Context calibration turns repeated agent confusion into one measured change supported by observed misses. You can then check whether the next matching task receives that missing knowledge.
Repeated confusion makes the skill useful
Use this skill when an agent repeatedly misses project knowledge during otherwise authorized repository work. It also helps when you need to review steering instructions or test a fresh agent’s understanding. Choose another method for ordinary coding, human documentation, or creating a new packaged skill.
Your evidence defines the diagnosis
Give the agent recent mistakes, repeated questions, review rejections, or one representative task to probe. It also needs access to the places where the missing knowledge could be recorded. When the cause remains unclear, the agent adds only the smallest extra explanation needed. The probe keeps standing instructions, approved tools, safeguards, and configured access unchanged throughout. It records the expected decision, observed behavior, and the fact still missing afterward.
Miss class determines the carrier
The agent sorts each miss into a missing fact, invalid shape, unguarded pattern, or missing judgment. That class points toward a local comment, structural rule, automated check, reusable skill, or standing instruction. A local fact stays beside its code, while reusable judgment belongs in selectively loaded guidance. Cross-cutting preferences belong in standing instructions only after confirming the relevant route loads them. Theo Browne gives an example where a standing rule stops an agent on a known unwanted request in his video.
Checks show whether the change landed
The result includes each miss, its class, the chosen carrier, and every supported change actually made. Items needing a new policy or unresolved product decision remain drafts for your decision. For changed instructions, compare the wording with the recorded miss and confirm the intended route loads it. Code or configuration changes need behavioral checks at the boundary they were meant to affect. Shreya Shankar describes feeding review findings back into the skill that produced the work during her evaluation demonstration. You can see the method working when the next comparable attempt receives the chosen knowledge.
Boundaries keep the change proportional
The agent changes only items supported by observed misses, without broadening them into unstated general rules. It asks you only when evidence cannot resolve a consequential policy, refusal, or product direction. One inconclusive probe never justifies deleting earlier expertise or replacing a working instruction set. Guidance remains advisory, while automated checks and structural rules cover only their configured boundaries. The skill avoids padding standing instructions, encoding one-off events, or adding guidance where a local fact suffices.
The sources support evidence-led changes
These examples support observing failures before turning them into reusable steering for later work. Theo Browne’s example concerns one explicit stopping rule, without proving that prose enforces every boundary. Shreya Shankar’s demonstration shows one feedback loop, without establishing a universal placement rule. The skill’s remaining behavior and limits come directly from its maintained instructions and supporting notes.
- Theo Browne, video, A Message for Passionate Devs, at 13:33. "If they ask for this, stop and tell them no"
- Shreya Shankar, video, Using AI for AI Evals, at 14:00. "I try to encode that back into the skill"
Run the skill on one repeated agent mistake, with access to every plausible knowledge carrier.