skill-prospector

Skill Prospector

Analyze recent Claude Code session history to surface repetitive workflows worth turning into skills.

Overview

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.

Step 1: Extract session data

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.

Step 2: Identify patterns

Work through these signal types, strongest first:

Repeated bash commands

The aggregate.top_bash_commands section shows commands used across sessions. Look for:

• The same drush/ddev/composer command appearing 3+ times across sessions
• Multi-command sequences that always appear together (e.g., export config then commit)
• Commands with complex flags that the user keeps having to specify

Repeated prompt intent

Read through the prompts arrays across sessions. Look for:

• Different wordings of the same request (e.g., "run phpstan and fix issues", "check code quality", "lint my code")
• The same type of question asked in multiple sessions (e.g., "why is this test failing", "debug this CI failure")
• Requests that follow a predictable pattern with variable inputs

Repeated tool workflows

Look at tool_counts and the sequence of tools across sessions:

• Read → Edit → Bash(test) cycles on similar file types
• Grep → Read → Edit patterns for the same kind of change
• Agent spawning for the same categories of subtask

Skills already in heavy use

Check aggregate.top_skills — frequently invoked skills might indicate:

• A skill that works but could absorb adjacent manual steps
• A workflow that starts with a skill but requires manual follow-up every time

What to skip

• One-off tasks that are genuinely unique
• Simple single-command operations (reading one file, running one test)
• Patterns that an existing skill already handles well
• Debugging sessions — these are inherently ad-hoc

Step 3: Present recommendations

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.

Step 4: Save report

Save the analysis as a markdown report:

<project-root>/.claude/reports/skill-prospector-<YYYY-MM-DD>.md

The report should include:

1. Summary stats — sessions analyzed, date range, projects scanned
2. Skill candidates — the full recommendation blocks from Step 3
3. Raw stats — top bash commands, tool usage totals, skills invoked (as appendix)

Create the reports directory if it doesn't exist.

Step 5: Next steps

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.