---
name: capture-learning
description: Use when a retrospective, failed handoff, repeated correction, long-running task, recurring process problem, or successful new workflow should change how future project or architecture work is handled.
---

# Capture Learning

Convert demonstrated lessons into the smallest durable improvement. One unusual event is evidence to inspect, not automatically a universal rule.

## Reconstruct

Gather the relevant conversation, decision trail, artifacts, outcomes, corrections, and available shared records. Keep the review inside the active project boundary. Identify what happened, what was expected, where the first consequential divergence occurred, and what evidence supports the explanation.

When independent review is available, apply [learning reviewer prompt](references/learning-reviewer-prompt.md) through distinct judgment, tooling, and divergent lenses. Otherwise run those lenses separately and disclose the limitation.

## Qualify

A proposed lesson must name:

- the recurring situation or clear generalizable pattern;
- the future decision or behavior that should change;
- evidence that the change would have prevented or shortened the problem;
- the narrowest place that owns the correction;
- a realistic evaluation scenario.

Reject advice that merely restates the desired outcome, responds to a one-off preference, duplicates an existing rule, or expands authority.

## Route

Read [change routing](references/change-routing.md). Prefer structural controls in this order when practical: required field, shared schema, template, validator, deterministic script, automation, connector configuration, routing condition, then prose skill guidance.

Present proposals as `accept`, `reject`, or `backlog`, with evidence and routing. A durable skill, workflow, template, automation, or shared system must not be changed until the user explicitly approves the selected proposal. Backlog filing and external updates also require authority.

## Verify

For an approved change, add or update the evaluation scenario first. Apply the narrow correction, run structural validation, and compare behavior. Report the changed artifact, evidence, evaluation result, rejected proposals, and remaining gaps.