Review recent Edu sessions and runs to propose lessons, hypotheses, skill changes and canonical proposals: read-only, evidence-linked, never applied without human approval. Triggers: 'reflect', 'weekly review', 'what have we learned', repeated mistakes, after a long run; 'reflexiona', 'revisión semanal', 'qué aprendimos', 'estamos repitiendo el mismo error'.
Installs into .claude/skills of the current project.
Are you the author of Edu Reflect?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/eduardomoraga-edu-reflect)
---
name: edu-reflect
description: "Review recent Edu sessions and runs to propose lessons, hypotheses, skill changes and canonical proposals: read-only, evidence-linked, never applied without human approval. Triggers: 'reflect', 'weekly review', 'what have we learned', repeated mistakes, after a long run; 'reflexiona', 'revisión semanal', 'qué aprendimos', 'estamos repitiendo el mismo error'."
---
# Edu reflect
## When
- The human asks to reflect, or a review period ends.
- The same problem or procedure showed up in several sessions.
- A run failed, needed a fix round, or a recalled note misled.
## Steps
1. Read recent episodes and runs: `edu_recall` with `tiers: ["episodic"]`, then
`edu_read` on the relevant ids.
2. Look for patterns with evidence: repeated fixes, repeated procedures, notes that
helped or misled.
3. Propose, linking every item to the episodes or runs that support it:
- new **L-** lessons as `candidate`;
- **H-** hypotheses for patterns that are suggestive but unproven;
- `edu_feedback` on existing lessons that outcomes proved helpful or misleading;
- skill diffs or new skills for repeated procedures;
- canonical proposals from lessons that are already `proven`.
4. Present the proposals with their claim bands and wait. Write them only after the
human agrees (`edu_remember`, `edu_propose_canonical`). For the full schema-validated
pass over runs, suggest `edu reflect --since 7d`, then `edu proposals accept <id>`.
## Rules
- Read-only until approved: reflection never edits notes, skills, or code directly.
- No evidence, no proposal. Prefer fewer, stronger items over many weak ones.
- When producing machine output, match the requested JSON schema exactly; malformed
output is rejected, not guessed.
## Failure handling
- The `edu_*` tools are missing: tell the user to run `edu setup`, then restart this CLI.