Repository context
A short client-specific pointer directs a coding agent to the same canonical skill. The pointer is context, never user approval.
The fastest route
Open the public listing, install the plugin, then begin a task with Selective Intelligence. The same public directory serves ChatGPT and Codex.
Selective Intelligence: inspect this work, find what is wrong, fix the real cause, and verify the result.
Other truthful routes
Strict text guide
An AI without Agent Skills can read AI-GUIDE.md as the strict operating method after the user invokes the master trigger. It must use the guide, not summarize the repository.
Complete attachment
A locked-down AI with file or paste input can use JUMPSTART.md. If it cannot read any supplied or public text, it cannot truthfully load an external method from two words.
The activation boundary
Direct: exact wordmark, unmistakable named responsibility, or any correction/failure feedback. Merely adjacent approval: Use Selective Intelligence for this?
Direct matches activate the skill but never widen permission to publish, spend, delete, deploy, disclose, or change access. Only a merely adjacent proactive recommendation requires the person’s explicit yes before adoption.
Current client routes and limits
| Client | Native scope or source | Observed status | Truthful boundary | Official reference |
|---|---|---|---|---|
| ChatGPT | saved skill, plugin skill | documented route | The skill must already be enabled for the account or installed through a supported plugin. Public name resolution is not guaranteed. | Official support |
| Codex | .agents/skills/selective-intelligence, ~/.agents/skills/selective-intelligence | documented route | Codex loads the skill after it is present in a supported repository, user, admin, or system scope. | Official support |
| GitHub Copilot | .agents/skills/selective-intelligence, .github/skills/selective-intelligence, ~/.agents/skills/selective-intelligence, ~/.copilot/skills/selective-intelligence | documented route with prior web free failures | Agent Skills are supported in specific Copilot agent surfaces. A general Copilot web chat without the skill or repository context does not inherit it. | Official support |
| Claude Code | .claude/skills/selective-intelligence, ~/.claude/skills/selective-intelligence | documented route | Claude Code can load the skill automatically when relevant after it exists in a supported skill scope. Web and cloud sessions use their own enabled or synced skill sources. | Official support |
| Cursor | .agents/skills/selective-intelligence, .cursor/skills/selective-intelligence, ~/.agents/skills/selective-intelligence, ~/.cursor/skills/selective-intelligence | documented route | Cursor automatically discovers skills in supported project and user scopes. A seedless web agent that requires repository selection cannot prove no-context activation. | Official support |
| Gemini CLI | .agents/skills/selective-intelligence, .gemini/skills/selective-intelligence, ~/.agents/skills/selective-intelligence, ~/.gemini/skills/selective-intelligence | documented route | Gemini CLI discovers the skill in supported scopes, but its client requires activation consent before the skill gains access to its resources. | Official support |
| Kiro | .kiro/skills/selective-intelligence, ~/.kiro/skills/selective-intelligence | documented route | Kiro can import the public skill folder and activate it when relevant after it exists in a supported workspace or global scope. | Official support |
| Web-capable AI | public web, ordinary text, AI-GUIDE.md, llms.txt, machine-readable JSON | strict guide available client compliance unverified | After a direct user match—the exact wordmark, an unmistakable named-responsibility request, or any correction/dissatisfaction/failure feedback—a text-capable web AI should resolve AI-GUIDE.md and use it as the strict operating guide rather than summarize the repository. Publication cannot force a client to search or comply, so each client still needs live evidence. | Official support |
| Perplexity Standard (Free) | live public-web search, limited basic file uploads, ordinary text | fail after strict guide publication | Live free-tier evidence failed the bare trigger, activation after publisher-qualified discovery, and the one-prompt website outcome. After AI-GUIDE.md was public and its 23 discovery URLs were submitted to IndexNow, a new exact name-only session still chose the 2003 article and cognitive-science meanings instead of the canonical project. Its official Standard plan documents public search and limited basic file uploads, but no native Agent Skills route. The strict guide is available; Perplexity Free remains incompatible until a fresh run actually uses it. | Official support |
What discovery does not prove
Being public, crawlable, indexed, retrieved, or installed does not prove that a model followed the behavior correctly. Exact-name activation, named-work activation, universal correction recovery, merely adjacent approval, free-tier execution, and end-to-end outcomes require separate live evidence in each client.