Field notes
Notes from building agentic systems.
Evidence-first, and sceptical. Most of these exist because a measurement contradicted an assumption.
Channels
Index →Start here
A side entrance, not a ranking
- Skeptic's Toolbox
Your harness is half the model
I ran four coder models through their officially-recommended harnesses and one came out 100x slower. The cause was my shell's rm -i alias leaking into Claude Code's tool subprocess. Your model comparison will credit that slowdown to the wrong variable.
- ML Journal
SFT Learns Your Median Employee, Not Your Best One
4,000 real support conversations, textbook loss curves, and a fine-tune that lost to its own base model 92% to 8%. SFT converges to the median of your dataset, and the median operator writes short answers with generic links.
The trace
TIME ↓ REVERSE-CHRON
- [09]Craft
The Most Annoying Prompt Is the One I Never Refactor
I spent a night judging another developer's uncleaned repo. Then I found the same rot in my own global instructions — a rule written three model generations ago, never once refactored.
- [08]Build Log
The quarterly K8s migration is now a Tuesday night
38 pods, ~150 GB of state, and an Elasticsearch cluster moved across two AZs in one evening, with 13 minutes of outage. The execution was never the bottleneck. Two months of planning was.
- [07]Skeptic's Toolbox/START HERE
Your harness is half the model
I ran four coder models through their officially-recommended harnesses and one came out 100x slower. The cause was my shell's rm -i alias leaking into Claude Code's tool subprocess. Your model comparison will credit that slowdown to the wrong variable.
- [06]ML Journal/START HERE
SFT Learns Your Median Employee, Not Your Best One
4,000 real support conversations, textbook loss curves, and a fine-tune that lost to its own base model 92% to 8%. SFT converges to the median of your dataset, and the median operator writes short answers with generic links.
- [05]People
The Gap Was Never Knowledge
A long-standing client almost cried at a Claude Code tutorial, bookmarked it, and went back to his 400 Excel files. The forty-five minutes that followed taught me that AI adoption was never a knowledge problem.
- [04]Knowledge Architecture
One Principle Beats 250 Words of Rules
A client's prompt carried 250 words of banned phrases. Every one was banned for the same reason, so eight words replaced the whole list.
- [03]Skeptic's Toolbox
Why My Local 80B Model Beats Claude at Coding Tasks, and It's Not About the Model
A local 80B on two 3090 Tis finished the benchmark in 1:02 against Opus's 10:28, and the only thing I changed was the client. Claude Code re-serializes the whole prompt every turn, which kills prefix reuse on hybrid SSM models and costs 20 seconds of re-prefill per tool call.
- [02]Skeptic's Toolbox
When 84% Token Reduction Means Zero Accuracy
Semantic code search cut my token spend by 84% on a real Laravel debugging task and found zero of the four files that mattered. I had been evangelizing the tool for weeks before I finally benchmarked it.
- [01]Build Log
I Built a Content Pipeline Because I Was Tired of Copy-Pasting
I was burning an hour after every post on copy-paste hygiene. So I built the pipeline I wished I had. Directus, n8n, Claude Code, all local-first.