TanStack Intent + npx skills: distribute leadtype as an agent capability

[KayleeWilliams]

Why

Our evals show agents discover docs only ~29% of the time when not explicitly pointed at them (2026-05-31 run, discovery arm). Agent-skill channels fix this by making the capability auto-discoverable and using progressive disclosure (no context bloat).

This issue tracks distributing leadtype through emerging agent-capability channels as an optional/additive path. The universal route (npm install + a root AGENTS.md pointer, #66) stays the fallback for tools that don't support these.

TanStack Intent — the main target (https://tanstack.com/intent/latest)

Intent is, essentially, a productized version of what leadtype already does: ship current, versioned procedural knowledge alongside an npm package so agents don't rely on stale training data. Mechanics:

• Agent Skills shipped in the npm package, auto-discovered from node_modules — no per-library manual setup. (This is a cleaner fix for our discovery gap than the manual root pointer in Make the root AGENTS.md pointer the headline setup (and scaffold it) #66.)
• A Skills Registry for discovery, staleness detection that flags when skill docs drift from source, and a CLI (intent scaffold, intent validate, intent install).

Relationship to leadtype: strong overlap. Rather than compete, leadtype should likely emit an Intent skill from the same MDX source (match intent scaffold's format) so leadtype-generated docs become an auto-discovered Intent skill. Intent's staleness check also overlaps with leadtype's existing lint/validation — worth aligning, not duplicating.

Also: npx skills / general skill registries

Same idea, different ecosystem — install a leadtype skill into a skill-aware tool (npx skills add inthhq/leadtype). Lower priority than Intent unless our audience uses it.

Two artifact types — emit from the same MDX source

1. Reference skill — progressive-disclosure leadtype docs (one-line description always visible; body/topic files load only when relevant). Keeps context lean even in monorepo-size projects, where reading the llms-full.txt monolith is the real bloat risk.
2. Setup skill — install it, say "set up leadtype docs," and it runs the wiring (leadtype generate --bundle, configure). Thin wrapper over the existing CLI.

Caveats / next

• Validate adoption + the auto-discovery claim. Intent's "agents find it automatically" depends on agent tools implementing the Intent runtime — confirm our target tools actually do before investing.
• Keep the universal AGENTS.md + pointer path (Make the root AGENTS.md pointer the headline setup (and scaffold it) #66) as the fallback for non-Intent tools.
• Spike: scaffold an Intent skill from leadtype's docs (intent scaffold), test auto-discovery end-to-end, and check whether Intent's staleness tooling can reuse leadtype lint.

Context: discovery finding in FINDINGS.md / docs/reference/evals.mdx; universal pointer in #66.

🤖 Generated with Claude Code