Dev tooling
A token-efficient development workflow for Claude Code
The discipline of heavyweight AI dev pipelines — at a fraction of the cost.
Working with a coding agent tends to fail in one of two ways
No process
The agent skips tests, pushes to main, patches symptoms
And loses context between sessions. Nothing stops it from cutting corners, because nothing was asked of it.
Heavyweight process
Every task costs 250–600k tokens — even a typo fix
Mandatory researcher subagents, plan documents, per-cycle review passes. You pay feature-sized overhead for one-line changes.
The insight: process proportional to task size
The skillset sizes every task first and applies only the process that size needs. Same rigor where it matters — TDD, worktree isolation, review before PR, persistent project docs — none where it doesn't.
Trivial
typo, config value, one-liner
- Edit
- Verify
~2–5k tokens
Small
contained bugfix
- Write failing test
- Fix
- Run tests
~8–20k tokens
Feature
crosses architecture layers or systems
- Explore
- Plan
- Human approval
- TDD
- Review
- PR
~60–130k tokens
Heavyweight pipeline
not thisresearcher subagents, plan docs, per-cycle review — applied to every task, regardless of size
250–600k tokens
The fourth lane is what heavyweight AI dev pipelines charge for a typo fix. The skillset never routes there unless the task actually is that big. Explore subagents scale the same way — solo by default, parallel only per separate system or disjoint .docs area too large to read directly (cap ~4).
What's in the box
A ~55-line router
The /code4food-skillset:setup command injects it into the project's CLAUDE.md. It sizes each task silently and routes to the right process. That's the whole "framework."
Twelve core skills
Each under ~500 words: dev-workflow, grill-me (batched Socratic questions that pin down scope and success criteria before planning starts), TDD, debugging, verify, worktrees, finishing (review + checks + PR), project docs, session handoff, plus domain guards for auth, database migrations, and deploys. Detailed references load only on demand — you never pay for guidance you don't use.
One review subagent
Spawned at most once per feature. Findings are confidence-scored; only high-confidence issues (≥80/100) surface. No seven-agent review committees.
Mechanical safety, zero tokens
A pre-tool hook that physically blocks git commit/git push on protected branches. Works everywhere the plugin is enabled, zero per-project setup, zero context-token cost.
Engine skills
For Unity and Godot, shipped inside the plugin, description-gated — they activate only when the task at hand is engine work. No detection step, no separate install.
.docs/ convention
Small per-area project docs agents actually read — each session loads only the slice its task touches, so knowledge persists across sessions without dragging history into context. No word-count quota, just a density rule: keep it dense, split an area once it outgrows one file. A living known-issues.md is now a standard file in every project, and /code4food-skillset:setup bootstraps an initial .docs adoption pass for pre-existing projects, not just new ones. The same discipline now covers user-facing docs too — READMEs, onboarding guides — not just agent-facing ones.
Quality-of-life
A one-shot /commit command and an optional status bar (branch, model, cost, tokens, lines changed). Runs on Node with zero external dependencies — works on macOS, Linux, and Windows.
Install: two slash commands
/plugin marketplace add MJTeixeira/claude-plugins, then /plugin install code4food-skillset@code4food. Per project, run /code4food-skillset:setup once. MIT/open source; uninstalls like any plugin, never touches your global Claude config, safe next to other skillsets.
Built on the same skills
The Skillset is interactive — you drive. The Factory is autonomous — it drives.
Factory sessions run the same core skills described here, unattended, on a schedule. See how it works.