Connect your AI
Briefings and the write gate
Before an AI writes to the brain, it reads a briefing: the blocks every task of that kind must see, in full, plus related blocks and what the graph is missing. The write gate checks that it did, and that it loaded the right expert.
This is context engineering done once, on the server, instead of in every prompt. Nobody has to remember to paste in the tone of voice or the design system: the task type decides what an agent reads, and the server refuses a write that skipped it.
What a briefing is
get_briefing takes a task type, an optional target (a block key, a name like "Table component" or a route like /pricing) and the task in a sentence. It returns:
- The required blocks, in full. For a UI component, for example: the stack profile, the component's spec, the colour system, the tone of voice and the page pattern.
- Related blocks as excerpts, up to a token budget (by default per task type, between 2,000 and 64,000 tokens).
- Every block as
key@version, so what was read can be cited and checked later. - What the graph is missing for this task.
- The expert for the task, and the sub-skill to read, if any.
- A
briefing_idto pass with the writes that follow.
The same input gives the same briefing.
Task types
| Task type | For | Expert |
|---|---|---|
ui-component | A component in the design system | Design |
ui-page | A page or screen, by route | Design |
content | Copy and content | Brand |
strategy | Strategy, roadmap, business case | Strategy |
research | Personas, insights, research | Users and research |
product | Features, glossary, information architecture | Product |
marketing | Campaigns, channels, content pieces | Marketing |
performance | Analytics, KPIs, feedback | Performance |
build-backend | APIs, data, architecture, decisions | Build |
org-setup | Setting up an organisation's shared context | – |
general | Anything else; takes the expert of its target | – |
The write gate
A write to a core block (one of the thirty-four entities) from an AI session is checked for four things:
- A recent briefing from the same caller, for this project, named by
briefing_idor found as its latest one. Recent means the last two hours. - The right expert. The briefing's expert must be the written block's (or one the briefing also names). A component spec written under a strategy briefing does not pass.
- The expert loaded. The session must have read the expert's method with
get_skilland the project. - The sub-skill. The first sub-skill the briefing names is read, or skipped with a reason in
sub_skill_skipped, which is shown with the change.
A content piece past draft must also carry its brief: who it is for, their need, the moment and the intent.
Rules, learnings and documents bypass the gate on purpose: save_rule and log_learning stay the easiest way into the brain.
Enforce, warn or off
Each project sets the gate under Settings › Brain › AI agents:
| Mode | What happens |
|---|---|
| Enforce (the default) | The write is refused, with the exact call that clears it |
| Warn | The write lands; the result says what to read first, and the session log counts it |
| Off | No gate |
get_adoption reports how sessions behave: how many read a briefing and loaded their expert before their first write.
Without MCP
- Over HTTP, the same briefing is
POST /api/v1/projects/{projectId}/v3/briefingwith{ "task_type": "…", "target": "…", "task": "…" }. A briefing fetched with an API key counts for MCP writes with that key. - For a sub-agent or a tool that cannot reach Epic Context,
get_context_packreturns a paste-ready summary of about 4,000 tokens.
Next
- Experts and skills: what the expert brings.
- MCP tool reference: every tool by group.
- Writing for AI agents: blocks that brief well.
Last updated: