Skill Finder — Behavioral Analysis of Claude Code Sessions

Analyze the user's Claude Code session history (~/.claude/history.jsonl and ~/.claude/transcripts/*.jsonl) to find repeated workflows worth automating.

What This Skill Does

Unlike claude-automation-recommender (which looks at codebase files to suggest automations), this skill looks at what the user actually does — their prompts, tools used, and projects — to find patterns worth turning into skills, plugins, agents, or CLAUDE.md entries.

Arguments

$ARGUMENTS — optional filters:

• No args: full analysis across all sessions
• A project name: filter to that project only
• "skills only" / "plugins only" / "agents only": focus on one recommendation type

Workflow

Phase 1: Extract Session Data

Read all user prompts from ~/.claude/history.jsonl (the display field has user messages, project has the working directory).

Also scan ~/.claude/transcripts/*.jsonl for tool usage patterns (lines with "type":"tool_use" have a tool_name field).

Strip any system prompt wrappers (e.g., <ultrawork-mode> blocks — extract the actual user message after ulw: or after the last ---\n\n).

Phase 2: Categorize & Cluster

For each user prompt, classify it into workflow categories by keyword matching. Use categories like:

• build_app_from_scratch (scaffold, create, new app, bootstrap, initialize)
• implement_from_ticket (linear, ticket, implement this)
• scrape_website (scrape, crawl, download contents, extract)
• write_perf_review (perf review, performance review, year end)
• debug_error (error, bug, fix, broken, failing, debug)
• git_commit_pr (commit, push, pull request, merge, branch)
• deploy_upload (upload, s3, deploy, publish)
• org_chart_people (who reports, staff engineers, org chart, direct reports)
• write_tests (test, spec, e2e, playwright, jest)
• cost_token_analysis (cost, token, usage, spend)
• design_ui (design, landing page, mockup, .pen, wireframe)
• refactor_code (refactor, clean up, simplify, reorganize)
• browser_automation (browser, navigate, click, open website)
• learn_claude_features (claude, skill, plugin, agent, mcp, hook)
• documentation (readme, doc, documentation)

Also track:

• Which projects each workflow appears in
• Total count per workflow
• Tool usage frequency across all sessions

Phase 3: Check Existing Setup

Read what the user already has:

• ~/.claude/skills/ — existing skills
• ~/.claude/CLAUDE.md — existing global instructions (may not exist)
• ~/.claude/plugins/installed_plugins.json — installed plugins
• ~/.claude/settings.json and ~/.claude/settings.local.json — hooks, permissions

Phase 4: Generate Recommendations

Apply this decision framework:

Phase 5: Output Report

Format the output as:

## Skill Finder Report

**Sessions analyzed**: X | **Prompts analyzed**: Y | **Projects**: Z

### Workflow Frequency
(table sorted by count, with project breakdown)

### Recommendations

#### CLAUDE.md (persistent instructions)
- What to add and why (things repeated across many sessions)

#### Skills (slash commands to create)
- `/skill-name` — what it does, why (N times you did this manually), example prompt it replaces

#### Agents (custom sub-agents)
- Agent name — what it does, why

#### Plugins (package for team)
- Plugin name — what skills/agents to bundle, why

### Already Automated
(list what the user already has as skills/plugins that covers a pattern)

### Quick Wins
(top 3 highest-impact recommendations ranked by frequency x effort saved)

Important Notes

• This is a read-only analysis — do not create any files, just output recommendations
• Focus on the top patterns — don't list every single message, surface the 5-10 most impactful opportunities
• Always compare against existing skills to avoid recommending what's already automated
• Show concrete examples from actual session history to justify each recommendation