Reference Tracker Skill

Automatically track research sources and major prompts for academic and project documentation.

When to Activate

This skill activates automatically when:

1. User runs /track:init command
2. You notice CLAUDE_SOURCES.md or CLAUDE_PROMPTS.md files in the project

Claude autonomously decides when to use this skill based on the description and context.

Once activated, check ./.claude/.ref-autotrack to determine if auto-tracking is enabled.

Activation Check

Before any tracking operation:

1. Check if ./.claude/.ref-autotrack exists

   • If exists → auto-tracking enabled, proceed with tracking
   • If missing → auto-tracking disabled, skip tracking

2. Read ./.claude/.ref-config for verbosity settings:

   PROMPTS_VERBOSITY=major|all|minimal|off
   SOURCES_VERBOSITY=all|off
   

About the .ref-autotrack File

Location: ./.claude/.ref-autotrack

Purpose: Marker file that enables/disables automatic tracking

Contents: Contains explanatory comments for other Claude sessions:

# Auto-tracking marker for ref-tracker plugin
# Presence = enabled | Absence = disabled
# Managed by: /track:auto command
# See: /track:help for details

Created by: /track:auto or /track:auto on command

Managed by: /track:auto command (toggles on/off, or explicit on/off)

NOT created by: /track:init - tracking starts disabled

If you find this file: The project has reference tracking initialized. Use the ref-tracker skill to automatically log research sources and major prompts according to the verbosity configuration in ./.claude/.ref-config.

Tracking Rules

When auto-tracking is enabled (.ref-autotrack exists), automatically track:

CLAUDE_SOURCES.md

Track after every:

• WebSearch operation
• WebFetch operation
• Local documentation search (Grep/Read for docs, API references)

Respect SOURCES_VERBOSITY:

• all (default) → Track all operations
• off → Skip source tracking

CLAUDE_PROMPTS.md

Track after completing:

• Feature implementations
• Complex debugging/refactoring
• Multi-step workflows
• Non-trivial research questions

Respect PROMPTS_VERBOSITY:

• major (default) → Only significant multi-step academic/development work
• all → Every user request
• minimal → Only explicit user requests to track
• off → Skip prompt tracking

Format Specifications

CLAUDE_SOURCES.md

Pattern: [Attribution] Tool("Query"): Result

Rules:

• [User] if user explicitly requested search ("search the web for...")
• [Claude] if you autonomously searched for missing information (renamed from [Auto])
• Tool name in PascalCase (WebSearch, WebFetch, Grep, Read)
• Query in double quotes (exact query used)
• Result is URL or brief key concept (1-2 sentences max)
• Single line per entry, no blank lines between entries
• No headers or markdown formatting (pure KV file)

Examples:

[User] WebSearch("PostgreSQL foreign keys documentation"): https://postgresql.org/docs/current/ddl-constraints.html
[Claude] WebFetch("https://go.dev/doc/", "embed.FS usage"): Use embed.FS to embed static files at compile time
[Claude] Grep("CORS middleware", "*.go"): Found in api/routes.go:23-45

CLAUDE_PROMPTS.md

Pattern: Two-line entry with blank separator

Rules:

• Line 1: Prompt: "<user request verbatim or paraphrased>"
• Line 2: Outcome: <concise result in present tense, 1-2 sentences>
• Blank line after each entry
• Header included if creating new file

Example:

# CLAUDE_PROMPTS.md

This file tracks significant prompts and development decisions.

---

Prompt: "Implement JWT authentication"
Outcome: Created auth middleware, login/logout endpoints, JWT token generation and verification, integrated with user model

Prompt: "Debug slow database queries"
Outcome: Added query logging, identified N+1 problem, implemented eager loading, reduced query time from 2.3s to 0.15s

Tracking Workflow

For CLAUDE_SOURCES.md

1. After operation completes (WebSearch/WebFetch/doc search)
2. Check activation: Look for ./.claude/.ref-autotrack
3. Read config: Check SOURCES_VERBOSITY in ./.claude/.ref-config
4. If enabled and verbosity allows:
   • Check if ./CLAUDE_SOURCES.md exists
   • If missing: Create empty file
   • Append entry: [User|Claude] Tool("query"): result
   • Be silent: Never announce tracking to user

For CLAUDE_PROMPTS.md

1. After completing major request
2. Check activation: Look for ./.claude/.ref-autotrack
3. Read config: Check PROMPTS_VERBOSITY in ./.claude/.ref-config
4. If enabled and verbosity allows:
   • Check if ./CLAUDE_PROMPTS.md exists
   • If missing: Create with header
   • Append two-line entry + blank line
   • Be silent: Never announce tracking to user

File Locations

• Tracking files: Project root (./CLAUDE_SOURCES.md, ./CLAUDE_PROMPTS.md)
• Configuration: ./.claude/.ref-config
• Activation marker: ./.claude/.ref-autotrack

Never create tracking files in subdirectories or ~/.claude/.

Error Handling

If Edit fails (file locked, permissions):

1. Read current contents
2. Write with contents + new entry appended

Attribution Decision Guide

[User] - User explicitly requested:

• "Search for X"
• "Look up Y documentation"
• "Find examples of Z"
• "Check the docs for..."

[Claude] - You decided to search (renamed from [Auto]):

• Verifying current syntax/API
• Checking best practices
• Looking up error messages
• Researching to complete a task
• Missing information to answer question

Verbosity Handling

PROMPTS_VERBOSITY=major (default)

Track only:

• Multi-step feature implementations
• Complex debugging sessions
• Significant refactoring
• Academic research questions requiring substantial work

Skip:

• Simple questions ("What is X?")
• Typo fixes
• Trivial changes

PROMPTS_VERBOSITY=all

Track every user interaction and request.

PROMPTS_VERBOSITY=minimal

Track only when user explicitly says "track this" or similar.

PROMPTS_VERBOSITY=off

Skip all prompt tracking.

SOURCES_VERBOSITY=all (default)

Track every WebSearch/WebFetch operation.

SOURCES_VERBOSITY=off

Skip all source tracking.

Best Practices

1. Check activation first - Always look for ./.claude/.ref-autotrack before tracking
2. Read configuration - Respect verbosity settings in ./.claude/.ref-config
3. Be immediate - Track right after triggering action completes
4. Be silent - Never announce "Tracking to CLAUDE_SOURCES.md"
5. Be accurate - Use exact queries and URLs
6. Be concise - Keep results brief (1-2 sentences)
7. Be selective - Respect verbosity settings for what to track

Detailed Examples

For edge cases, multi-line results, concurrent tracking, and comprehensive examples, see references/examples.md.