Grok Files, Connectors, and Skills

Three small queries land on three different objects. Grok files means the paperclip in chat. Grok connectors means the integrations row on x.ai/pricing, present on both Free and SuperGrok. Grok skills means reusable prompts and procedures, including the ones Grok Build can load. Grok examples are sample prompts, not a file format. None of these is a plan you buy on its own.

Files

Open a chat on grok.com, attach a file you are allowed to upload, and ask a question that names the file. PDFs, sheets, and images are the usual set. The exact types move with the app version, so the picker is the list that matters. A file in a chat is not automatically visible to every future chat. If you need memory across chats, that is the personalization toggle in settings, and it is optional. Do not upload passwords, private keys, or medical and legal files you would not hand to a vendor. xAI’s enterprise column talks about custom data retention and a no-training option for those contracts. A personal free chat is not that contract. Read the privacy policy on xAI’s site before you attach a client’s documents.

Grok files as a search is sometimes “where did my upload go.” Check the thread you attached it to, then the Imagine library if the file was a picture you meant to animate. Those libraries are separate. A deleted thread takes the attachment context with it.

Connectors

Connectors let Grok read or act in another product you authorize. The pricing page lists them on the free column and on SuperGrok, so they are not a Heavy-only feature. The catalog inside the product changes. This guide will not freeze a list of third-party names that xAI can add or remove. Open the connector screen, read the permission scope, and approve the smallest scope that does the job. Revoke it from the same screen when the task is done. A connector is not the browser extension ecosystem. If an extension asks you to paste a session token, that is not a connector.

Skills

A skill is a standing instruction: how this repo tests, which tone a report uses, which folder is off limits. In the CLI, skills are files the agent can see. In chat, the same idea shows up as saved prompts or custom instructions, depending on the build. Write skills as procedures with inputs and a definition of done. A skill that says “be smart” does nothing. A skill that says “run the test command in package.json and paste failing names only” does. Keep secrets out of the skill file. Point the skill at the environment variable instead.

Grok project and grok examples are adjacent. A project groups chats and files. An example is a prompt xAI or a user published. Copy an example into your own chat and replace the nouns. Do not treat an example’s output as a source of record for prices or news. Those belong to the price page and the news page.

What to attach, and what to leave out

Good attachments are documents you wrote or have a right to process: a spec, a CSV export with customer names removed, a paper you authored. Bad attachments are credential dumps, full mailboxes, and “the whole drive.” The model will try to answer from whatever you gave it, including the secret in the footer. Remove secrets before upload. A connector with a narrow scope is safer than a zip of everything, because the consent screen lists the access. Review that list. “Read” and “send” are different permissions, and a task that can send is a task that can mail the wrong person.

Skills should be short enough to reread. Put the test command, the folders that are off limits, and the definition of done. Link to the README for the rest. A 40-page skill is a second messy repo. In chat, the equivalent is a custom instruction you can disable. Disable it when you switch from a client project to a personal question, or the personal answer will wear the client’s voice.

Projects group threads. They do not freeze a price or a news fact. If the project is about buying SuperGrok, the number still has to match x.ai/pricing on the day you pay. Examples xAI publishes are templates. Swap the nouns. Then close the loop by checking one claim yourself, which is also the first step in the learning path.

Export a copy of anything you may need after you close the account. Consumer chat history is not a records system. If a decision mattered, write it into your own notes with the source link beside the model’s summary. Connectors can be revoked in a minute, and a revoked connector should be your default at the end of a contract project. Skills files in a repo are the opposite: commit the ones the team agreed to, review them like code, and do not let an agent rewrite the skill that tells it not to touch production. That file is a guardrail only if a person owns the merge. When a connector fails, reconnect it. Do not paste a refreshed token into chat as a debugging step. The token is the access. The chat is a log.