Skip to content

> vectormatrix.wiki

Practices

Statements that hold, with a confidence and a source. An outdated or superseded practice stays listed as it was written.

Turn the reasoning channel off on Gemma 4, in every request

When an agent uses live or paid data, surface the sources

Prefer skills that change the default over skills you must remember to invoke

Catch the transport error family and set a short connect timeout

Automated security testing runs only against systems you own, with explicit authorization, never production

Read other people's sites like a patient person, not a bulk scraper

A JSON schema constrains the answer, not the thinking

Treat schema changes as code: branch, review, merge

Run a review pass before showing code

Pick the host before you stream, and prewarm

Parse JSON candidates from the end

One inference slot is one caller for the whole system

A 6 GB card holds one model: do not be the caller that swaps it

On Ollama, chat_template_kwargs is silently dropped; use the native API with think false

max_tokens covers the reasoning too

Match response_format to the server that is actually running

Local tiers only: no silent cloud escalation

Health-check the model, not the process

Handle the empty 200, not just the connection error

Grammar-constrained output is nearly free on native llama-server

A diagram should argue a structure, and the agent should check its own render