mcollina/init
Creates, updates, or optimizes an AGENTS.md file for a repository with minimal, high-signal instructions covering non-discoverable coding conventions, tooling quirks, workflow preferences, and project-specific rules that agents cannot infer from reading the codebase. Use when setting up agent instructions or Claude configuration for a new repository, when an existing AGENTS.md is too long, generic, or stale, when agents repeatedly make avoidable mistakes, or when repository workflows have changed and the agent configuration needs pruning. Applies a discoverability filter—omitting anything Claude can learn from README, code, config, or directory structure—and a quality gate to verify each line remains accurate and operationally significant.
85
94%
Does it follow best practices?
Impact
72%
1.14xAverage score across 5 eval scenarios
Passed
No known issues
Evaluation results
100%
44%Task
Excludes tech stack summary
0%
100%
Excludes directory structure overview
0%
100%
Excludes linting rules already enforced by ruff
0%
100%
Excludes generic best-practice advice
33%
100%
Excludes architecture or design descriptions
80%
100%
Includes the uv run pytest caveat
100%
100%
Includes the legacy/ directory warning
100%
100%
Includes the non-standard dev server port (8001)
100%
100%
Content is specific and actionable
50%
100%
Signal-to-noise ratio is high
20%
100%
Does not duplicate README installation instructions
20%
100%
85%
Task
Subdirectory AGENTS.md files created
100%
100%
PCI sandbox flag captured in payments AGENTS.md
100%
100%
GPU memory env var captured in ml AGENTS.md
100%
100%
Security review requirement in payments AGENTS.md
100%
100%
Root AGENTS.md is short and cross-cutting only
70%
50%
Module-specific details absent from root AGENTS.md
30%
20%
ML model weights path captured
100%
100%
No tech stack summaries or architecture descriptions
62%
100%
No generic best practices
100%
100%
Reasoning or explanation for hierarchical placement
66%
66%
100%
7%Project Guide for AI Agents
Removes tech stack summary
100%
100%
Removes or omits directory structure listing
100%
100%
Removes code style rules duplicated by linter/formatter config
100%
100%
Removes generic development workflow platitudes
100%
100%
Removes or collapses obvious/redundant command descriptions
42%
100%
Retains the --legacy-peer-deps flag warning
100%
100%
Retains the scripts/migrate.sh requirement
100%
100%
Retains the payments/ module separate deployment pipeline warning
100%
100%
Output is shorter than input
100%
100%
Retained lines are specific and actionable
70%
100%
Does not introduce new inaccuracies or hallucinated content
100%
100%
56%
4%Input files
Non-discoverable commands section present
16%
0%
Landmines / do-not-touch section present
75%
100%
Scope or routing or task-specific constraints section present
50%
0%
No generic documentation sections
25%
100%
make db-reset caveat captured
16%
0%
SKIP_AUTH=1 flag captured
0%
0%
billing/ package landmine captured
100%
100%
config/local.yaml auth URL landmine captured
83%
100%
No generic development advice
100%
100%
Sections are short and high-signal
50%
100%
Content is specific enough to execute
83%
66%
85%
7%Agent Instructions
Consults .cursorrules
100%
100%
Consults .github/copilot-instructions.md
100%
100%
Incorporates non-discoverable guidance from source files
100%
100%
Does not blindly copy all cursor rules content
66%
100%
Recommends root-cause fix for wrong test command
0%
0%
Removes stale Docker setup instruction
100%
100%
Retains the production migration caveat
100%
100%
Does not include generic discoverable guidance
33%
100%
Output is shorter or more focused than existing AGENTS.md
50%
100%
Shows incremental update rather than blind replacement
83%
50%
Correct test command present in updated AGENTS.md
100%
100%
Next.js App Router / server component guidance included
100%
100%
- Evaluated
- Agent
- Claude
- Model
- Claude Opus 4.6
Table of Contents