My AIsy Toolkit
I spent a while looking for a framework simple enough to actually use every day, but complete enough to cover the whole journey: Spec-Driven Development in the style of spec-kit, and worktree-based agentic-loop management in the style of Sandcastle. I never found one that did both while staying simple, so after a lot of trial and error across different development patterns, I ended up publishing my own. My AIsy Toolkit is a catalog of skills and subagents that install at the repository level and that I use every day.
If you've read my post on the 8 levels of AI, this toolkit covers everything from level 3 (Agent mode) and level 4 (CLI first) of Spec-Driven Development, up through level 5 (Subagents) and level 6 (Multiagents): a lead agent delegating verifiable waves of work to specialists on isolated branches.
1. Installation
I could have packaged this as a library with proper install packages, but since it's a public repo and I wanted maximum simplicity, I set it up so all you need to do is paste one prompt into your agent's conversation: it fetches and follows the setup instructions itself.
Paste this prompt into your CLI agent's chat
Fetch and follow the setup instructions at https://raw.githubusercontent.com/charlstown/my-aisy-toolkit/main/setup-ai.md
Compatible with the following agents:
It can be installed on other agents too, just tell it in the prompt to adapt the skills to that agent's own documentation.
2. How it works
The whole skill cycle is designed as a closed loop, and it starts with /constitution: a living document that stays with the product for its whole life and pins down what it is and how you're going to build it. Once that's in place, the loop starts turning: you specify the features you want to add, plan them, and implement them, one at a time. When a feature is done, /clean-feature realigns that living document with what you actually just built, and the whole thing starts over from there.
The workflow
My AIsy Toolkit
SDD + agentic loops, in one tool.
Diagram of the My AIsy Toolkit feature loop, arranged like an athletics track around the constitution. /specify-feature scopes the work against the project's constitution — a living set of product-spec, tech-spec and roadmap documents. /specify-feature hands off to /plan-feature, which breaks the work into eight tasks across two features, each in the @agent · task format and assigned to a subagent from the catalog. All eight tasks converge on /implement-feature, which runs the plan in its own isolated git worktree and feeds directly into /clean-feature, which reconciles the specs and closes the loop back to the constitution, ready for the next feature.
/constitution scaffolds the specs every skill in the loop reads from and writes back to.
/specify-feature scopes what you're building and why into a requirements file.
/plan-feature breaks the requirements into tasks, assigned to the right subagent.
/implement-feature builds the plan inside its own isolated git worktree.
/clean-feature updates the specs to match what shipped, and closes the loop.
2.1 Skill catalog
The first four bootstrap the loop (or regenerate a single piece of the living document on its own); the next five are the cycle that repeats for every feature:
| Phase | Skill | What it does | Example |
|---|---|---|---|
| 1 | /constitution |
Bootstraps product-spec.md, tech-spec.md, and roadmap.md |
/constitution |
| 1 | /product-spec |
Regenerates just the product-spec.md |
/product-spec |
| 1 | /tech-spec |
Regenerates just the tech-spec.md |
/tech-spec |
| 1 | /roadmap |
Recalculates the roadmap.md phases |
/roadmap |
| 2 | /specify-feature |
Scopes the feature: the what and the why | /specify-feature "add dark mode to the dashboard" |
| — | /clarify-feature |
Closes decision gaps before anything gets planned | /clarify-feature specs/003-dark-mode/requirements.md |
| 3 | /plan-feature |
Breaks the feature down into a plan.md |
/plan-feature specs/003-dark-mode/requirements.md |
| 4 | /implement-feature |
Builds it, isolated in a git worktree | /implement-feature |
| 5 | /clean-feature |
Aligns the specs with what actually shipped | /clean-feature |
2.2 Agent catalog
/implement-feature doesn't do the work alone: /plan-feature already decided which subagent owns each task, and /implement-feature just spins them up and hands out their share.
| Agent | What it does |
|---|---|
architect |
Researches, evaluates alternatives, and designs the solution before anything gets implemented |
code-developer |
Implements backend or general-purpose code from an already-clear plan |
ui-developer |
Designs and implements full screens, from visual concept to component |
test-developer |
Writes tests without running them |
tester |
Runs the tests and verifies the app's actual behavior |
judge |
Reviews another agent's work and returns PASS or CHANGES_REQUESTED |
If any of this sounds like your daily workflow, the install prompt above is the whole barrier to entry. I use this exact catalog every day and keep extending it as new gaps show up; if you hit one yourself, or something behaves differently than described here, open an issue on the repo.
3. Further reading and resources
| Resource | Description |
|---|---|
| Claude Code CLI: The complete guide | Blake Crosley's complete guide to Claude Code CLI |
| The 8 levels of AI | The framework this toolkit puts into practice |
| Spec-Driven Development with Coding Agents | DeepLearning.AI × JetBrains course on the constitution → spec → plan → implement workflow |
| Sandcastle | The git-worktree agentic-loop tool that inspired the isolation model here |
| Claude Code Ultimate Guide | Beginner-to-power-user reference with production-ready templates and a cheatsheet |
References
- DeepLearning.AI. (2026). Spec-driven development with coding agents [Course]. https://www.deeplearning.ai/courses/spec-driven-development-with-coding-agents
- FlorianBruniaux. (2026). Claude Code ultimate guide [GitHub repository]. GitHub. https://github.com/FlorianBruniaux/claude-code-ultimate-guide
- GitHub. (2026). Spec-kit [GitHub repository]. https://github.com/github/spec-kit
- Pocock, M. (2026). Sandcastle [GitHub repository]. GitHub. https://github.com/mattpocock/sandcastle