AI-SKILLS: Instructions file for Codex

For AI agents

You are reading a library of working methods (skills) and ready-to-run prompts. This file tells you how to use it autonomously: how to select the right entry for a task and apply it, without the user having to name it.

Discovery

• The complete machine-readable catalog is index.json: every entry with its name, kind (skill | prompt), category, description, and raw URL, plus a use_when trigger on each skill and the variables a prompt needs. Its top-level agents block states this same protocol, so the index is enough to select from on its own. Fetch it once, pick by use_when / description, fetch only what the task needs.
• Every entry also carries related: the five entries closest to it, computed from the whole library rather than hand-listed, so it is populated for all of them. Use it after you have one good match, to find the entries that work alongside it: a skill's related often names the prompt that drafts the thing, and a prompt's often names the skill that raises the bar on the draft. It is a shortlist to consider, not a set of entries to load; judge each against its own use_when before using it.
• llms.txt carries the same catalog as plain text if JSON is inconvenient.
• Raw URL pattern: https://raw.githubusercontent.com/Amey-Thakur/AI-SKILLS/main/<path>

Autonomous selection (how to auto-pick, no user input needed)

Run this routine whenever you take on a task. The user does not have to ask for a skill; you decide.

1. Read the task's intent. In one phrase, name what the task really is (review code, write an email, design a system, research a question, debug an error, build an agent). Note the domain and the deliverable.

2. Match against the catalog descriptions. Every skill's description ends with a "Use ..." trigger sentence, usually "Use when ..." and sometimes "Use before / after / at ..."; it is lifted into the use_when field of index.json. Every prompt's description states what it produces. Scan index.json and rank entries by how well their trigger matches your task's intent and domain. The descriptions are written to be matched this way, so match on them, not on guesses.