Guides & resources
Pages worth keeping: what each one helps with, who it suits, and what it does not cover.
Choose a model for a workflow, then verify what answered
A short procedure for the two decisions that sit behind most model trouble: which model name your workflow calls, and whether the thing that answered is the thing you tested. It uses the settings the providers document - auto-selection tiers, follow-the-family aliases, gateway pricing and effort caps - and points to the two narrower guides that go deeper.
For Anyone wiring a model into a repeatable task: a scheduled job, an agent, a document pipeline or a small internal tool where you need to know what ran and what it cost.
Does not cover: This covers documented settings only, on the pages linked below, and providers change defaults and catalogue entries without notice - check the release notes for the version you actually run. It does not measure output quality and it does not compare models on a benchmark; it only helps you confirm you are running what you tested. Not every client exposes every setting, and the alias behaviour described here is what the catalogue entry states, not a guarantee about timing.
Check what a latest model alias points to before you ship
OpenRouter's catalogue lists entries that always follow the newest model in a family. This guide shows how to see what such an entry points to today, and when to call a fixed version name instead.
For Anyone whose script, app or agent calls a hosted model by name and needs to know which model actually answered.
Does not cover: Only OpenRouter's catalogue is covered, and only what its public pages and API state. The pages do not say how quickly an alias follows a new release, and a catalogue entry is a listing rather than a quality claim. Alias targets and prices change, so re-check both before relying on either.
Set a Copilot auto-model tier, then check which model answered
GitHub Copilot's automatic model selection now has three tiers. This guide covers what each tier prioritises, where the setting applies, how usage is charged, and how to see which model produced a response.
For Copilot users who let the tool choose the model and want a lever on cost, quality or speed.
Does not cover: The sources are GitHub's own changelog and documentation, so this is vendor-stated behaviour, not an independent test, and neither source reports scores per tier. Tiers exist only in three Copilot surfaces, and auto may serve evaluation models on individual plans, which those users can switch off.
Decide how much autonomy to give an AI agent
How to place a task on the autonomy spectrum, why one bad step spreads, and where to borrow a threat model instead of writing one from scratch.
For Operators and small teams about to let an agent act on real tools, records or money.
Does not cover: The n8n guide is a vendor explainer, not a test of your setup, and it offers no number for how much autonomy is safe. OWASP's document is a threat-model reference that the site offers as a download, not a compliance rule or a checklist for your own stack.
Check for AI changes once a week with four public endpoints
A fixed, short check that catches product changes, model releases and list-price moves without a daily news habit.
For Anyone who wants to notice material changes in the AI tools they already use, on a schedule they can keep.
Does not cover: Vendor feeds announce changes and test nothing; preprint feeds are unreviewed claims; catalogue prices are list prices that can carry overrides. These endpoints can change shape without notice, so re-check any URL before automating against it.