AI Coding Playbook

AI Coding Playbook

Five skills that turn a capable LLM into an orchestrator that ships.

The Problem

You have a smart AI coding assistant. It can write code, answer questions, and follow instructions. But when you give it a non-trivial task — build a feature, refactor a module, design a system — it:

• Drifts into half-thoughts instead of finishing
• Grinds through boilerplate in the main thread
• Picks the first plausible approach instead of exploring
• Doesn't know when to stop and ask you for help
• Repeats the same mistakes across sessions

You spend more time correcting it than it saves.

The Fix

The skills in this pack are the operating system we layer on top of the raw model. Each one captures a discipline that's hard to maintain in the moment but easy to forget when you're heads-down:

• Plan first — write down what you're about to do, with a Brief and a Next Action
• Phase the work — name what kind of step this is (explore / design / apply / verify)
• Delegate execution — dispatch well-defined tasks to a cheaper model, watch the loop
• Debate the hard parts — when the answer is open-ended, run multiple agents and synthesize
• Author the patterns — when something recurs, capture it as a reusable skill

What's In The Pack

planner — The discipline of writing it down

• Triggers on any 3+ step task, multi-file edit, or "let's tackle X" phrase
• Writes a persistent plan to .pi/plans/<date>-<slug>.md
• Format: Brief / Status / Last touched / Next action header + free-form body
• One plan per block of work — never merge unrelated work

sdd — Spec-Driven Development as a vocabulary

• Defines four phases: explore → design → apply → verify
• "A task can skip phases. A task can revisit a phase. The phases are vocabulary, not gates — but the order is the default."
• Includes a delegation trigger table — when to stop and dispatch to a worker

worker — The organizer pattern

• "The worker is not a fire-and-forget subprocess. It's a loop."
• Worker writes checkpoint → Organizer reads → Organizer feedback → Worker continues
• Default for execution — if it involves more than a couple of tool calls, dispatch a worker
• Includes stopping rules (max attempts, explicit abort)

team — Multi-agent debate for hard decisions

• Runs multiple models in parallel for structured debate
• Falls back to simulated personas if only one model is available
• "Architecture, design, API, or data-model decisions" all qualify
• Triggers: "compare A vs B", "decide between", "help me choose", "trade-offs between"

create — When to author your own extensions

• Decision tree: skill vs. agent vs. command vs. rule vs. hook
• Skill anatomy (the markdown + scripts format)
• Anti-patterns to avoid

Adapters Included

These skills were authored for pi-agent, but we include drop-in adapters for the major AI coding tools:

The adapters translate the methodology into the format each tool expects. Same playbook, different packaging.

Real Usage

This is the playbook we used to build every library in this storefront:

• 3 React/TypeScript libraries, scaffolded and shipped
• The Stripe product/pricing/payment-link setup automation
• The source repository + automated build pipeline
• The storefront itself (Next.js, Tailwind, static export)

It works. It's not magic — it's discipline.

License

Personal-use commercial. One purchase = one seat. Install in your dev environment, modify freely, use in client work. Do not resell the pack or the unmodified skills as a competing product.

What's in the download

A .tgz containing:

ai-coding-playbook-0.1.0/
├── README.md              ← this overview
├── INSTALL.md             ← per-tool setup walkthrough
├── METHODOLOGY.md         ← the longer write-up
├── skills/                ← the 5 skill files (SKILL.md)
│   ├── planner/
│   ├── sdd/
│   ├── worker/
│   ├── team/
│   └── create/
├── adapters/              ← drop-in for other tools
│   ├── claude-code/
│   ├── cursor/
│   ├── cline/
│   └── generic/
└── examples/              ← worked examples from real projects

Updates

Free for life — re-download any time.