Conversation To Skill Miner
Overview
Convert messy conversation history into a safe skill backlog. The goal is not to
publish every idea; it is to identify repeatable workflows, remove private or
high-risk material, and package only the candidates that can be useful without
leaking personal context.
Read references/publishability-rubric.md when deciding whether a candidate can
be published.
Workflow
Define the scan boundary.
- Prefer durable archives:
conversation-memory/, conversation-archive/,
daily-ideas/, project READMEs, and existing local skill folders.
- Do not scan private messaging exports, browser profiles, secrets, or raw
logs unless the user explicitly asks and the output will remain private.
Find repeated workflows.
- Look for phrases such as
skill, workflow, pipeline, SOP, agent,
继续, 整理, 归档, 自动化, weekly, GitHub, PPT, paper,
Simulink, and screen.
- Prefer workflows that have been requested more than once or already have
artifacts, scripts, reports, or topic-memory entries.
Classify each candidate.
publish_now: generic, safe, repeatable, and already proven by at least
one real task.review_first: promising but may contain personal data, domain risk,
weak examples, missing validation, or unclear audience.keep_private: depends on private contacts, credentials, sensitive
profiling, personal health/finance decisions, or private chat content.
Package only safe candidates.
- Strip raw conversation text and personal examples.
- Keep general workflows, safety boundaries, trigger phrases, and reusable
scripts.
- Add references only when they are public, lawful, and one level deep from
SKILL.md.
Validate and report.
- Run
quick_validate.py for each skill.
- Produce a concise report with candidate name, source pointer, decision,
risk, next packaging action, and recommended repository.
- Keep generated reports local unless they are manually sanitized.
Script
Use the bundled scanner for a first pass:
python3 scripts/scan_conversations_for_skills.py \
--workspace /path/to/Codex \
--output /path/to/weekly-skill-candidates.md
The script outputs a markdown report and a sibling JSON file. Treat its result
as a triage aid; the agent still makes final publishability decisions.
Weekly Review Rule
For weekly operation, schedule only the scan/report step automatically. Do not
auto-publish directly from private conversations. Public upload should happen
after an explicit review pass that confirms no secrets, raw chats, private
profiles, or copyrighted material are included.
1---2name: conversation-to-skill-miner3description: Mine conversation archives, topic memory, idea inboxes, and project notes for reusable agent-skill candidates, then classify which should be packaged, reviewed, or kept private. Use when Codex is asked to search chats/conversations for possible skills, turn repeated workflows into skills, run a weekly skill-mining review, or prepare a safe public skill backlog from private conversation history.4---56# Conversation To Skill Miner78## Overview910Convert messy conversation history into a safe skill backlog. The goal is not to11publish every idea; it is to identify repeatable workflows, remove private or12high-risk material, and package only the candidates that can be useful without13leaking personal context.1415Read `references/publishability-rubric.md` when deciding whether a candidate can16be published.1718## Workflow19201. Define the scan boundary.21 - Prefer durable archives: `conversation-memory/`, `conversation-archive/`,22 `daily-ideas/`, project READMEs, and existing local skill folders.23 - Do not scan private messaging exports, browser profiles, secrets, or raw24 logs unless the user explicitly asks and the output will remain private.25262. Find repeated workflows.27 - Look for phrases such as `skill`, `workflow`, `pipeline`, `SOP`, `agent`,28 `继续`, `整理`, `归档`, `自动化`, `weekly`, `GitHub`, `PPT`, `paper`,29 `Simulink`, and `screen`.30 - Prefer workflows that have been requested more than once or already have31 artifacts, scripts, reports, or topic-memory entries.32333. Classify each candidate.34 - `publish_now`: generic, safe, repeatable, and already proven by at least35 one real task.36 - `review_first`: promising but may contain personal data, domain risk,37 weak examples, missing validation, or unclear audience.38 - `keep_private`: depends on private contacts, credentials, sensitive39 profiling, personal health/finance decisions, or private chat content.40414. Package only safe candidates.42 - Strip raw conversation text and personal examples.43 - Keep general workflows, safety boundaries, trigger phrases, and reusable44 scripts.45 - Add references only when they are public, lawful, and one level deep from46 `SKILL.md`.47485. Validate and report.49 - Run `quick_validate.py` for each skill.50 - Produce a concise report with candidate name, source pointer, decision,51 risk, next packaging action, and recommended repository.52 - Keep generated reports local unless they are manually sanitized.5354## Script5556Use the bundled scanner for a first pass:5758```bash59python3 scripts/scan_conversations_for_skills.py \60 --workspace /path/to/Codex \61 --output /path/to/weekly-skill-candidates.md62```6364The script outputs a markdown report and a sibling JSON file. Treat its result65as a triage aid; the agent still makes final publishability decisions.6667## Weekly Review Rule6869For weekly operation, schedule only the scan/report step automatically. Do not70auto-publish directly from private conversations. Public upload should happen71after an explicit review pass that confirms no secrets, raw chats, private72profiles, or copyrighted material are included.
walterkken