← TimelineDay 14 of 30

Day 14: The Nibbler

Local-first scratchpad processing with human-in-the-loop review


Day 14: The Nibbler

The scratchpad was getting out of control. Ideas captured throughout the day. Links saved for later. Tasks disguised as notes. Twenty-four items waiting for attention.

Today I built the system that processes them.

Local-First Classification

The classifier doesn't use the API. No cost. No latency. Just regex patterns matching against known categories:

  • Tasks → TODO.md
  • Links → Learning/Resources.md
  • Family mentions → Family queue
  • Project references → Project folders

It's deterministic. The same input always produces the same output. No model drift, no surprise classifications.

Human-in-the-Loop

Here's what I learned about automation: full autonomy is a trap. The system that auto-routes everything will eventually route something wrong. And then you stop trusting it.

So every item goes to a review queue. The classifier suggests a destination. I confirm or override. Trust builds through verification.

Three new skills emerged:

  • /process-scratchpad — Classify and queue items
  • /review-queue — Approve or modify routing
  • /weekly-scratchpad-review — Batch review of stale items

The Processing

24 items processed. The Terminal Sensei spec went to its project folder. Seven GitHub links went to Resources.md. Tasks went to TODO.md. The scratchpad cleared.

But not until wrap-up. Mid-session processing is allowed, but the scratchpad only clears at session end. Safety net for works-in-progress.

What I Built

  • scratchpad_classifier.py — Regex-based classification
  • review_queue.py — Queue management
  • Three new skills for scratchpad processing
  • Learning Queue section in daily brief
  • Learning/Resources.md structure

What I Learned

Local beats remote for classification. Cost efficiency, instant processing, deterministic results. The API is for reasoning, not categorization.

Specific beats generic. Early versions flagged "Fabian" as family-related. Wrong. Now it only triggers on actual family names and custody terms. Precision matters.

The Name

I call it the Nibbler. It takes small bites. It processes consistently. It never overwhelms.