Customer sites don't belong on your domain
We moved 28 customer sites off our main domain to stop sharing one reputation with them. Why the risk is real, and how the move took about half an hour.
Read →
Field notes
Field notes
Moving config into the database — without trusting the database
We moved billing-currency config into an admin-editable table while keeping a copy in code — bound together so any drift between them fails loudly.
Read →
Field notes
Instruction rot: why your AI agent quietly does the wrong thing
We audited the instructions our agents rely on and found around forty places where they no longer matched the codebase. Nothing failed loudly — but the drift was changing behaviour. Here's what we found and what we changed.
Read →
Field notes
An engineer you manage from a group chat
We built a coding-agent loop from Claude Code, Linear, Telegram, and git worktrees — no framework. Tickets become reviewable PRs, with a human approving every step. Day one, it shipped two merged fixes.
Read →
Field notes
What actually breaks when you drop the old column
Dropping our mirrored status column was easy in the model code. The real work was scripts, migration-graph conflicts, deploy timing, and NULL semantics.
Read →
Field notes
The cache that never hit
A unique job ID slipped into our cache key, giving a paid-API cache a 0% hit rate for weeks — no error, just quietly paying twice. Here's how we found it and what belongs in a cache key.
Read →
Field notes
Is Meta sending your WhatsApp ad click IDs? Your database already knows
We thought our Click-to-WhatsApp ads weren't sending ctwa_clid. One query on our own leads table showed the exact day it switched on.
Read →
Field notes
When your AI agent's memory overrules the user
Our build agent kept a persistent memory. A customer asked to change a dull background colour; the agent recalled an older "no bright colours" preference and ignored them. Here's why that happens, and what we changed.
Read →
Field notes
How we isolate Claude Code skills per client (it wasn't --setting-sources)
We assumed --setting-sources project isolated Claude Code skills per client. It only scopes settings, not skill discovery. The real boundary came from per-workspace git roots and a fail-closed allowlist copy step.
Read →