# Context calibration: applying the diagnosis

Boris's post describes shared knowledge carriers around [09:56](https://www.youtube.com/watch?v=xmGY276gEFY&t=596s). Theo endorses the direction while disputing stronger claims around [10:31](https://www.youtube.com/watch?v=xmGY276gEFY&t=631s). His newcomer-question practice starts around [11:14](https://www.youtube.com/watch?v=xmGY276gEFY&t=674s).

## Use a probe when the cause is unclear

Choose a representative task from the observed misses. Supply the minimum extra prompt explanation while retaining standing instructions, admitted tools and runtime safeguards. Record the expected decision, observed behavior and missing fact. Repeat only if needed to resolve the diagnosis; no run count or sentence-level proof quota is implied.

A code change needs an appropriate behavioral check. An instruction-only edit needs comparison to the actual requirements and, where loading matters, verification that the changed instruction is actually loaded. A skill file on disk does not prove the agent's runtime loaded it.

## Place the knowledge close to the decision

If agents repeatedly simplify an intentionally unusual query, a local comment can preserve the measured reason and point to the relevant regression test. If the problem is missing domain vocabulary, improve the glossary and its decision links. If it is a compiler-expressible invariant, use the type system. A cross-cutting owner preference belongs in the root guidance on routes verified to load it.

A useful navigation map names the task, entry point and reason. Keep implementation locations accurate, but avoid replacing judgment with a directory inventory. The video's criticism of bare maps starts around [16:38](https://www.youtube.com/watch?v=xmGY276gEFY&t=998s).

## Respect recorded decisions and preserve expertise

Theo advocates human ownership of root instructions around [12:29](https://www.youtube.com/watch?v=xmGY276gEFY&t=749s). Authorized agent drafting is compatible with that ownership: apply agreed decisions and bounded wording improvements, and ask only for consequential unresolved direction. Never invent a product refusal or elevate old guidance above a later authorized change.

Resolve conflicting instructions from their scope and provenance. Preserve exact originals, keep uncertain source versions inactive when migrating, and use your project's existing change journal for instruction changes. A single inconclusive probe does not justify deleting expertise.