Learn From Corrections
Update the relevant skills so future agents can avoid the same mistake. Work backward from the user's corrections to the skills that shaped the exchange. A better sample response helps judge a skill change; carry justified findings into the skills themselves within the user's authorized scope.
A skill is its SKILL.md and the supporting references, scripts, and
configuration it owns. Read the relevant parts and change where the behavior
is actually governed. AGENTS.md and other active instructions can explain a
conflict or show that a lesson belongs outside a skill; distinguish those
findings from skill changes. The conversation's prompt is evidence of intent,
not the default destination for the improvement.
Establish the purpose and evidence
Use dialectic to establish the intended outcome before diagnosing the correction or suggesting changes. State what the user wants the reviewed conversation or workflow to help them understand, decide, or do. Put that proposed intent in a blockquote and bold the words that state it; emphasizing a question such as "Is that the job you want?" leaves the intent itself hard to find. Let the user sharpen that understanding before building on it. Carry forward intent already established; a settled correction can proceed directly.
Once the purpose is understood, examine one concrete issue or suggestion at a time against it. Ground the diagnosis in the conversation and the skills actually invoked, and let the user's reactions shape the next attempt. Develop the corresponding skill change once the diagnosis is understood. Revisit it if considering the change exposes something it missed.
Read the referenced conversation through the available conversation tools. For
Codex chat references, use read_thread and follow pagination far enough to
understand the relevant requests, responses, corrections, and resolutions.
Use the record before asking the user to reconstruct it. They may identify a
correction they should never have needed to make without knowing the missing
instruction.
Verify which skills were actually invoked, alongside other active instructions. Distinguish loaded guidance from skills merely available or mentioned. Recover the wording in effect at the relevant turn when possible; separate later changes from the baseline. State gaps in that evidence.
Work backward from a correction
Connect the user’s correction, the agent’s behavior, and the skill wording that may explain it. Show the relevant excerpts and explain your current reading in plain language. Use dialectic to examine that connection with the user, including discrepancies between what the skill says and what the agent did. Develop one concrete issue at a time, letting their reactions and the evidence revise the explanation before deciding what to change.
Distinguish a misunderstanding the agent could have avoided from a preference or idea that emerged through useful exploration. A disagreement is evidence to investigate, not proof that the agent was wrong.
Inspect the implicated skills and relevant surrounding instructions together. Determine whether the problem was missing or misleading guidance, competing instructions, skill selection, or failure to apply guidance already present. Loading a skill does not establish that it caused the mistake. If the record cannot support attribution, keep the diagnosis provisional rather than assigning blame or adding a rule.
Improve the responsible skills
Use skills-and-instructions to place and make the resulting skill edits. Prefer changing the existing owner and removing guidance that encouraged the failure over adding compensating rules. Do not spread the same lesson across every skill involved. Some findings warrant no skill change.
Show what the changed guidance would produce on the same request, using only context available at that turn.
Show the actual skill changes and the behavioral consequence together. For prose, show the exact wording changed; for scripts or configuration, show the relevant change and its effect.
Keep comparison inputs consistent, distinguish illustrative responses from actual runs, and check a materially different request for transfer or regression. Use skills-and-instructions' evaluation guidance when designing behavioral comparisons. Report what remains unmeasured; success on the motivating exchange alone does not establish improved accuracy.
Judge improvement by whether future agents can avoid the same misunderstanding while preserving useful discovery, not by minimizing conversation turns. Carry authorized changes through and report what changed, why, and what was checked. Respect a request to diagnose or propose without editing. Creating this workflow does not itself authorize running it on a referenced conversation.