Official documentation, and when to go there instead

This site exists for the parts of using AI that vendor documentation does not cover: which decision to make, how it fails, whether to trust the output. For anything that is a matter of record — a limit, a price, a parameter, a contractual term — their pages are authoritative and ours would be a copy that goes stale.

Every link below was checked on the date in the frontmatter.

Start here for a specific product

Each vendor splits its documentation the same way: the consumer apps, the coding agent, the API. They are genuinely different products with different behaviour, and the commonest way to get a wrong answer is to read the right sentence in the wrong one.

Anthropic

Anthropic Help Center. The apps. Plans, limits, Projects, memory, Cowork, connectors, file handling. The first place to look for "how does this feature behave", and consistently more current than anything written about it elsewhere.

Claude Code docs. The terminal agent. The best-practices page is unusually good and repays reading in full rather than searching.

Platform docs. The API. Parameters, their deprecations, and the exact behaviour of each.

Anthropic's skills repository. The Agent Skills spec, a template, and example SKILL.md files by category. The reference set to read before writing your own.

OpenAI

OpenAI Help Center. The ChatGPT apps. Data controls, memory, file uploads, retention, plan behaviour. Blocks automated clients, so it opens in a browser and not from a script.

Codex documentation. The coding agent across CLI, IDE and cloud, including permission modes, AGENTS.md and skills.

OpenAI API docs. Parameters, tools, structured output, and the pricing table.

Google

Gemini Apps Help. The apps. Activity and retention, personalisation, Connected Apps, uploads. Google's help pages carry a visible last-updated date, which is more than most vendors manage.

Gemini CLI documentation. The terminal agent: sandboxing, the policy engine, hooks, checkpointing, extensions.

Gemini API docs. Models, capabilities, and the API surface.

Not owned by any of them

Model Context Protocol. The protocol itself, maintained separately from any vendor's product docs, and the right reference when the three implementations disagree.

For a fact that changes

Model overview. What exists, context windows, capabilities.

Pricing. Plan and API pricing. Our model matrix is generated from a data file we maintain by hand, which means it can lag theirs by a day. Theirs wins.

Release notes and news cover what changed and when.

For OpenAI, plan pricing sits on the ChatGPT pricing page and token pricing in the API docs. For Google, subscription tiers are on Google AI plans, whose prices render in the browser and cannot be read by any script, and model pricing sits with the API docs.

This is the section where our pages deliberately say least. Plan advice ages badly, so the guides here name the tiers and the shape of the limit and send you to the vendor for the number.

For what you are agreeing to

Consumer terms and commercial terms carry the distinction that decides whether inputs may train models. Our what data may go in explains why the distinction matters; the terms themselves are what your legal team needs.

Privacy Center. Retention, deletion, the training toggle.

Trust Center. Certifications, subprocessors, security documentation. The place to send a vendor questionnaire before writing one yourself.

For OpenAI, the policies hub holds the terms, how your data is used states the training position per plan, and enterprise privacy carries retention and the business commitments.

For Google, the Gemini Apps Privacy Hub covers activity retention and human review on consumer terms, and the Workspace privacy hub covers the different commitments that apply to a work account. Those two differ in ways that decide a data-protection assessment, which is why we treat them as separate sources rather than one.

For learning properly

Anthropic Academy and OpenAI Academy. Free structured courses. If you want to be taught rather than to look something up, start there. We deliberately do not reproduce either. We have found no Google equivalent to point you at. That is a gap on their side, and it is named here so the absence reads as deliberate.

Two engineering write-ups worth reading whole, because most practical advice elsewhere is downstream of them: Building effective agents and Effective context engineering.

What this site is for instead

Deciding between options. Knowing what goes wrong before it does. Verifying output. Governance and the DACH regulatory picture. Where their documentation says what a feature does, we try to say when to use it and how to tell it worked.

When the two disagree, believe theirs and tell us. Every page on this site carries a reporting link at its foot for exactly that.

What goes wrong

Trusting a third-party paraphrase over the source. Including ours. AI documentation changes in weeks, and a confident secondary summary is the most common way people end up acting on a fact that stopped being true.

Searching the wrong product's docs. Every vendor splits agent, app and API documentation, and the behaviour genuinely differs. A Claude Code answer applied to the apps is often wrong, and so is a Codex answer applied to ChatGPT.

Reading a blog post as documentation. Anthropic's engineering posts are excellent and are not reference material. They describe approaches at a point in time; the docs describe current behaviour.

Assuming a dated page is wrong. Something published eighteen months ago may still be accurate. Check it against current docs rather than discarding it, and check our pages the same way — the verified_on date on every guide here says when someone last looked.

How to check it worked

Take a claim from this site that matters to a decision you are making, and find it in the vendor's documentation. If it holds, you have a source you can cite to someone who asks. If it does not, theirs is correct and ours needs fixing — the reporting link at the foot of this page opens an issue with the page and date already filled in.