Results are triaged with you — critical issues must be fixed, warnings are your call.
Output: inline review report with file:line references
Trigger: "we're done", "save session", "commit and push"
Walks you through closing the session: summarizes what was done, proposes a commit message (you approve or edit), pushes to GitHub, optionally merges to main, and updates your Linear issues with the new status. Nothing happens without your confirmation.
Output: .agent/session-notes.md updated + git commit + push
All skills write to .agent/ in your project root:
.agent/
├── specs/ ← from /dev-intake
├── research/ ← from /dev-1-research
├── plans/ ← from /dev-2-plan
├── implementation/ ← from /dev-4-implement
└── session-notes.md
Add .agent/ to your .gitignore or commit it — your choice.
- Download the
.pluginfile from Releases - Open Claude Desktop → Settings → Plugins → Install from file
- Select the downloaded
.pluginfile - Done — the
/dev-*commands are now available in any Cowork session
Requires: Claude Desktop with Cowork mode enabled.
The plugin auto-detects project context from CLAUDE.md, GEMINI.md, or README.md in your project root. Works with any stack — the research and planning steps adapt to whatever they find.
For best results, maintain a CLAUDE.md in your project with:
- Stack and framework
- Key directories
- Test commands
- Conventions and guardrails (e.g. "never use
anyin TypeScript")
MIT