Cf Learn

$cf-learn

Extract learning points for human understanding. Topic: $ARGUMENTS

Purpose

In vibe coding, AI writes the code and human approves it — but the human doesn't actually learn anything. This skill extracts knowledge from the conversation and writes it as clear, educational docs so the human builds real understanding over time.

Workflow

Step 0: Custom Guide

bash "${PLUGIN_ROOT}/lib/load-custom-guide.sh" cf-learn

If the block above printed anything, apply only the ## Before, ## Rules, and ## After sections; if it shows the raw command instead of output, re-run that exact load-custom-guide.sh fence now.

Step 1: Load Config

Read config from global only:

1. Global: ~/.coding-friend/config.json

If learn.disabled is true in the global config, stop immediately and inform the user: "CF Learn is disabled. Re-enable it with cf config or by setting learn.disabled: false in ~/.coding-friend/config.json."

Extract settings with these defaults:

learn settings:

Language resolution: Read learn.language from global config. If not set, fall back to top-level language. If neither is set, default to en (English).

Path resolution for outputDir:

• Starts with / → absolute path, use as-is
• Starts with ~/ → expand ~ to home directory
• Otherwise → treat as relative to home directory
• Always pass file_path as an absolute path to the cf-writer agent

If outputDir directory doesn't exist, create it.

Auto-commit git check:

Before deciding whether to ask about or apply autoCommit, check if outputDir is inside a git repository:

git -C "<resolved_outputDir>" rev-parse --is-inside-work-tree 2>/dev/null

• If exit code is 0: the dir is a git repo — autoCommit from config applies
• If exit code is non-zero: the dir is NOT a git repo — skip auto-commit entirely (set autoCommit = false), do not mention it to the user

Step 2: Identify Knowledge Points

Scan the conversation for things the human might not fully understand:

• New concepts: Design patterns, algorithms, architecture principles
• Language features: Syntax, idioms, type system features
• Library/tool usage: API patterns, configuration, best practices
• Debugging techniques: How a bug was found and fixed
• Best practices: Why something was done a certain way

Step 3: Categorize

Use categories from config. Default categories:

File path: {outputDir}/{category}/{name}.md

Discovery: Check Existing Files

Before creating a new file, check if a relevant file already exists in the target category:

bash "${PLUGIN_ROOT}/skills/cf-learn/scripts/list-learn-files.sh" "/absolute/path/to/outputDir"

Replace /absolute/path/to/outputDir with the resolved absolute path from Step 0 before running.

For each potentially relevant file, read its first 20 lines to understand what it covers. If the new knowledge fits an existing file, append to it instead of creating a new one.

Step 4: Assess Complexity

Before delegating to the cf-writer agent, assess the complexity of the content to write:

Use cf-writer agent (low reasoning effort) when:

• Simple, factual concepts (e.g., "how to use X tool", "naming convention for Y")
• Straightforward tool/library usage notes
• Short content with clear structure
• Single-concept explanations

Use cf-writer-deep agent (medium reasoning effort) when:

• Content requires deep reasoning about nuanced technical concepts (e.g., explaining race conditions, distributed system trade-offs, complex type system features)
• Very long context needs to be synthesized into a coherent doc
• Multi-concept synthesis is needed (connecting several ideas into one explanation)
• Advanced architecture explanations or subtle debugging insights

Step 5: Delegate to cf-writer Agent

Choose cf-writer or cf-writer-deep per Step 4 (Assess Complexity), then build the spec (do not dispatch yet).

Determine Task Type

• If an existing file was found in Step 2 (Discovery) that matches → task: append
• Otherwise → task: create

Build the Write Spec

Include ALL of these in your delegation prompt to the cf-writer agent:

WRITE SPEC
----------
task: create | append
file_path: {outputDir}/{category}/{name}.md
language: {learn.language from config (with fallback chain)}
content: |
  <The full markdown content to write, including frontmatter for new files>
readme_update: {readmeIndex from config: false | true | per-category}
readme_path: {outputDir}/README.md (and {outputDir}/{category}/README.md for per-category)
auto_commit: {autoCommit from config}
commit_message: learn: <brief summary of topics added>
existing_file_action: append
category_description: {description of the category from config}

Content to Include in the Write Spec

For new files (task: create), include full content with frontmatter:

---
title: "<Concept Name>"
category: "<category-name>"
tags: [tag1, tag2, tag3]
created: YYYY-MM-DD
updated: YYYY-MM-DD
---

# <Concept Name>

## What

<1-2 sentences: what is this concept?>

## Why

<When would you use it? Why does it matter?>

## How

<Code example from the actual project — not a generic tutorial>

## Gotchas

- <Common mistake 1>
- <Common mistake 2>

## Read More

- <Link to official docs>
- <Link to good tutorial>

For appending (task: append), include only the new content to add:

• New section under a ## Heading
• Note to update updated date in frontmatter
• Note to add new tags if applicable
• If file is getting long (>300 lines), switch to task: create with a new file name

Language Rules (include in write spec)

• en: Write everything in English
• vi: Write explanations in Vietnamese. Keep all technical terms, code, commands, library names, and variable names in English
• Other: Write explanations in the configured language. Keep technical terms in English

README Index Formats (include if readme_update is not false)

Mode true (single README) — {outputDir}/README.md:

# Learning Notes

## Categories

### <category-name>/

- [file-name.md](<category>/file-name.md) - Brief description

---

_Last updated: YYYY-MM-DD_

Mode "per-category" — two files:

1. Category README at {outputDir}/{category}/README.md:

# <Category Name>

<category description>

## Notes

- [file-name.md](file-name.md) - Brief description

---

_Last updated: YYYY-MM-DD_

2. Main README at {outputDir}/README.md:

# Learning Notes

## Categories

### [<Category Name>](<category-folder>/)

## <category description> — _N notes_

_Last updated: YYYY-MM-DD_

Only include categories with at least one note file (excluding README.md).

Invoke the Agent

Dispatch cf-writer or cf-writer-deep (per Step 4) with the complete write spec as the prompt.

Step 6: Confirm

Read back the cf-writer agent's output and show the user:

1. What was learned and where the doc was saved (full path)
2. Whether the file was created new or appended to
3. If auto-committed, show the commit message

Rules

• Write for a human who can read code but might not understand WHY it works
• Use examples from the ACTUAL conversation/project, not generic textbook examples
• ELI5 style — explain like the reader is smart but unfamiliar with this specific topic
• One concept per file — keep it focused (unless appending a closely related point)
• If $ARGUMENTS is provided, focus on that topic
• If no topic given, pick the most interesting/valuable learnings from the session
• Create directories as needed
• NEVER delete or overwrite existing content — only append or create new