HoneyConf 2026 Talk

I gave a talk at HoneyConf 2026 (a dev meetup in Montreal) about seven personal projects and what they taught me about AI. I went in with naive hope, fell into the trough of disenchantment, and eventually got to a point where I could use AI as an ordinary tool.

The deck and artifacts are at the end.

The curve

I started top-left: "AI's gonna write my code for me", vibe-coding, miracle... And then I fell into the trough. It doesn't work. Or it works one time out of two. I'm thinking "damn, this is garbage".

So I decided I wanted to push through the trough. On the other side, AI becomes useful, though it never lives up to that initial hope. I could use it without treating it as magic or an existential threat.

Two fears in the background

Two fears kept me trying:

• Right now, I'm afraid of losing my job. Not in five years, but in September. That keeps me trying AI even when it disappoints me.
• Longer term, I'm afraid of being left behind by the tech. I love this field and want to still be hirable in 5-10 years. That puts the tool back in my hands after I've dropped it.

The cautious start: Window Café

First serious project with AI from day 1: I was becoming a dad, I wanted it to work. I started slow, extremely cautiously, using AI to add things like PostgreSQL + Prisma, Docker, linting, tests, etc... but no major features.

At the start, that meant a lot of churn (and wasted tokens): AI would write the feature, the tests would break, it'd fix them, we'd start over. Many, many times. I learned to encode the guardrails into the rules to steer it from the start. Otherwise, it kept having to correct itself at the end and dig itself out with re-runs.

→ window.cafe (opens in new tab)

The YOLO middle

Once I got past that caution, I pushed hard to see where AI could take me when I gave it more control and freedom:

• This anti-framework blog has 8 invariants. It's built statically, from scratch, with no static-blog framework, just Vite + 7 in-house plugins. This was my first use of pre-built agents (agency-agents (opens in new tab) by Mike Sitarzewski, with UI, UX, designer and researcher agents called in sequence).
• With Destiny, I learned astrodynamics by shipping a 3D browser game. I wanted to push AI usage and use it for everything: Gemini for the 2D ship design, tripo3d for the mesh, Gemini for the HUDs, Lyra for the code rules, Claude.AI for the domain research (KSP's patched conics), Claude Code to execute. I learned to use the right AI tool for each job instead of staying locked in a single LLM, and to use tests + git history as externalized memory.
• CityForge was my first time with /plan. I threw away the v1 prototype and started over, encoding the mistakes from the start. Six agents in parallel, five days, zero reverts.

Three more projects, three more lessons

• Encaisse helped me formalize what I'd learned. It was my first use of /get-shit-done, and I went through three AI prototypes before the final boilerplate. I learned CASL mid-build. /evolve was born here, /commit-message got formalized, and I tried git worktrees and multi-agent flows for the first time. Three path-gated rules let AI load the right context based on the layer being touched. encaisse.ca (opens in new tab)
• With Home Nexus, I solved three household problems using AI: an INFRASTRUCTURE.md + runbooks, an SSH CA so my hosts can sign each other, and a Traefik services.conf with compiled rules. I learned to pass context between tools to get fresh eyes, and encode the mistakes into the CLAUDE.md.
• My Daughter's Little Book was a picture book for my daughter, with consistent characters across 20 pages. I used seven or eight Gemini threads in parallel and found that the model drifts within the same thread. A new thread + the same BABY_CORE kept my daughter visually consistent. I could switch threads as long as I encoded what needed to stay stable.

A typical session

Those lessons became my session loop:

A SETUP once per project feeds the loop: PLAN → SPEC → BUILD → DECIDE → REVIEW → COMMIT. At DECIDE, you can throw it away and start over with the learnings. After COMMIT, what you keep goes back into CLAUDE.md and the rules via /evolve, following the big red arc back to the start.

Wrap-up

I need a fast feedback loop to learn what works and what doesn't, then encode it into the system for the next session. I also need a human in the loop to decide when to follow the AI, when to correct it, and when to throw it away and start over. What works for humans works for AI too: it needs rules, guardrails, context, memory, a plan, a spec...

See the deck (only in French)

Thanks to the HoneyConf folks for being there!