What Building Email 👍/👎 Taught Me About Feedback Loops
Any 'learn from me' feature needs two layers — an instant deterministic one and a slow LLM one — both reading the same persisted store. Build both. Neve
Models, agents, memory and the systems around them.
Any 'learn from me' feature needs two layers — an instant deterministic one and a slow LLM one — both reading the same persisted store. Build both. Neve
If your app is multi-page and you fire a long operation as an inline fetch, navigation kills it. Move the work into a server-side background job and
My first two model comparisons lied to me — in opposite directions. One made a good model look broken. One made a bad model look like a winner. Here's t
An insight-regeneration cooldown and a row-count cap (returning 413) stopped a potential runaway bill cold. Add both before any AI feature ships —
Skill instruction files must use absolute paths. Relative paths silently misresolve when the shell's cwd isn't what you think. And if your skill liv
An HTTP 200 from an LLM API tells you the request arrived. It says nothing about whether the model finished. Always check stopreason and cont
the setup
Poll your live data fast. Cache the LLM's ranked order separately. Invalidate the cache on an explicit trigger — not on every data change.
I wanted a fast local model ranking my tasks. The runtime had to stay pure stdlib. The answer was a sidecar in its own venv — and the real lesson was: b
When building an attention gate with an LLM, fail-closed beats fail-open. Precision is the differentiator — not recall.
Grep finds the word you typed. RAG finds what you meant. Neither alone was enough — and fusing them costs almost nothing extra.
When Anthropic moved Agent SDK usage off the Claude Max subscription onto a metered credit pool, I had one day to rethink the email triage in Apollo
Every LLM that touches ingested content — at chunking time AND at synthesis time — is holding untrusted input. Don't paste documents into your instr
When your AI agent needs to filter relational data, reach for structured SQL column-filters — not embeddings. The column whitelist is your security
Stock claude speaks Anthropic /v1/messages natively — no proxy needed. If your inference provider does too (Ollama v0.14+ does), just set ANTHR
Slack's new rate-limit regime for non-Marketplace apps can be brutal — but whether your app is on it isn't in the docs. You have to probe live.
The Setup
The Setup
I split my AI infrastructure by consent axis. Daemons that fire without me → Western providers only. Me at the keyboard → my call, per-call.
Pure semantic RAG sounds like the upgrade. It's not sufficient on its own. The gap isn't recall quality — it's freshness and coverage. Run grep alongsid
~/.claude/settings.json doesn't just set your interactive model — it leaks into every background claude -p subprocess (headless Claude call) tha
Before you add a translation proxy, grep your binary. The tool you're wrapping might already speak the format you need.
HTTP 200 from an AI provider is NOT proof the call succeeded. You have to verify stopreason == endturn AND non-empty content — every time. Differe
A static system prompt makes your embedded AI agent blind to the record the user is actually looking at. Pass the page entity ID in the POST body, f
If your RAG synthesis layer strips the date from a chunk before answering, you're not building a memory system — you're building a confident liar.
When a client departs, 'remove them from everything' is the wrong instruction to give your AI agent. The right move is surgical. I learned this when
When onboarding a non-technical user to Claude Code, the entry point isn't a checklist. It's one prompt they paste — and the agent does the setup for th
Texting 'call me' to my AI agent now rings my actual phone. Getting there required three non-obvious decisions I'll make again on every agent I build.
Don't cache cheap, user-mutable reads in a long-lived execution context. The staleness cost will eat you alive.
Cold-start bugs don't live in your shell — they live in the real launch environment. You can't find them until you run there.