AgentSkillsCN

gen-mindset

挖掘 Codex 会话，提炼出你最顶尖的软件设计启发式经验（无需引用，也无需佐证）。

SKILL.md
--- frontmatter
name: gen-mindset
description: "Mine Codex sessions to synthesize your top software design heuristics (no quotes, no evidence)."
metadata:
  short-description: "Generate a design mindset from session logs"

gen-mindset

Intent

Synthesize a ranked list of software design heuristics by mining your Codex session logs.

This skill is optimized for:

  • •Design-only heuristics (types/invariants, testing/validation, error handling, readability/complexity, abstraction discipline).
  • •Signal-weighted ranking: what shows up often and what the user appears to approve.
  • •Low leakage output: list only (no scores, no quotes, no file paths).

When to use

  • •You want to distill the design principles that actually recur in your working style.
  • •You’re updating or drafting another mindset skill (e.g., $tk) and want evidence from your own sessions.

Defaults

  • •Sessions corpus: ~/.codex/sessions
  • •Output: top 20 heuristics
  • •Evidence: suppressed (no counts, no quotes)
  • •Approval detection: explicit positives + short approvals + next-step directives

Inputs

Ask only if the user hasn’t specified:

  1. •top N (default 20)
  2. •sessions path override (default ~/.codex/sessions)

Do not ask about scoring/evidence unless the user requests it.

Output contract

  • •Output only a bullet list of the top N heuristics.
  • •Do not include scores, weights, quotes, or snippets.
  • •Do not include paths into sessions.
  • •If fewer than N heuristics are detected, output what you have and state that fewer were found.

Method (what the script does)

  1. •Scan ~/.codex/sessions/**/*.jsonl.
  2. •Extract assistant messages (stripping the leading Echo: line when present).
  3. •Match against a curated set of design-heuristic patterns.
  4. •Weight by:
    • •frequency across assistant messages
    • •plus additional weight when the next user message looks like approval

Approval detection rules

A user message counts as approval if it is not a clear complaint and matches any:

  • •Approval keywords: thanks, great, lgtm, ok, yes, continue, etc.
  • •Short acknowledgements (very short messages)
  • •Next-step directives (imperative “do X next” instructions)

Execution

Run the helper script and return the list.

Recommended invocation:

  • •python3 ~/.codex/skills/gen-mindset/scripts/gen-mindset.py --top 20

If you prefer uv:

  • •uv run python ~/.codex/skills/gen-mindset/scripts/gen-mindset.py --top 20

Hard rules

  • •Explicit-only skill: do not auto-trigger.
  • •Don’t invent principles that aren’t in the corpus.
  • •Keep the list phrased as heuristics (“Prefer…”, “Treat…”, “When…”) not slogans.
  • •No scope creep: don’t update other skills unless asked.