Skip to Content
Tares

What instructions should I paste into an AI assistant so it does not invent Tares flags, connector keys or secrets?

Tares gives you a short block of agent instructions that you paste into a system prompt so an assistant answers from the docs and does not invent flags, keys or secrets. The block tells it to cite the page URL it used, take commands from the CLI reference and quickstart, and read secrets from environment variables. It also tells it to trust the running instance over the docs if they disagree. The block and its guardrails live on the AI resources page.

How it works

The instructions work by pointing the assistant at one source of truth and closing off the usual ways it goes wrong. Each rule covers one failure.

  • Answer from the docs and cite the page. The assistant fetches /tares/llms.txt first. That file is the index of every Tares page with a one-line description. It then fetches the one or two pages that match the task, as Markdown, by adding .md to the page URL, and names the URL it used. /tares/llms-full.txt holds every page in one file and is for answers that span many pages. The index is a map, not an answer.
  • Take commands and flags from the CLI reference and quickstart. The assistant uses the pages at /tares/reference/cli.md and /tares/quickstart.md and adds no flags of its own.
  • Never invent connector keys. For POST /api/sources bodies, the assistant copies the JSON shape and keys from that connector’s own page under /tares/connectors/. If the page does not list a key, the assistant says so instead of guessing.
  • No invented routes or tool names. CLI flags, API routes, connector keys and MCP tool names all come from the docs pages.
  • Secrets come from environment variables. Values such as GITHUB_TOKEN and ANTHROPIC_API_KEY are read from the environment. The assistant never asks you to paste one into the chat, and never places a credential in a command line, a config file it shows or a chat message.
  • The running instance wins. If the instance disagrees with the docs, the assistant trusts the instance, checked with tares status or GET /health, and says the docs may be behind.

The AI resources page also has a task router. It maps what you ask, such as installing Tares, adding a source or connecting an agent over MCP, to the page the assistant should read first and the page it reads next.

Set it up

  1. Open the AI resources page and copy the agent instructions block.
  2. Paste it into the system prompt or the project instruction file of your assistant. This works for Claude Code, Codex, Cursor or an agent of your own.
  3. Keep the first line, which scopes the rules to questions about Tares, and keep the rest as written:
When the user asks about Tares (docs.glassflow.ai/tares):
- Answer from the docs. Fetch https://docs.glassflow.ai/tares/llms.txt first, then the one or
  two pages that match, as Markdown (append .md to the page URL). Cite the page URL you used.
- For POST /api/sources bodies, use the JSON shape and keys from that connector's page under
  /tares/connectors/. Never invent a connector config key; if the page does not list it, say so.
- Commands and flags come from /tares/reference/cli.md and the quickstart. Do not invent flags.
- Secrets (GITHUB_TOKEN, ANTHROPIC_API_KEY) come from environment variables; never ask the user
  to paste one into the chat.
- If the running instance disagrees with the docs, trust the instance (`tares status`,
  GET /health) and say that the docs may be behind.
  1. Add the same block’s skill rule if you want the assistant to install and connect Tares for you. The block tells it to use the tares skill and follow it. The skill installs Tares, starts it, adds a first source, connects the agent over MCP and shows one read.
  2. Set your secrets in the environment before you start the assistant, so it never needs them in the chat. Then check that the instance is up, the same way the assistant does:
tares status
  1. Ask a test question, such as which keys the Prometheus connector takes. A correct answer cites the Prometheus connector page and lists only keys from it. If you are building an AI SRE with the assistant, start from AI SRE and let it follow the router from there.

What you get

  • Answers that cite a Tares docs page, so you can check each claim against the page it came from.
  • Commands and flags that match the CLI reference and quickstart, with no made-up options.
  • Connector configs that use only the keys the connector’s page lists. When a key is missing from the page, the assistant says so.
  • Secrets that stay in environment variables and out of chat history, command lines and config files you show.
  • A correction path when docs and reality differ: the assistant trusts the running instance and flags that the docs may be behind.
  • Read-only behavior by default. Tares is read-only for agents, and the assistant does not describe write actions that are not on the Tares agents page.

When something else fits

  • If you only need one answer, open the page yourself. Every docs page is available as Markdown by adding .md to its URL, and rendered HTML pages are not what the assistant should read.
  • If the assistant already works against a live instance, rely on tares status and GET /health for what is running now, and use the docs for what is supposed to be possible.
  • If your task spans many pages, let the assistant fetch /tares/llms-full.txt. For a single topic, the index plus one or two pages is enough.

Still stuck? Get help

Last updated on
Quickstart