merge-feedback

Merge Feedback

One job: take what this session revealed about how the user wants to work and merge it into tasks/feedback.md as a new dated section, preserving everything already there. This is a Utility — the feedback artifact, every time. It does NOT write session-notes or the memory vault (end-session owns those and calls this only for the feedback piece).

Steps

1. Read the existing tasks/feedback.md if it exists. You must see the current content before writing — the merge rule depends on it.

2. Synthesize this session's observations into the five sections below. Scan the FULL conversation, not just recent turns. Be specific — paraphrase the actual correction/preference, not vague generalities. Keep each bullet ≤ 20 words. No code snippets or implementation details — behavioral/communication patterns only. (A genuinely technical, non-behavioral gotcha belongs in the memory vault, which end-session handles — don't force it into feedback.md.)

3. Write a new dated section at the TOP of the file (most recent first), using today's date:

# Session Feedback — {YYYY-MM-DD}

## Preferences
- [concise bullet per preference observed]

## Corrections I Made
- [exact correction → what to do instead]

## What Worked Well
- [approaches the user accepted without pushback or praised]

## What to Avoid
- [behaviors the user pushed back on or corrected]

## System Instructions for Future Sessions
- [imperative rule a future session can follow immediately]
- ...

1. Patch the skill that was running. A correction recorded only as prose in feedback.md does not reach the skill that produced the mistake — the next invocation repeats it. So for each correction captured in step 2, ask: was a named skill driving when this happened? If yes, also append the lesson as one line to a ## Gotchas section in THAT skill's own SKILL.md (create the section at the end of the file if absent).

   Write the gotcha as an imperative rule with its trigger, not a story: "Before X, check Y — Z fails silently otherwise." One line each. Skip this step entirely when no skill was driving; a correction from ordinary conversation belongs in feedback.md only.