Generate AI tool configuration for an existing project. Explores the codebase and produces context files, path-scoped pattern rules, landmine rules, and agents for Claude Code and VS Code Copilot. Run once per project.
Instalação
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Generate AI tool configuration for an existing project. Explores the codebase and produces context files, path-scoped pattern rules, landmine rules, and agents for Claude Code and VS Code Copilot. Run once per project.
Project Bootstrap — AI Tool Configuration Generator
Purpose
Explore an existing project and generate native configuration for the AI tools in use so they work effectively from the first task. Runs once, produces artifacts that each tool auto-loads and auto-enforces.
Outputs per tool:
Artifact
Claude Code
VS Code Copilot
Project context
CLAUDE.md
AGENTS.md
Pattern rules
.claude/rules/*.md
.github/instructions/*.instructions.md
Landmine rules
.claude/rules/*.md
.github/instructions/*.instructions.md
Agents
.claude/agents/*.md
.github/agents/*.agent.md
Module context
CLAUDE.md (in module dir)
—
Universal
AGENTS.md
AGENTS.md
Pattern and landmine rules use the same content — only the frontmatter format differs per tool.
Generate artifacts for the tool currently in use. If running inside Claude Code, generate Claude Code artifacts. If running via VS Code Copilot, generate Copilot artifacts. AGENTS.md is always generated (universal).
Pre-flight: Auto-sync Check
Before starting, read the root CLAUDE.md (if it exists and is not a starter template). If ## How to Approach Tasks appears AFTER ## Tech Stack or ## Architecture, the file has an outdated section order. Silently run /genai-sync to fix it before continuing with this skill.
When to Use
First time working on an existing project with any AI coding tool
When project-specific AI configuration is empty or missing
When onboarding to a new codebase
Prerequisites
The project root must be accessible
Git history is helpful but not required (used for churn analysis)
Content Depth
Generated rules must carry enough depth to be genuinely useful. Use two content structures from the project's knowledge base:
Pattern structure (skills/_references/patterns/_TEMPLATE.md) — for conventions and reusable approaches. Each pattern rule follows the full template: Virtues, When to Use, Problem It Solves, Core Approach with step-by-step code examples, Complete Example, Best Practices, Common Mistakes with wrong/fix code, Variations, Testing This Pattern, Performance Considerations, Related patterns and landmines.
Landmine structure (skills/_references/landmines/_TEMPLATE.md) — for danger zones and known traps. Each landmine rule follows the full template: Severity with evidence, Threatens (virtues endangered), Symptoms, Root Cause, The Trap (why devs fall in), Safe Approach (Don't/Do with code), Validation, Real-World Impact, Prevention, Related patterns and landmines, Origin.
Code Virtues (skills/_references/virtues/code-virtues.md) — the 7 code virtues (Working, Unique, Simple, Clear, Easy, Developed, Brief) used to tag patterns and landmines. Higher virtues take precedence.
Line limits:
Context files (CLAUDE.md, AGENTS.md): MUST stay under 100 lines (max 125 if critical context would be lost). Split into multiple scoped files rather than exceeding.
Pattern and landmine rules: Should be as detailed as needed to follow the full template — typically 50-150 lines. Depth and working code examples matter more than brevity. These are the project's knowledge base.
Workflow
Step 1: Explore & Detect
Goal: Understand the shape and stack of the project before reading code.
Actions:
Detect the technology stack — Read config files at the root to identify:
Language & version: look for Package.swift, pyproject.toml, setup.py, requirements.txt, Cargo.toml, go.mod, package.json, tsconfig.json, Gemfile, build.gradle, pom.xml, *.csproj, Makefile, CMakeLists.txt, etc.
Package manager: SPM, pip/poetry/uv, cargo, go modules, npm/yarn/pnpm, bundler, gradle/maven, NuGet, etc.
Build system: Xcode, webpack/vite/esbuild, cargo, go build, make, gradle, etc.
Test runner: XCTest, pytest, cargo test, go test, jest/vitest, RSpec, JUnit, etc.
CI/CD: .github/workflows/, .gitlab-ci.yml, Jenkinsfile, .circleci/, etc.
Guardrail placement — where input/output filters sit in the pipeline, PII detection, injection defenses
Fallback strategy — what happens when primary model is unavailable, degraded mode behavior
Cost architecture — where the expensive operations are, batch vs real-time decisions, token budgets per component
Step 3: Discover Patterns
Goal: Build a comprehensive catalogue of "how things are done here" so new code follows conventions and the knowledge survives across sessions.
Scope: Analyze ALL modules, not a sample. Walk the entire codebase systematically. A 3-5 module sample misses cross-cutting patterns, infrastructure conventions, and operational practices that only surface when looking broadly.
Actions:
Map every module — list all top-level directories/packages. Group them by role (feature modules, shared libraries, infrastructure, configuration, tests, scripts, deployment).
Scan each group for recurring patterns. Look for conventions in these areas — not all will apply to every project, focus on what the codebase actually uses:
Structure — how features, modules, or components are organized and laid out
Data flow — how data enters, moves through, and exits the system
Dependencies — how components get what they need (injection, imports, configuration)
Error handling — how errors are caught, surfaced, and recovered from
State — how state is managed, shared, and synchronized
External boundaries — how the system communicates with anything outside itself
Testing — how tests are organized, what's mocked, what's tested end-to-end
Build & deploy — how the project is built, packaged, and shipped
Look for shared utilities, base classes, protocols, or helpers reused across modules — these often encode implicit patterns worth documenting explicitly.
Cross-reference patterns — build a relationship map before writing any rule files:
Which patterns compose (used together in the same flow)?
Which patterns are alternatives (different approaches to the same problem)?
Which landmines does each pattern prevent?
Which landmines result from misapplying a pattern?
Every rule's "Related" section must reference at least one other rule. If a pattern has no relationships, it's either too generic or you missed a connection.
For each discovered pattern, capture using the full pattern structure from skills/_references/patterns/_TEMPLATE.md:
Virtues — tag with the primary virtue(s) the pattern serves (from skills/_references/virtues/code-virtues.md: Working, Unique, Simple, Clear, Easy, Developed, Brief)
When to Use / Not Good For — specific situations
Problem It Solves — what goes wrong without it, what improves with it
Core Approach — step-by-step with code examples from the actual codebase
Complete Example — a full working example demonstrating the pattern end-to-end
Best Practices — do this, why
Common Mistakes — wrong approach with code, fix with code
Variations — alternative forms of the pattern found in the codebase
Testing This Pattern — how to verify correct application, with test code example
Performance Considerations — any performance implications and mitigations
Related — links to other patterns and landmines
Aim for completeness. A thorough bootstrap should discover 15-40 patterns depending on project complexity. If you find fewer than 10, you likely stopped too early — revisit areas beyond the core feature code (build, deploy, testing, configuration, shared infrastructure).
Step 4: Detect Danger Zones
Goal: Build a comprehensive map of every area where developers can get burned — from code-level traps to operational pitfalls.
Scope: Look everywhere, not just code hotspots. Danger zones exist in configuration, deployment, infrastructure, third-party integrations, and operational procedures — not only in source code.
Actions:
Complexity hotspots — large files, deeply nested logic, functions with many parameters, types with many responsibilities. These areas break easily and are hard to modify safely.
Misleading abstractions — code that doesn't do what its name suggests, dead code paths, unused parameters that look required. These trap developers into wrong assumptions.
Silent failures — swallowed errors, empty catch blocks, default fallbacks that hide problems. These make debugging nearly impossible.
Developer warnings — search for TODO, FIXME, HACK, XXX, WORKAROUND, TEMPORARY comments. Each one is a documented landmine left by a previous developer.
Files changed most frequently often contain instability or poorly understood behavior.
External boundaries — anywhere the system communicates with something outside itself (APIs, services, SDKs, hardware, file systems). These are where assumptions break and failures cascade.
Configuration sensitivity — settings, credentials, feature flags, or environment-specific behavior where a wrong value causes silent or catastrophic failure.
Resource management — anything the system allocates, opens, or acquires that must be released, closed, or returned. Leaks here cause gradual degradation.
Test gaps — modules or features with no test coverage. Untested code is a landmine waiting to detonate.
Hardcoded prompts — prompts embedded in code instead of managed externally. These are hard to version, test, and iterate on without code changes.
Missing error handling on model calls — no retry, no fallback, no timeout. Model API calls fail regularly and must be handled gracefully.
Unbounded context — user input or RAG results passed to model without size limits. This causes token limit errors, unexpected costs, and degraded output quality.
No output validation — model output used directly without parsing or validation. LLM outputs are non-deterministic and must be validated before use.
PII in prompts or logs — user data passed to external model APIs or logged. This creates compliance risk and potential data breaches.
Missing cost controls — no token limits, no rate limiting, no cost alerts. Runaway model calls can generate significant unexpected costs.
Stale embeddings — vector indexes without refresh mechanism. As source data changes, embeddings become stale, causing retrieval quality to silently degrade.
Single model dependency — no fallback if primary model is unavailable. Model providers have outages; applications need graceful degradation paths.
For each danger zone, capture using the full landmine structure from skills/_references/landmines/_TEMPLATE.md:
Severity — CRITICAL / HIGH / MEDIUM / LOW based on real-world impact. Justify with evidence (file counts, specific instances found, blast radius).
Threatens — tag with the virtue(s) this landmine endangers (from skills/_references/virtues/code-virtues.md: Working, Unique, Simple, Clear, Easy, Developed, Brief)
Symptoms — observable signs you've hit this
Root Cause — technical explanation of why this happens
The Trap — why it seems correct, what makes it non-obvious
Safe Approach — Don't (dangerous code with explanation) / Do (safe code with explanation)
Validation — how to verify you're safe, detection in existing code
Real-World Impact — what actually happens when this goes wrong (cite specific instances found in the codebase)
Prevention — habits, code review checks, and validation steps
Related — safe patterns that avoid this, other related landmines
Origin — how/when this was discovered during bootstrap analysis
Aim for completeness. A thorough bootstrap should discover 5-15 landmines depending on project maturity. If you find fewer than 3, you likely stopped too early — revisit areas beyond the core feature code.
Step 4b: Quality Gate — Self-Review Before Writing
Goal: Catch gaps and inconsistencies before generating files. This step is mental — no files written.
Checks:
Coverage check — Does every major module have at least one pattern or landmine? If a module has neither, revisit it — you may have missed something.
Template completeness — Does every pattern have ALL required sections from the reference template (Virtues, When to Use, Problem, Pattern with code, Complete Example, Best Practices, Common Mistakes, Variations, Testing, Performance, Related)? Does every landmine have ALL required sections (Severity with evidence, Threatens, Symptoms, Root Cause, Trap, Safe Approach, Validation, Real-World Impact, Prevention, Related, Origin)?
Path glob review — For each rule, verify the glob matches only the files where the pattern actually applies. If you used a broad glob like **/*.ext or **/src/**, narrow it.
Cross-reference integrity — Every pattern's "Related" section should link to at least one other pattern or landmine. Every landmine's "Related" should link to the safe pattern that avoids it.
Code example audit — Every code example must use the project's actual types, imports, and conventions. No placeholder names like MyService or doSomething() unless the project itself uses those names.
Deduplication — No two rules should cover the same concern. If two rules overlap, merge them or clarify their distinct scopes.
If any check fails, fix it before proceeding to Step 5.
Step 5: Generate Configuration
Goal: Produce configuration files for the current AI tool. Generate all sections below. The content is the same — only the file locations differ per tool.
A. Project Context (root-level)
Write the project context to these locations:
CLAUDE.md — Claude Code (auto-loaded). Uses plain text prompts for user approval gates.
AGENTS.md — Universal / VS Code Copilot (auto-loaded). Uses vscode_askQuestions for user approval gates.
All context files share the same project content (Tech Stack, Architecture, Grain, Module Map). The only difference is the managed sections — CLAUDE.md uses plain text prompts while AGENTS.md uses vscode_askQuestions for interactive approval gates.
Section order matters — instructional sections (How to Approach Tasks, When to Learn) come FIRST so the AI tool sees them before project context.
Managed sections must be copied verbatim. Read the template file first, then copy the managed sections (everything between <!-- [NORTH-STARR-GENAI:*] --> and <!-- [/NORTH-STARR-GENAI:*] --> markers, including the markers themselves) character-for-character into the generated file. Do NOT paraphrase, simplify, summarize, or rewrite these sections — they contain precise instructions (assessment tables, decision rules, workflow steps) that must be preserved exactly.
For CLAUDE.md: read templates/CLAUDE.md and copy the how-to-approach-tasks and auto-learn managed sections verbatim
For AGENTS.md: read templates/AGENTS.md and copy the managed sections verbatim
Project context sections are identical across all files:
## Tech Stack
[List languages with versions, frameworks, key dependencies, build tools, package manager, test runner, CI/CD — be specific, not generic. Include: LLM provider SDKs, vector stores, embedding models, eval frameworks, prompt management tools.]
## Architecture
[Name the pattern (MVVM, Clean, etc.), topology (monolith, modular, etc.). List each layer with its responsibility and dependency direction. Include DI approach and state management strategy. Include model selection strategy, caching approach, and guardrail placement.]
## Grain
[What changes easily (e.g. adding a new feature screen) vs. what is hard (e.g. changing navigation pattern). State what to avoid going against and why.]
## Module Map
[List each top-level module with one-line purpose. Show key dependencies between modules. Note shared infrastructure, model configs, and prompt registries.]
If any of these files already exist with project-specific content, merge rather than overwrite.
B. Module-Level Context Files
For each danger zone or complex module found in Step 4, write context in that directory:
CLAUDE.md — Claude Code (auto-loaded when working in that directory)
# [Module Name]
[What this module does, how it fits in the architecture]
## Caution
[Specific warnings: race conditions, fragile logic, missing tests, known bugs]
## Patterns
[How this module does things, if different from the project defaults]
C. Pattern Rules & Landmine Rules
Generate pattern and landmine rules for the current tool. The content is the same — only the file location and frontmatter format differ per tool.
File formats per tool:
Claude Code — .claude/rules/*.md:
---
paths: ["glob/pattern/**"]
---
[Rule content]
VS Code Copilot — .github/instructions/*.instructions.md:
Landmine Rules — one rule file per danger zone discovered in Step 4.
Follow the full landmine template from skills/_references/landmines/_TEMPLATE.md. Each landmine rule file must include:
Severity (CRITICAL / HIGH / MEDIUM / LOW) with evidence justification, Category, and Threatens (which code virtues this landmine endangers — see skills/_references/virtues/code-virtues.md)
## Quick Summary — one-line description
## Symptoms — observable signs you've hit this
## Root Cause — technical explanation of why this happens
## The Trap — why developers fall in, what makes it non-obvious
## Safe Approach — Don't (dangerous code with explanation) / Do (safe code with explanation). Code examples must use the project's actual types and conventions.
## Validation — how to verify you're safe, detection patterns in existing code (include grep/search commands)
## Real-World Impact — what actually happens when this goes wrong. Cite specific instances found in the codebase (file names, line counts, blast radius).
Create one rule file per pattern or landmine discovered in Steps 3 and 4. Patterns become pattern rules, danger zones become landmine rules. The specific concerns depend on the project — generate rules only for what was actually found in the codebase.
Guidelines:
Generate only rules that reflect real patterns or dangers found in the codebase — never invent conventions
Use specific path globs — broad rules waste context on irrelevant files. Path globs must be as narrow as possible to match only files where the pattern applies. Anti-patterns:**/*, **/Sources/**, or **/*.py (matches everything). Good examples:**/Sources/Feature/**/*ViewModel*, **/tests/integration/**, **/src/api/routes/**/*.ts, **/models/**/*repository*. When a pattern applies to a specific file naming convention (e.g., files ending in ViewModel, Service, _test.py), include that in the glob. When it applies to a specific directory subtree, scope the glob to that subtree.
Keep each rule file focused on one concern
Include code examples in every rule — abstract descriptions without code are not actionable
Pattern and landmine rules should be as detailed as the templates require — typically 50-150 lines. Depth matters.
The content is the same across tools — only the frontmatter format differs
Include a _TEMPLATE.md in the rules directory for future contributions via /learn. The generated _TEMPLATE.md MUST be derived from the reference templates (skills/_references/patterns/_TEMPLATE.md and skills/_references/landmines/_TEMPLATE.md). It must include ALL sections from both templates — Virtues/Threatens tags, Testing This Pattern, Performance Considerations, Complete Example, Changelog, and Origin fields. Combine both pattern and landmine structures into one unified template with comments indicating which sections apply to which type. The template's Language field and code fence language must match the project's detected language — never hardcode a specific language.
D. Agents
Generate agents for the current tool. Do not hardcode the agent list — dynamically list all .md files in the plugin's templates/claude/agents/ (for Claude Code) or templates/github/agents/ (for VS Code Copilot) and copy each one to the project's agent directory.
Claude Code — copy each agent from templates/claude/agents/*.md to .claude/agents/*.md:
Core agents (always install):
genai-layoutplan — spawned by /ai-invert for implementation planning
genai-storymap — spawned by /decompose for PRD decomposition
chief-ai-po — spawned by /decompose for AI-specific story decomposition (3 modes: decompose, refine, incorporate-feedback)
Analysis agents (install for AI projects):
ai-invert-analyst — spawned by /ai-invert skill for AI-specific risk analysis (Q1/Q2 gate)
baseline-capturer — spawned by /baseline skill for pre-change performance snapshot (Q3 gate)
auto-improver — spawned by /autoimprove skill for hill-climbing prompt optimization
Build specialists (install for AI projects):
prompt-engineer — spawned during BUILD phase for prompt design and versioning
rag-advisor — spawned during BUILD phase for RAG pipeline design
integration-planner — spawned during BUILD phase for external system integration
Validation agents (install for AI projects):
eval-designer — spawned during HARDEN phase to run eval suites
guardrails-designer — spawned during HARDEN phase to test safety guardrails
prompt-adversary — spawned by guardrails-designer or standalone for red-teaming
ai-ops — spawned during HARDEN phase for monitoring and observability setup
Orchestration agents (install for AI projects):
orchestrator — spawned by /orchestrate for multi-story pipeline management
ai-architect — spawned by orchestrator during DESIGN phase
cost-estimator — spawned by orchestrator during DESIGN phase
Delivery agents (install for AI projects):
demo-builder — spawned during DELIVER phase for client packaging
VS Code Copilot — copy each agent from templates/github/agents/*.agent.md to .github/agents/*.agent.md.
Why all agents are installed: The workflow in CLAUDE.md/AGENTS.md (v3.0) automatically spawns specialist and validation agents during the gate (Q1/Q2/Q3/Q4) and BUILD/HARDEN phases. Every specialist has MUST-cite peer consultations — if an agent isn't installed in the project, the cross-consult chain breaks. Installing all agents during bootstrap ensures the full pipeline works from the first task.
Generate additional project-specific agents only if the project clearly warrants them (e.g., an explorer agent for very large codebases).
Post-Bootstrap Checklist
Files:
AGENTS.md at root (always)
CLAUDE.md at root (if running in Claude Code)
Module-level CLAUDE.md for each identified danger zone (if running in Claude Code)
Pattern rules in the current tool's format — aim for 15-40 depending on project complexity
Landmine rules in the current tool's format — aim for 5-15 depending on project maturity
_TEMPLATE.md in the rules directory — must include ALL sections from reference templates
At least one project-tuned explorer agent (optional, for large codebases)
Quality:
Context files (CLAUDE.md, AGENTS.md) are under 100 lines (max 125)
Every pattern rule has: Virtues, When to Use, Problem, Pattern with code, Complete Example, Best Practices, Common Mistakes, Variations, Testing, Performance, Related
Every landmine rule has: Severity with evidence, Threatens, Symptoms, Root Cause, Trap, Safe Approach, Validation, Real-World Impact, Prevention, Related, Origin
No path glob matches the entire codebase (e.g., **/* or **/*.ext)
Every rule's "Related" section links to at least one other rule
All code examples use the project's actual types, imports, and conventions — no generic placeholders
_TEMPLATE.md Language field matches the detected project language
