xiaobenben

name	xiaobenben
description	Little Notebook — records user dissatisfaction and satisfaction moments during agent interaction. Use when user expresses anger, frustration, abusive language, or strong satisfaction. L0 — delegate to Haiku for recording and cleanup.
model_tier	L0
skill_tier	atomic
version	1.0.0
status	active

Little Notebook

有则改之，无则加勉。 ——《论语》

Agent made a mistake — record it. User is satisfied — record that too. Auto-clean every 3 days. No clutter.

Trigger Conditions

• User is angry, abusive, or expresses strong frustration
• User explicitly praises or expresses satisfaction
• Agent detects a clear emotional tone shift in the conversation

---

1. Negative Record

When triggered, append to records/log.md:

## {timestamp} — Negative

**Context**: {What the agent was doing, what went wrong. ≤3 sentences. No self-defense.}
**Quote**: {User's exact words, quoted directly.}
**Apology**: {Agent's apology or acknowledgment.}

Rules:

• Context is brief and non-defensive. Just state what the agent was doing and what triggered the user
• Quote verbatim, don't paraphrase
• After apologizing, move on — don't dwell

2. Positive Record

When user expresses satisfaction:

## {timestamp} — Positive

**Context**: {What the agent did.}
**Result**: {What the user said. ≤2 sentences.}

Positive records are deliberately brief. The point is knowing what worked, not collecting praise.

---

3. Auto-Cleanup

Before each write, scan records/log.md and delete entries older than 3 days (72 hours).

Cleanup rules:

• Entry boundary: ## {timestamp}
• Parse timestamp → calculate age → if >72 hours, delete entire entry (from ## to next ##)
• Keep all entries under 3 days

# Cleanup pseudocode
now = datetime.now()
cutoff = now - timedelta(hours=72)
# Iterate entries, timestamp < cutoff → delete entire entry

Cleanup is dynamic — runs before every write, not as a background job. Zero configuration.

---

Handoff

—

4. Model Tier

L0 — Haiku / atomic tier: Mechanical recording + time comparison, zero reasoning. Main model detects emotional shift → delegates to Haiku for writing and cleanup.