From dev-general
Analyze recent Claude Code sessions to find repetitive tasks that should become reusable skills. Use when the user says "find skills", "what should be skills", "analyze my sessions", "repetitive tasks", "skill candidates", "mine my sessions", "what am I doing repeatedly", "skill opportunities", "automation opportunities", or wants to discover patterns in their Claude Code usage that could be automated. Also use proactively when a user has been doing the same kind of work across multiple sessions.
How this skill is triggered — by the user, by Claude, or both
Slash command
/dev-general:skill-prospectorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Analyze recent Claude Code session history to surface repetitive workflows worth turning into skills.
Analyze recent Claude Code session history to surface repetitive workflows worth turning into skills.
Session data lives in ~/.claude/projects/{encoded-project-path}/ as JSONL files — one per session. Each line is a JSON object containing user prompts, assistant responses (with tool calls), timestamps, and metadata. This skill uses a bundled extraction script to parse that data, then you analyze the results for recurring patterns.
Run the extraction script bundled with this skill. Determine the skill's directory from the path of this SKILL.md file.
python3 <skill-directory>/scripts/extract_sessions.py \
--cwd <project-root> \
--days <N, default 7> \
-o /tmp/skill-prospector-extract.json
Add --all-projects if the user asks to scan across all their projects.
Read the output file. If it's very large (>3000 lines), read the metadata and aggregate sections first to get bearings, then sample a few sessions.
Work through these signal types, strongest first:
The aggregate.top_bash_commands section shows commands used across sessions. Look for:
Read through the prompts arrays across sessions. Look for:
Look at tool_counts and the sequence of tools across sessions:
Check aggregate.top_skills — frequently invoked skills might indicate:
For each skill candidate, present a block like this:
### 1. <skill-name> (kebab-case)
**What it automates:** 1-2 sentences describing the repetitive workflow.
**Evidence:** Which sessions/prompts demonstrated this (dates + brief quotes).
**Frequency:** N occurrences across M sessions in the last D days.
**Context needed:** What inputs/config the skill would require.
**Complexity:** Simple | Medium | Complex
**Impact:** Low | Medium | High (frequency × effort-saved-per-invocation)
Rank by impact, highest first. Aim for 3-7 candidates — enough to be useful, not so many it's overwhelming. If you find fewer than 3, say so honestly rather than padding with weak candidates.
Save the analysis as a markdown report:
<project-root>/.claude/reports/skill-prospector-<YYYY-MM-DD>.md
The report should include:
Create the reports directory if it doesn't exist.
After presenting the report, ask which skill(s) the user wants to create. When they choose one, invoke /skill-creator with the context you've gathered — describe what the skill should do, what triggers it, what inputs it needs, and any example prompts/commands from the session data.
Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub foxtrotcharlie/claude-skills --plugin dev-general