ai-coding

A complete, spec-driven development workflow for pi. Instead of letting the agent jump straight to code, it guides every change through the same disciplined path: talk through the design first, agree on a spec, break it into small tasks, implement and review them one at a time, confirm the result matches the spec, and update the docs. The result is more predictable output, fewer surprises, and code you can actually trust — even on large changes.

It runs mostly on its own: the agent picks up the right phase at the right time, keeps its own context clean, and hands work to specialized sub-agents so the main agent stays sharp.

Installation

First install ai-base — the foundation layer of extensions, sub-agents, and web access this workflow builds on. Install it globally so it's available everywhere:

# ai-base — global, writes to ~/.pi/agent/settings.json
# replace vx.x.x with the desired release tag
pi install git:github.com/cristi-lng/ai-base@vx.x.x

Then install ai-coding per project:

# ai-coding — per project, writes to .pi/settings.json
# replace vx.x.x with the desired release tag
pi install git:github.com/cristi-lng/ai-coding@vx.x.x -l

Usage

The first time you use it in a project, let the agent set things up. It runs setup automatically — recording your project's coding conventions and choosing which AI models handle which jobs. If for some reason it doesn't, run /skill:ai-coding-setup yourself.

After that, just tell the agent what you want to build or change. It automatically starts by talking through the design with you and won't touch code until you approve a plan, then works through the rest of the steps on its own. If it ever skips straight to coding, run /skill:ai-coding-brainstorming to nudge it back on track.

Features

Workflow skills

The agent moves through these in order:

• setup — Learns your project's coding style and picks which AI models do which jobs.
• brainstorming — Talks through the design with you and writes a spec. No code until you approve it.
• spec-to-tasks — Splits the spec into small, ordered tasks that can each be checked on their own.
• implementation — Builds the tasks one at a time: one sub-agent writes the code, another reviews it.
• verification — Checks that what was built actually matches the spec.
• documentation — Updates the project's docs when behavior changed.

Under the hood

• Setup guard — Won't let the agent start working until the project has been set up.
• Progress widget — A line above your input box shows the current step, e.g. [Brainstorming 1/2: Clarifying questions].
• Specialized sub-agents — Each job gets the right model: reviewers are careful deep thinkers, the coder and doc-writer are balanced all-rounders, the explorer is a fast, lightweight scout. Chosen during setup.
• Small, focused context — When a phase finishes, its conversation collapses into a short summary. This keeps the main agent's context small so it stays in the smart zone and doesn't drift. Your full conversation is still saved and survives /resume.
• Auto-commits — In auto mode, the agent commits after each finished task.