Skills finder

Skills Finder

Scan today's Claude session history to spot repeated manual patterns, workflows done more than once, or multi-step processes that would benefit from being packaged as a reusable skill.

How to Run

Manual: /skills-finder

Automated: Runs daily at 9PM BST via launchd (com.0xj7r.skills-finder)

Process

Step 1: Extract Today's History

Read ~/.claude/history.jsonl and filter to today's messages only (midnight to now). Parse each line as JSON; use the timestamp field (milliseconds since epoch).

Group messages by sessionId and project. Include the full message text from the display field.

Step 2: Identify Skill Candidates

Look for these patterns across all of today's sessions:

Repeated workflows:

• The same sequence of steps performed 2+ times (even across different projects)
• Similar prompts with minor variations ("deploy X to Y", "deploy A to B")
• Copy-paste patterns where the user provided the same boilerplate instructions

Multi-step manual processes:

• Tasks that required 3+ back-and-forth messages to accomplish
• Processes where the user had to guide Claude through steps that could be codified
• Workflows that started with "first do X, then Y, then Z"

Tool/format patterns:

• Generating the same type of output repeatedly (reports, emails, PR descriptions)
• Using the same tools in the same sequence
• Applying the same conventions or rules manually each time

Near-miss skills:

• Tasks where the user said "like last time" or "same as before"
• Workflows adjacent to existing skills that could extend them

Step 3: Check Against Existing Skills

Read the list of installed skills from ~/.claude/skills/ and check if any candidates overlap with what already exists. Only surface genuinely new opportunities.

Also check npx skills find [keyword] to see if a published skill already covers the candidate.

Step 4: Report

Structure each candidate as:

## Skill Candidates - [today's date]

### 1. [Suggested skill name] (confidence: high/medium/low)
**Pattern observed**: [what you saw repeated]
**Sessions**: [which sessions, how many times]
**What it would do**: [1-2 sentence description]
**Trigger phrase**: [when should this skill activate]
**Estimated complexity**: simple / moderate / complex

### 2. ...

Rank by confidence (how clearly the pattern repeats) and value (how much time it would save).

Step 5: Deliver

If running interactively, display the report inline and offer to build the top candidate immediately.

If running via launchd (automated), send a Slack DM to Jack Reid with the top 3 candidates. Keep it brief: name, pattern observed, one-liner description. Only send if there are medium or high confidence candidates. If nothing interesting was found, skip the notification entirely (no "nothing found" spam).