skill-suggester – skill Agent & Tự động hóa cho agent AI

What I do

Reads your opencode prompt history, finds repeated multi-step workflows, and recommends skill-worthy candidates. Saves you from having the same conversation twice.

When to Use

Use this skill when the user wants to mine opencode prompt history for repeated workflows, recurring unmet needs, or candidates for new reusable skills.

How to invoke

Run /skill skill-suggester to scan the full history. Optionally pass --since <date> (e.g. --since 2026-05-01) to limit the window.

Analysis method

1. Locate prompt history files at ~/.local/state/opencode/prompt-history*.jsonl

2. Parse each entry's message content

3. Score for skill potential by looking for:

• Repetition: similar phrasing or topic used 3+ times ("scan all repos", "check my inbox")
• Multi-step sequences: a request that required 5+ tool calls to complete
• Unsupported requests: things you asked for that don't have a dedicated skill yet
• Workaround patterns: instructions you give every time instead of a one-shot command

4. For each candidate, note:

• How many times the pattern appeared
• How many tool calls it consumed
• The estimated time savings if it were a skill

Output format

## Skill Candidates (last N entries)

### 1. "<candidate name>" (PRIORITY)
- **Pattern**: <what you keep asking for>
- **Frequency**: X times in history
- **Avg complexity**: Y tool calls per instance
- **Estimated savings**: ~Z minutes/week
- **Evidence**:
  - "[excerpt from prompt history]"
  - "[another excerpt]"
- **Recommendation**: <create as skill | add as command template | not worth it>

### 2. ...

Key rules

• Only flag patterns that happen more than twice. One-offs are not skills.
• Use minimal redacted excerpts only when needed for evidence; summarize sensitive prompts and never disclose secrets. Analyze only the user-authorized history window locally. Tool-call counts and time savings must be marked unavailable or estimated when prompt-only records do not contain them.
• Rate each candidate: high (clear ROI, use weekly), medium (nice to have), low (rare but worth noting).
• If nothing qualifies, say so and explain why.
• After presenting candidates, ask if you want to create any of them.

Limitations

• Prompt history can contain sensitive local context; summarize patterns without exposing unnecessary private excerpts.
• Recommendations are suggestions only and still need human review before creating or publishing a new skill.