THE LAB

Everything we've learned using coding agents.

The information hub and a place where we examine capabilities and boundaries of LLM-s and agentic work

  • 8 WRITE-UPS
  • 4 MODELS COVERED
  • 13 OSS TOOLS
  • UPDATED 25 JUL 2026
LATEST BENCHMARK

21 JUL 2026 · BENCHMARKS · KIMI K3 · CLAUDE FABLE 5 · CLAUDE OPUS 4.8

Kimi K3 vs Claude Fable 5 and Opus 4.8: a benchmark you can run yourself

Everyone posts Kimi K3 benchmark scores; the benchmark itself is harder to find. So we built one vs Claude Fable 5 and Opus 4.8, read every line all three wrote, and open-sourced it.

RUN IT YOURSELF →
ALL WRITE-UPS
25 JUL 2026MODELS

Claude Opus 5: everything you need to know

Claude Opus 5 is out at half the price of Fable 5, with thinking on by default — and one change that breaks existing code. The facts, from the docs.

  • ai
  • claude
  • llm
21 JUL 2026BENCHMARKS

Kimi K3 vs Claude Fable 5 and Opus 4.8: a benchmark you can run yourself

Everyone posts Kimi K3 benchmark scores; the benchmark itself is harder to find. So we built one vs Claude Fable 5 and Opus 4.8, read every line all three wrote, and open-sourced it.

  • ai
  • kimi
  • claude
  • llm
  • benchmark
20 JUL 2026

Building an MCP server in Python (and connecting it to Claude Code)

An MCP server lets an AI model call your own tools. Here's how to build one in Python with the official SDK and connect it to Claude Code — tested end to end.

  • mcp
  • python
  • claude
  • ai
17 JUL 2026MODELS

How to use Kimi K3 with Claude Code (subscription, not API)

Kimi K3 just shipped and it rivals Claude Fable 5. Here's how to run it inside Claude Code on a Kimi subscription — no Anthropic account, no API bill.

  • ai
  • kimi
  • claude
  • llm
14 APR 2026INFRA

Kubernetes kills your pod? Here's why

Your pods keep getting killed. Not crashing — killed. One moment they're running fine, the next they're gone and Kubernetes is spinning up replacements. You check the logs and there's nothing useful…

  • kubernetes
  • devops
  • infra
06 APR 2026INFRA

How to see running queries in Postgres and kill them

Something is slow. Maybe a page takes forever to load, maybe a migration is hanging, maybe your Supabase dashboard just spins. You suspect a query is stuck somewhere in your database…

  • postgres
  • database
  • infra
09 MAR 2026AGENT WORKFLOWUPDATED 18 JUN 2026

HTML/CSS Animation to Video (MP4): the Headless, Deterministic Way (incl. Claude)

Convert an HTML/CSS animation to MP4 video the deterministic way: headless Chrome plus FFmpeg, frame-perfect every time — including animations Claude generated for you.

  • ai
  • claude
  • devtools
27 FEB 2026INFRA

How to toggle realtime on a table in Supabase migrations

You set up a table locally, opened it in Supabase Studio, flipped the Realtime toggle on, and everything worked — your frontend was getting live updates, subscriptions firing perfectly…

  • supabase
  • database
  • infra
MODEL HUB

Every model we've written about, and the pieces on each — benchmarks, briefs, how-tos.

LAST UPDATED 25 JUL 2026
MODELWRITE-UPS
Claude Opus 5Claude Opus 5: everything you need to knowREAD →
Kimi K3Kimi K3 vs Claude Fable 5 and Opus 4.8: a benchmark you can run yourselfREAD →How to use Kimi K3 with Claude Code (subscription, not API)READ →
Claude Fable 5Kimi K3 vs Claude Fable 5 and Opus 4.8: a benchmark you can run yourselfREAD →
Claude Opus 4.8Kimi K3 vs Claude Fable 5 and Opus 4.8: a benchmark you can run yourselfREAD →
VOCABULARY

Useful AI-engineering-adjacent terms for your reference.

Agentic engineering
Building software by directing agents that plan, edit, run and check their own work, instead of typing every change yourself. The job moves from writing to specifying and reviewing.
Context rot
The slow drop in quality as a long session fills up with stale detail. Starting fresh with the right three files usually beats carrying on.
Context window
How much text a model can hold in mind at once.
Determinism
This is when the same input yields the same output every time. LLMs are not deterministic by nature, but can be brought closer to that by e.g. prompting or providing them with callable tools.
Eval
A repeatable, scored test for a model or an agent. Without one, "it feels smarter" is your only metric, and it's wrong about as often as it's right.
Guardrails
What a change has to survive before it counts: types, tests, lint, a staging gate. Agents move fast enough that the gate is the design.
Harness
The code around the model: the tools it can call, the files it can see, the limits it runs under. Most of what people call model quality is actually harness quality.
Model Context Protocol (MCP)
A widely-adopted standard for enabling LLMs to integrate with external tools and data sources.
One-shot
Getting a whole task right from a single prompt, without follow-ups.
Prompt-to-PR
The loop from an instruction to a diff someone can review.
Slop
Unreviewed AI-authored output that might work, but it's not necessarily good long term and/or is largely not human-digestible.
Subagent
A second agent with its own clean context, handed one narrow job so its output doesn't flood the main session.
Token
The unit models read, write and bill in — roughly a word-piece. Both your cost and your context are counted based on these.
Vibe coding
Prompting your way to working software without reading most of what came back. Fast for a demo, expensive the moment someone else has to maintain it.
TOOLS WE BUILT

Here are some of our tools that we created just because we needed. We hope you will find them useful too.

  • supabase-plusBatteries-included command-line utility that extends the official Supabase CLIREPO →READ THE WRITE-UP →
  • htmlrecRender HTML animations to video — frame-perfect, headless and deterministicREPO →READ THE WRITE-UP →
  • cant-happenA lib and a manifesto of panicking gracefully and only if it can't happenREPO →
  • vue-modalVue version of our modal lib; call the collector globally, define a modal, trigger in typed wayREPO →
ALL 13 TOOLS →
MACHINE-READABLE

Built for the crawlers too: /rss.xml · /feed.json · /llms.txt.

Everything here was written by developers working with AI. If you want them on your team — say hello.