Skip to content
Cameron Mills

All projects

Practice sessions that save every answer before the next question

The practice engine

Part of Life Tracker

A question-by-question practice engine that follows up until understanding breaks, keeps a queue of weak spots, and saves every answer to a sourced tracker before marking the next, in a chat session or on my dashboard.

Code on GitHub

Problem

Reading notes back is a poor test of understanding. I wanted practice that asks one question at a time, pushes until the answer breaks, remembers what I got wrong, and never invents material the sources do not contain. A session can also crash or lose its context, so no answer should ever be lost when it does.

Approach

Split the work cleanly. Everything predictable (what to ask next, spacing, re-tests, saving) is code. Only judgement, writing a question and marking an answer, goes to a model, as one call per turn that returns structured data the page can display.

Architecture

Each subject has a tracker built from source material, and every point in it names the file it came from. Anything an agent inferred rather than read is marked as inferred and kept separate. Each tracker opens by listing what the material does not cover, and a session says so rather than inventing an answer. A point heard only in a recording is marked unverified, and is never used as a model answer until written material backs it up.

A session works through course content that has not yet been tested, rotating between topic groups, and returns to re-tests and weak spots once the course is covered. Every reply ends with a question. Re-tests are capped so they never crowd out new material, and a missed point comes back a few questions later.

What I built

The same rules run in a chat session and on my dashboard, where the tutor can be Claude Code, Codex or a model running locally on the Mac. Both save through one writer, which changes only the answered point and the queue of weak spots, and schedules a re-test for anything answered badly. Saving is safe against two writers at once, and a save that fails is retried at the end of the session. A self-test runs a scripted session and checks that both routes leave identical files. Worked exercises take their model answers from a small solver, and a test fails if any figure drifts or the totals stop adding up. The dashboard's practice screen came off its tab bar on 9 October, as Polaris turned from P1 to P2, and stays reachable.

Results

The first question of a session used to wait half a minute on a model call. A pool written in advance now serves it instantly. Later questions are written while I am answering the current one, and a prepared practice card fills any gap. Practice cards never change the tracker.

What didn't work

A smaller, cheaper model was tried for writing questions and was no faster under load, so the speed came from writing ahead rather than from the model.

Stack

  • Python
  • Claude Code
  • Claude Sonnet
  • Markdown

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit 0d5d4a9 on 14 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.