find-skills ยท diff
git:20260902.32224e7 to git:20260906.90262a6
13 added, 0 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: find-skills
description: Discover installable agent skills for a concrete capability, verify their source and contents, compare them with already installed skills, and recommend only additions that materially improve the current skill set. Use when the user asks whether a skill exists, wants skill recommendations, or wants to extend Codex.
---
# Find skills
Find the smallest trustworthy addition that solves the user's recurring need. Do not recommend a skill merely because its title matches the query.
## Discover
1. Identify the task, target environment, and behavior the user lacks.
2. Inventory installed skills before searching. Treat overlapping triggers and conflicting instructions as costs.
3. Search first-party repositories, reputable ecosystem indexes such as [skills.sh](https://skills.sh/), and the source repository named by the user.
4. Inspect the actual `SKILL.md` and any scripts, references, hooks, agents, or manifests that affect behavior.
Use current web sources for repository contents, popularity, maintenance status, and compatibility claims. Install counts and stars are weak signals, not proof of quality.
+ When an addition overlaps installed skills or context cost motivates the request,
+ read [usage and context evidence](references/usage-and-context.md). Check whether
+ the existing skill is discoverable before treating missing usage as a capability
+ gap. Separate installed files, host-visible listings, and observed invocations.
+
## Evaluate
Recommend a skill only when it adds non-obvious procedures, maintained domain knowledge, deterministic tooling, or reusable assets. Check:
- source ownership and recent maintenance
- license and bundled executable code
- trigger precision and overlap with installed skills
- assumptions about Claude Code, Codex, or another host
- destructive actions, network access, credentials, and external side effects
- whether a small rewrite is required for the target environment
Reject generic prompt collections, stale migrations, duplicate workflows, and host-specific skills that would misroute Codex.
+
+ When revisiting an audited candidate, recover its source revision, verdict, and
+ reason from existing records. Recheck the facts that could change that verdict,
+ such as compatibility, license, or overlap with newly installed skills. Separate
+ fresh evidence from repeated promotion, and trace apparent corroboration to its
+ original source. Retain the earlier decision and explain what supersedes it;
+ an old assessment is not automatically false or current. Correct a mistaken
+ source identity or evaluation criterion where it originated, within task scope.
## Report
Give a short verdict for each serious candidate: what it adds, what overlaps, compatibility risks, and the source link. Say plainly when nothing is worth installing.
## Install
Do not install until the user asks. In Codex, use the native `skill-installer` workflow for GitHub paths or curated skills. Preserve the complete skill directory, not only `SKILL.md`, and validate the installed result. Use another package manager only when the user requests it or the source requires it.
After installation, state the destination and that the skill becomes available on the next turn.