Read this as What does the model see before it chooses a skill?
- Failure Trap
- Eager-loading every skill body and burying the relevant instruction in irrelevant context.
- Decision Rule
- Keep descriptions on the shelf, open exactly one body on demand, and let unused bodies drop out.
The model sees a shelf of descriptions
Progressive disclosure starts with short skill descriptions in context, not full instruction bodies.
- The description is the public API.
- The body stays on disk.
- The candidate set stays small enough to scan.
The task matches one description
The model chooses a skill by reading descriptions that name the action, input, and boundary.
- Good descriptions are verb-first.
- Examples disambiguate adjacent skills.
- Generic descriptions are invisible.
Only the chosen body opens
After selection, the harness loads the full skill body into the working context.
- The body can be detailed.
- It arrives only when useful.
- The harness owns the gate.
Unused bodies drop out
Other skill bodies remain out of context, preserving headroom for the actual task and files.
- The catalog can be large.
- The working set stays small.
- Capability no longer equals always-on tokens.
Eager loading pays the 35K trap
The source example uses ten skills with bodies around 3,500 tokens each: eager loading spends about 35K tokens immediately.
- Irrelevant bodies still tax attention.
- The right instruction is harder to find.
- The cited eager run scored 29 percent.
Lazy loading keeps the shelf cheap
Descriptions plus one opened body in the source example consume about 5.5K tokens and preserve the path to the 95 percent run.
- Same model.
- Same instructions.
- Only retrieval architecture changed.