session-corrections-to-skill

Turns corrections made during a session into durable knowledge in the matching skills. It accepts two sources: a user explicitly rejecting the agent's existing behavior, or reality disproving an approach when the root cause and correct approach have already been verified.

One-off scope instructions, choices between valid options, general preferences about how the agent should work, and unresolved problems are not deposited. Each accepted correction is checked against the target skill's description; corrections outside that scope are never forced into the nearest skill.

Deposited content

One correction may produce one or more of these forms:

• Definition: name + requirement
• Implementation detail: scenario + action
• Error case: scenario + error

Drafted content is shown to the user before any file is changed. Only confirmed content is written, using an existing matching table or a minimal new two-column table. The workflow does not opportunistically restructure or expand the target skill.

When to use it

• At the end of a session where you corrected a skill's behavior and want the correction to stick
• After a pitfall whose root cause and correct approach were verified in the session

What you get back

• Confirmed definitions, implementation details, or error cases added to the matching SKILL
• A list of rejected one-off, general, or unresolved items
• A separate list of corrections that need a new skill, with a recommended name
• A ledger mapping each original correction to the exact content written