Skill Incubator

Skill Incubator

Goal

Turn repeated work into durable Codex capabilities without bloating the skill system.

This skill does not silently create production skills. It detects reusable patterns, drafts skill candidates, and waits for explicit user approval before installing or updating any skill.

When To Run

Run this at the end of a task when one or more is true:

• The same kind of task has appeared 3 or more times.
• The user says "以后这类事就这样做", "沉淀一下", "写成 skill", or similar.
• A workflow now has stable inputs, outputs, tools, validation, and failure handling.
• The task required a non-obvious sequence that would be expensive to rediscover.
• A previous memory rule has grown into a repeatable operating procedure.

Do not run it for one-off errands, pure opinions, volatile facts, or tasks whose success depended on a temporary website state.

Classification

Classify the learning into exactly one bucket:

| Bucket | Use When | Output | |---|---|---| | ad_hoc | One-time fact, temporary path, or narrow event | No skill candidate | | memory_rule | Stable preference or decision rule, but no multi-step workflow | Memory note only if user explicitly asked to update memory | | workflow_note | Reusable sequence, but still too context-dependent or under-tested | Candidate note, not a skill | | skill_candidate | Trigger, inputs, steps, validation, and boundaries are clear | Draft skill candidate |

Prefer workflow_note until there is enough evidence. A bad skill is worse than no skill because it hijacks future routing.

Promotion Gates

A task can become a skill_candidate only if all gates pass:

1. Trigger clarity: the future user phrasing is predictable.
2. Stable contract: inputs and final deliverables are clear.
3. Operational steps: the workflow can be written as ordered actions.
4. Validation: there is a concrete way to check success.
5. Boundary: the skill says when not to run.
6. Repeat evidence: at least 2 real runs, or explicit user instruction to create it now.
7. Risk control: no silent deletion, account action, legal/financial commitment, or long-term memory write without confirmation.

Candidate Queue

Write candidates to:

~/.codex/skill_candidates/

Use filenames:

YYYY-MM-DD_.md

Use scripts/new_candidate.py if available:

python3 ~/.codex/skills/skill-incubator/scripts/new_candidate.py \
  --name "候选 skill 名" \
  --trigger "用户怎么说会触发" \
  --why "为什么值得沉淀" \
  --evidence "真实任务证据或路径" \
  --boundary "什么时候不能用" \
  --validation "怎么验收"

Candidate Template

Every candidate must contain:

---
type: skill_candidate
status: pending_user_review
created_at: YYYY-MM-DD
---

# Candidate: 

## Why This Should Exist
- Repeated pain:
- Existing evidence:
- Expected leverage:

## Trigger
- Use when:
- User phrases:
- Do not use when:

## Input Contract
- Required inputs:
- Optional inputs:
- Missing input behavior:

## Workflow
1.
2.
3.

## Validation
- Success checks:
- Failure signals:
- Artifacts to inspect:

## Skill Shape
- Target runtime: Codex / Claude / shared agents
- Suggested path:
- Needs script: yes/no
- Needs references: yes/no

## Risks
- Over-trigger risk:
- Privacy/account risk:
- Staleness risk:

## Install Decision
- Approve:
- Reject:
- Merge into existing skill:

Installing Or Updating A Skill

Only install after the user approves.

1. Check whether an existing skill already covers it:

• ~/.codex/skills/
• ~/.agents/skills/
• ~/.claude/skills/

1. If covered, update the existing skill instead of creating a duplicate.
2. If new, create a minimal skill folder:

• SKILL.md required
• agents/openai.yaml recommended for Codex-visible skills
• scripts/ only when deterministic execution is needed
• references/ only when detailed context must stay out of SKILL.md

1. Keep SKILL.md under 500 lines.
2. Add a completion checklist with real validation.
3. Do not update memory unless the user explicitly asks to update memory.

End-Of-Task Prompt

When a task appears reusable but not yet approved, end with a short line:

这类任务已经具备 skill 候选特征；我可以把它写入 skill_candidates，等你确认后再晋级。

If the task clearly does not qualify:

这次更像一次性任务，我不沉淀成 skill。

Completion Checklist

• Classified as ad_hoc, memory_rule, workflow_note, or skill_candidate.
• Did not silently write long-term memory.
• Checked for existing overlapping skills before proposing a new one.
• Candidate includes trigger, boundaries, workflow, validation, and risks.
• Candidate is queued under ~/.codex/skill_candidates/ unless user requested direct installation.
• User is told whether this is only a candidate or an installed skill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: smwswk
• Source: smwswk/xiaoming-agent-skills
• License: MIT
• Homepage: https://smwswk.github.io/skill-store/

Install and usage instructions live in the source repository linked above.