suggest-skills

name	suggest-skills
description	Analyze past Claude Code and Codex sessions to suggest new skills, workflow improvements, or changes to existing skills. Identifies repetitive patterns and unmet automation opportunities.
argument-hint	[--days <N>] [--project <name>]
allowed-tools	Read, Bash, Glob, Grep, Agent

Analyze recent conversation history across Claude Code and Codex to identify opportunities for new skills, workflow improvements, or changes to existing skills.

@shared/SESSION_PARSING.md

Arguments

• --days <N> — how far back to look (default: 14)
• --project <name> — focus on sessions from a specific project

Workflow

1. Inventory Existing Skills

Read the skill directories to build a list of what already exists:

ls ~/.claude/skills/ | grep -v shared

For each skill, read just the frontmatter (first 6 lines) of SKILL.md to get the name and description.

2. Discover Recent Sessions

Find session files modified in the last N days:

find ~/.claude/projects -name "*.jsonl" -not -path "*/subagents/*" -mtime -${DAYS} -type f
find ~/.codex/sessions -name "*.jsonl" -mtime -${DAYS} -type f

If --project is set, filter to matching paths.

3. Sample Sessions

For efficiency, don't read every session end-to-end. Instead:

1. Sort files by modification time (newest first)
2. Take up to 20 sessions
3. For each session, extract only user messages (skip system messages, tool results, and assistant responses). These reveal what the user is asking for and doing repeatedly.

Use a Bash one-liner per file to extract user text:

python3 -c "
import json, sys
with open(sys.argv[1]) as f:
    for line in f:
        try:
            obj = json.loads(line)
            if obj.get('type') == 'user':
                msg = obj.get('message', {})
                content = msg.get('content', '') if isinstance(msg, dict) else ''
                if isinstance(content, str) and len(content) > 20 and not content.startswith('<'):
                    print(content[:500])
            elif obj.get('type') == 'response_item':
                p = obj.get('payload', {})
                if p.get('role') == 'user':
                    for b in (p.get('content') or []):
                        t = b.get('text', '')
                        if len(t) > 20 and not t.startswith('<') and not t.startswith('#'):
                            print(t[:500])
        except: pass
" FILE

4. Analyze Patterns

With the sampled user messages, identify:

A. Repetitive requests — Things the user asks for repeatedly across sessions that could be automated into a skill. Look for:

• Similar phrasing across sessions
• Multi-step workflows the user orchestrates manually
• Recurring "can you..." / "I need to..." patterns

B. Tool usage patterns — If assistant messages show repeated tool call sequences (e.g., always running the same grep → read → edit pattern), that's a candidate for a skill.

C. Pain points — Moments where the user expresses frustration, restarts, or corrects the assistant. These suggest the current workflow is inadequate.

D. Cross-session themes — Topics or tasks that span multiple sessions, indicating ongoing work that would benefit from better tooling.

5. Cross-Reference with Existing Skills

For each identified pattern, check whether an existing skill already addresses it. If a skill exists but the user doesn't seem to use it, that's also worth noting (discoverability issue or skill gap).

6. Generate Recommendations

Present findings as a ranked list:

# Skill Suggestions

Based on analysis of N sessions over the last D days.

## New Skills

### 1. /SKILL-NAME — SHORT DESCRIPTION
**Pattern observed:** What the user does repeatedly
**Sessions:** List 2-3 session dates/projects where this appeared
**Effort:** Low / Medium / High
**Impact:** How much time/friction this would save

### 2. ...

## Improvements to Existing Skills

### 1. /EXISTING-SKILL — WHAT TO CHANGE
**Gap observed:** What the skill doesn't handle that the user needs
**Sessions:** Where this gap appeared

## Workflow Observations

- Any cross-cutting insights about how the user works that don't map to a single skill

Rank by impact (time saved × frequency). Limit to 5-7 recommendations total — quality over quantity.