Will Coach change my project files without asking?

No. Coach operates on a 'no silent writes' principle. It generates proposals that you must review and approve using the `/coach approve` command.

What kind of 'friction' triggers the Coach system?

It tracks four main categories: COMMAND_FAILURE (stderr/non-zero exits), USER_CORRECTION (explicit language like 'I said'), REPETITION (similar tasks repeated), and TONE_ESCALATION (frustration indicators).

Can I use Coach across different projects?

Yes. Coach can determine if a rule should be specific to your current repository or applied globally across all your Claude Code interactions based on the context of the learning signal.

How does the Coach skill learn from my interactions?

Coach monitors 'friction signals' such as when you say 'no' or 'stop,' when you repeat an instruction, or when a shell command fails. It then aggregates these events into proposed rules.

Where does Coach store the rules it creates?

Approved rules, checklists, and snippets are typically added to your project's CLAUDE.md file or your global Claude configuration directory.

Coach

Name: Coach
Author: netresearch

bynetresearch

•

Productivity & Workflow

Implements a self-improving learning loop for Claude Code by monitoring friction signals and proposing behavioral rule updates.

About

Coach is a specialized meta-skill designed to make Claude smarter over time by capturing and analyzing friction points during development sessions. It monitors signals such as user corrections, repeated instructions, and tool failures to extract actionable improvement candidates like rules, checklists, and snippets. With a strict 'no silent writes' policy, Coach ensures that all behavioral changes are transparent and require explicit user approval via the /coach command before updating your CLAUDE.md files. This creates a personalized, self-documenting environment where the AI continuously adapts to your specific coding standards and workflow preferences.

Key Features

Intelligent candidate generation for project rules, workflows, and code snippets
Multi-scope application to apply rules at either the project or global level
9 GitHub stars
Deduplication of learning events using fingerprinting and local SQLite storage
Automated signal detection for user corrections, tool errors, and tone escalation
Human-in-the-loop approval system to prevent unauthorized configuration changes

Use Cases

Building robust checklists for repetitive workflows like pre-deployment testing or linting
Refining Claude's understanding of custom project architectures and internal library preferences
Capturing 'Never do X' constraints immediately after a developer corrects a specific AI mistake

About

Key Features

Intelligent candidate generation for project rules, workflows, and code snippets
Multi-scope application to apply rules at either the project or global level
9 GitHub stars
Deduplication of learning events using fingerprinting and local SQLite storage
Automated signal detection for user corrections, tool errors, and tone escalation
Human-in-the-loop approval system to prevent unauthorized configuration changes

Use Cases

Building robust checklists for repetitive workflows like pre-deployment testing or linting
Refining Claude's understanding of custom project architectures and internal library preferences
Capturing 'Never do X' constraints immediately after a developer corrects a specific AI mistake