Weekly Notes: June 29 – July 5
Weeknotes - the AI tools, workflow experiments, and reading that actually stuck this week.
An agentic workflow video
A principal engineer walking through how they actually run their day around agents. Most “AI workflow” content is filler - this one isn’t. Worth the 45 minutes.
Opencode GO
Came across Opencode GO. It has access to low-cost models, and I run it through the Opencode VS Code extension. Has Deepseek V4 Pro, Kimi K2.7, etc… and is super cheap.
I had ChatGPT spin up a cheatsheet for me:
Superwhisper
Found Superwhisper. Voice-to-text on the Mac that just works, no setup rabbit hole. I’ve been dictating prompts instead of typing them and it’s faster than I expected. Worth a look.
ESLint as the feedback loop, not more agents
A coworker dropped a good one: stop using agents to check the agent’s code - use ESLint. It’s deterministic, instant, and free. Why pay a second model to eyeball style and correctness when the linter already knows the rules cold?
I ran with it and built a loop with Claude (Fable 5) where lint output feeds straight back into the agent until the code comes back clean. The output is noticeably better - the agent stops relitigating conventions because they’re enforced, not politely suggested.
Fair pushback I got: is Cursor already doing this deeper than VS Code under the hood, and do we even need custom rules for it? Don’t know yet, still digging in. Wrote it up on LinkedIn.
Per-project knowledge bases
A tweet sent me down this one: a self-building, per-project knowledge base. In CLAUDE.md you tell the agent to keep updating AGENTS.md as it works - pitfalls, decisions, gotchas, anything worth remembering. The project documents itself, the agent quits rediscovering the same things every session, and you burn fewer tokens because the context lives in one file instead of getting re-derived on every run. Obvious in hindsight, working well so far.
Obsidian + a personal monorepo
Same idea, wider net. Spun up a personal monorepo and wired it into Obsidian as a knowledge base. Whenever an AI chat produces something worth keeping, I have it write a markdown summary and drop it in the vault. Now every project I touch pulls from the same brain.
The /docs trick
Saw someone keep a /docs folder in their projects and have the agent write weekly updates into it - a running changelog they share with their team. Simple, smart, stealing it.
On the road: AI Engineering
Road-trip audiobook: AI Engineering by Chip Huyen. It’s about building real apps on top of foundation models - prompting, RAG, finetuning, and the part everyone skips: how to actually tell if the thing works (evals). Four or five hours in and it’s holding up. Recommend.
Design engineering reading
Three that go together:
- Design Engineering at Vercel - clearest answer I’ve seen to “what even is a design engineer.”
- Invisible Details of Interaction Design by Rauno Freiberg - the stuff most people never notice, and the reason good UI actually feels good.
And a good talk on the subject: