north-starr

north-starr

Agentic Development Workflow for AI Coding Tools

You don't improve productivity by going faster — you improve it by improving control.

Claude Code (marketplace) · VS Code Copilot (Homebrew)

agentlinter.com/r/w7Jo9X0qVOV8

---

Inspiration

north-starr is inspired by Idea Flow by Janelle Arty Starr — a framework for making invisible friction visible in software development.

The core insight: you don't improve productivity by going faster, you improve it by improving control. north-starr applies this to AI-assisted development — your AI partner learns your patterns, remembers your landmines, and works with control, not just speed.

The book · Talk · Podcast · "The most underrated book in software engineering management"

---

How it works

Every task goes through a complexity gate. Simple tasks flow fast. Complex tasks get automatic risk analysis and structured planning. The AI decides — you approve.

Task given
    │
    ▼
Complexity Assessment
    │
    ├─ Low ──────────► State files → Wait for approval → Code
    │
    ├─ Medium/High ──► /invert (risk analysis)
    │                      │
    │                      ▼
    │                  layoutplan agent (separate thread)
    │                      │
    │                      ▼
    │                  Approval gate → Code
    │
    ▼
RED (failing tests) → GREEN (implementation)
    │
    ▼
Completion menu:
    ├─ /generate-commit
    ├─ /generate-pr
    ├─ /learn (capture patterns & landmines)
    └─ Done

The complexity gate

Before any code change, the AI prints this assessment:

#	Question	Answer
0	Is current behavior covered by tests?	Yes / No
1	How many files will this change?	1-2 / 3+
2	Am I creating new types or protocols?	No / Yes
3	Does this require cross-module integration?	No / Yes

• All low → state the files and proceed
• Q0 = No → write tests for current behavior first
• Any medium/high → run /invert automatically, then plan before coding

No configuration needed. This is built into every project north-starr bootstraps.

PRD decomposition

For full product specs, /decompose breaks the work into manageable pieces before the complexity gate kicks in:

PRD received
    │
    ▼
/decompose → Scan & detect AI project
    │
    ├─ Non-AI project ──► storymap agent → .plans/STORIES-<name>.md
    │
    ├─ AI project ──────► chief-ai-po agent → .plans/STORIES-AI-<name>.md
    │                      (inverted stories, safety stories SA.1-SA.6,
    │                       human oversight checkpoints, graceful degradation)
    │
    ▼
Optional: GitHub Issues script
    │
    ▼
Pick a story → /invert → layoutplan → Implement

---

Install

Claude Code (marketplace):

/plugin marketplace add selcukyucel/north-starr
/plugin install north-starr

VS Code Copilot (Homebrew):

brew tap selcukyucel/north-starr https://github.com/selcukyucel/north-starr.git
brew install north-starr
cd your-project && north-starr init

Then run /bootstrap — north-starr explores your codebase and generates everything your AI needs.

---

What /bootstrap generates

All output is tool-native — the exact files each tool already reads:

Artifact	Claude Code	VS Code Copilot
Project context	CLAUDE.md	AGENTS.md
Universal context	AGENTS.md	AGENTS.md
Pattern rules	.claude/rules/*.md	.github/instructions/*.instructions.md
Landmine rules	.claude/rules/*.md	.github/instructions/*.instructions.md
Agents	.claude/agents/layoutplan.md, storymap.md, chief-ai-po.md	.github/agents/layoutplan.agent.md, storymap.agent.md, chief-ai-po.agent.md
Module context	CLAUDE.md per module	—

Pattern rules document how things are done in your codebase. Landmine rules document what to watch out for. Both are scoped by file path — they fire only when the AI touches matching files.

---

The learning loop

When your AI gets corrected, discovers a convention, or breaks something — /learn captures it as a pattern or landmine rule. These rules feed directly into how the AI operates next session.