Building a Habit App That Never Shames You

Most habit trackers are built on a lie: that the way to make you consistent is to punish you when you’re not. Break a streak and watch the number reset to zero. Miss a day and get a guilt-tinted notification. It works right up until the day you slip — and then the shame does the opposite of what you needed, and you delete the app.

I have ADHD. I’ve deleted a lot of those apps. So I built the one I actually wanted: CYBER//FIT — a cyberpunk self-improvement tracker with a single non-negotiable rule. It never shames you.

Forgiving by design

The whole product is organized around removing guilt as a mechanic:

• Skip is not fail. Missing a day doesn’t zero your progress.
• Streak shields auto-absorb misses, so a bad week doesn’t erase a good month.
• A reset is a “reboot,” never a “failure” — you come back to “Good to see you, choom. Ready for a new gig?”, not a pile of red X’s.
• It never shows you a count of missed days or notifications. That number helps no one. It’s not in the app.

Gamification is there — XP, levels, cosmetic cyberware unlocks — but it’s scaffolding, meant to get a habit started and then get out of the way as real progress data takes over. Crucially, every reward is cosmetic. Nothing useful is ever locked behind progress or payment.

Off-grid by design

I’m a privacy-first, local-first engineer, and the app reflects it:

• After first load it works 100% offline, forever — an installable PWA with a local database. There’s no account. There’s no server for your data to go to.
• No analytics, no telemetry, no cookies, no tracking. Backup is a local file you control.
• Notifications are strictly nice-to-have. If you opt in, the relay stores only an anonymous push token and the time slots you picked — it’s blind to your schedule and your completions by design.

It’s free forever, open source (MIT), and the whole thing runs on Cloudflare Workers for about the cost of a domain.

The AI sidecar: a coach you can text

The piece I’m most proud of as an engineer is private and personal: cyber-trainer, an SMS coach that runs on my own hardware.

I text it in plain English — “just finished the workout” — and it logs the right entry for me. The architecture is deliberately paranoid:

1. A text hits a Twilio number, forwarded through a Cloudflare Tunnel (no inbound ports open anywhere).
2. The request lands on a service running on my Jetson, co-located with a local Qwen3-8B model served by llama.cpp.
3. The model’s only job is to turn one message into one strict JSON intent. It never touches storage and never touches the network.
4. My own code owns every read and write to the encrypted data vault, then texts back a confirmation.

That separation — the LLM interprets, my code acts — is what lets a small local model run a real workflow safely. No cloud AI, no data leaving my house, the same local-first ethic as the app it feeds.

Why build it this way

Because the technology should serve the person, not the metrics dashboard. A habit app’s job is to make it a little easier to be who you’re trying to be — on your good days and the day after a bad one. Punishment-free isn’t a soft feature; for an ADHD brain it’s the difference between a tool you keep and one you delete.

If any of that resonates, try it — no signup, nothing to lose, and it’ll never make you feel bad for missing Tuesday.

Update — screenshots and a user guide

When I published this, the repo had no screenshots — a privacy-first app has a bootstrapping problem: I wasn’t going to post my own vitals, and empty-state screenshots sell nothing. The fix was to treat screenshots like a build artifact. A seed script generates a fully fictional demo profile — three weeks of habits, workouts, reading logs, even PHQ-9 results, all invented and schema-checked against the real app — and a Playwright pipeline installs it through the app’s own backup-import path, then walks every screen and captures it. npm run shoot:demo, nine deterministic PNGs, no human data anywhere.

So here’s what the app actually looks like:

And the screen I most wanted people to see — the crash kit. One tap from anywhere: paced breathing, grounding, and call/text 988. Nothing in it is logged, scored, or tracked, and it works with zero signal:

The same pipeline now feeds a full user guide — every screen, documented with the same fictional profile. When the UI changes, I re-run one command and the docs stop lying.