One thing that didn't make it into the post: the context file spec uses semantic section tags ([PROJECT], [ROLE], [THESIS], [CERT], etc.) on every heading specifically so an agent can navigate to the relevant section without reading the full file. A CV generation task loads only [ROLE] and [PROJECT] entries. An interview prep task loads [THESIS] and [COMPETITION] entries. The token budget stays predictable regardless of how large the file grows over time.
There is also a VERIFIED FACTS HTML comment in the scope declaration that lists every atomic fact that must never be hallucinated: grades, competition rankings, certification IDs, dates. It is invisible in rendered Markdown but visible to the agent reading raw text. If the fact is not in that comment, the agent is not allowed to state it as certain.
The provider adapter problem was genuinely annoying to solve. Codex, Claude Code, Gemini CLI, and OpenCode each have different conventions for where skills live, how they are activated, and what naming they expect. The install CLI resolves that per-provider rather than shipping four separate packages.
Still early and the spec is not frozen. Feedback on the file structure or the provider support is welcome.
One thing that didn't make it into the post: the context file spec uses semantic section tags ([PROJECT], [ROLE], [THESIS], [CERT], etc.) on every heading specifically so an agent can navigate to the relevant section without reading the full file. A CV generation task loads only [ROLE] and [PROJECT] entries. An interview prep task loads [THESIS] and [COMPETITION] entries. The token budget stays predictable regardless of how large the file grows over time.
There is also a VERIFIED FACTS HTML comment in the scope declaration that lists every atomic fact that must never be hallucinated: grades, competition rankings, certification IDs, dates. It is invisible in rendered Markdown but visible to the agent reading raw text. If the fact is not in that comment, the agent is not allowed to state it as certain.
The provider adapter problem was genuinely annoying to solve. Codex, Claude Code, Gemini CLI, and OpenCode each have different conventions for where skills live, how they are activated, and what naming they expect. The install CLI resolves that per-provider rather than shipping four separate packages.
Still early and the spec is not frozen. Feedback on the file structure or the provider support is welcome.