Chalkbook

The problem

I wanted a training log that felt like a journal rather than a spreadsheet. Every climbing app I tried wanted an account before it would show me anything, and most of them looked like a gym at night — dark, neon, loud. I wanted the opposite: warm, calm, paper-like. Something closer to the 5 Minute Journal than to a fitness tracker.

So: local-first, single user, no accounts, no server, no social features. SQLite on the device is the only source of truth.

How it works

The core loop is one tap. Open the app, and today is already the subject — greeting, a week strip with status dots, and a large “Climbed today” card. Tap once and the session exists. Everything after that is optional: discipline, gym, how it felt on a 1–5 scale, notes, individual climbs with grade, count, and style tags.

The parts that took actual design work:

• Versioned migrations. Four schema versions so far. Adding workouts alongside climbing meant new tables, seeded preset types, and foreign keys with RESTRICT so a workout type in use cannot be deleted out from under its rows.
• A layered structure. app/ is routing only; src/db holds repositories and migrations, src/domain is pure functions, src/theme owns every visual token, src/stores is deliberately thin because SQLite — not zustand — is the truth.
• Goal-driven dashboard. Yearly grade goals with progress, streaks, a grade pyramid, style analysis. Body metrics and cycle phases overlay onto performance data, which surfaces something a spreadsheet never showed me: which phase I actually climb strongest in.

Working with AI on this

Chalkbook is the project where I stopped treating the model as an autocomplete and started treating it as something that needs a spec.

Three files do the work. PLAN.md is the product spec and phase roadmap — data model, screens, non-goals, and an explicit rule that work stays inside the current phase. CLAUDE.md holds the conventions the model is not allowed to deviate from: the exact stack, the design tokens down to the hex values, the rule that dates are local-time YYYY-MM-DD strings and never timezoned timestamps, and “no new dependencies without asking first”. CLAUDE_HANDOFF.md gets rewritten at the end of every session so the next one can pick up cold.

That last file is the one I did not expect to matter. A session ends when context runs out, not when the work is finished, and without a handoff the next session confidently rebuilds something that already exists. Writing the summary also forces me to state what is actually done — which is how I caught, more than once, that “done” meant “compiles” rather than “verified on device”.

The pattern underneath all three: the constraint is the deliverable. The design tokens file exists so no component ever hardcodes a hex value. The phase rule exists because a capable model will happily build Phase 5 while you are still verifying Phase 2. Deciding what the model is not allowed to do turned out to be more valuable than describing what it should.

Where the source is

The repository is private for now — it holds personal health data alongside the code. Happy to walk through the architecture or the milestone docs in a conversation.