How Herald works

Most teams already have the context their AI needs. It's just in the wrong places: a voice guide in Google Drive, pricing in a spreadsheet, the best answer to a tricky objection in one person's head. Everyone ends up pasting their own version into ChatGPT, and every version is a little different.

Herald gives that knowledge one home: a shared context layer for your team and your agents. You write it down once as plain pages. Then every agent your team uses (Claude, ChatGPT, Cursor, Claude Code, an n8n workflow, a script you wrote on a Friday afternoon) connects to the same workspace and asks it for what it needs.

The important word there is asks. Agents don't get your whole library pasted into their context window. They describe the job they're about to do, and they get back the parts that matter for that job, with your rules on top.

The short version

1. You write pages. Company overview, product, pricing, personas, competitors, how you sound, what you can and can't claim. Plain Markdown, edited in the app, or imported from the docs you already have.
2. You publish them. Drafts are yours. Agents only ever see published pages, so you can work on something without it leaking into a customer email halfway through.
3. Agents connect. Teammates sign in from their AI tool. Contractors and automations get an API key. Either way, it's one URL.
4. Agents call get_context before they write. They pass the task in plain words, like "reply to a customer who was charged twice after downgrading". They get back your rules, the pages that kind of task always needs, and the sections of everything else that are most relevant, sized to fit.
5. You watch what happens. Every request is logged, so you can see what was asked, what came back, and where the workspace had nothing to say.

That's it. The rest of these docs are detail on each of those steps.

Three kinds of context

A prompt file treats everything you know as one block of text. A workspace keeps three kinds of thing apart, because each one should reach an agent at a different time.

Facts are most of what you write: what you make, who it's for, how it works, what it costs, what customers have said. Facts are sent when the task calls for them. A reply about refunds gets the refund policy, not your API conventions.

Rules are the things that must never be skipped: how you sound, the words you don't use, the promises you can't make. Mark a page as a rule and it goes out with every single request, whoever is asking and whatever the task. Keep them short. There's more on this in Rules, pinned pages and packs.

Packs say which pages a kind of job needs. Your brand voice page never mentions refunds, so a search for "refund reply" would never find it. A pack for support replies makes sure it arrives anyway.

What an agent actually receives

Here's a trimmed example of what comes back when an agent at a company called Tidewater asks for help with a support reply:

# Instructions from the Tidewater team

Always pull context before writing to a customer.

# Rules (always follow)

## Style rules
_style · full page · page_id: 4c1e…_

#### Words we never use
- "Appointments". We say "bookings".
...

# Reference

## Refund policy
_faq · excerpts · page_id: 9a72…_

> **Usage:** Quote amounts and timelines from this page only.

### Refund policy › Refunds after a downgrade

When a customer moves to a cheaper plan mid-cycle, we credit the difference to their next invoice...

It's ordinary Markdown. Rules first, then reference material. Each page says why it was included, whether it's the whole page or just excerpts, and carries any usage note you attached to it. The agent can follow up by reading a full page or searching for something more specific.

Words you'll see

Workspace. One body of context with its own pages, tags, packs, keys and connection URL. Most companies need one. Agencies and multi-brand companies often have one per client or brand. A workspace belongs to an organization, which is where your people live.

Page. A Markdown document with a title and a few properties (type, tags, a usage note and so on). Pages live in folders.

Section. A piece of a page under one heading. Pages are split at their #, ## and ### headings, and sections are what search actually finds. This is why headings matter so much: see Writing pages agents can use.

Published. The version agents see. Editing a published page creates unpublished changes, and agents keep getting the old version until you publish again.

Rule. A page sent with every request as a hard constraint.

Pinned. A page set to "Always include". Sent with every request, but as reference material, not as a constraint.

Pack. A named kind of task (outbound email, release notes, support reply) plus the pages it always needs.

Tag. A word from your workspace's vocabulary, like a competitor's name or a persona. Sections are tagged automatically when they mention one, and tasks that mention it rank those sections higher.

Private note. A page only you can see. Your own agents get it; nobody else's do.

Where to go next

If you're setting up a workspace for the first time, the Quickstart takes about fifteen minutes and ends with an agent answering from your context.

If someone has already set it up and you just want to connect your tools, go straight to Connecting your tools.

If you're the person who'll own the content, read Writing pages agents can use. It's the page that makes the biggest difference to the quality of what your agents produce.