Prompt guidance

A Process-Flow AI step is usually authored as one or two sentences — "Create a project charter and scope statement". That is the right thing to write in an extension: it is readable, and it stays correct across every tenant that installs it. It is also thin, and the same sentence has to serve an organization with its own policies, terminology, constraints and prior decisions.

Prompt guidance closes that distance. When it is enabled, the platform takes the step's short prompt and sharpens it against knowledge the run already has access to — the process's own knowledge bases and, when the step asks for one, your Company Context briefing — and records the result as a page you can read.

What it changes for the person running the flow

Nothing they have to do. The preparation runs ahead of the step, not when someone presses Generate, so the step itself is no slower than before. When guidance is ready, the step's panel shows it and links to the page it came from; when it is not, the panel shows nothing rather than an empty state.

The guidance is not editable before the step runs. Because the preparation happens off the interactive path, there is no moment where a person is holding the compiled prompt with the run still pending. The page is the durable record instead — it survives a refresh, and it is what an auditor reads afterwards.

Which steps it applies to

Only the step types that carry an operator-authored prompt worth sharpening:

What lands in the knowledge base

One page per prepared step, in the run's own instance knowledge base. It holds the compiled guidance, the retrieval questions that were asked, the retrieval intent, and the knowledge-base sources that answered — so "why did the step say that" has an answer that does not depend on re-running anything.

The page is classified, and it can outrank the step. It is written at the higher of the knowledge base's own baseline clearance and the clearance of the identity the step ran as. A colleague whose clearance does not reach that level does not see the page, and does not see the guidance in the step panel either. Classification wins over transparency when the two disagree.

An agent step never reads guidance above its own clearance ceiling. When an agent step runs, the page is read at the lower of the reader's clearance and the step's clearance ceiling. That is the same cap that applies to every document the step is grounded on. If the page is classified above that ceiling, the step runs with its original prompt.

Enabling it

Prompt guidance is an organization policy, written through the organization-policy API under the key agentic.process_flow.query_expansion. An organization admin enables it; the change is recorded with the user who made it. With no policy row present the feature is disabled, which is also what happens if the policy cannot be read for any reason.

Failure is always quiet

Every failure mode degrades to the behaviour you already have, rather than failing the step:

• The toggle is off, or the policy cannot be read → the step runs with its original prompt.
• Preparation errors, or finds no relevant knowledge → no page is written and the step runs with its original prompt.
• The knowledge base changed since the page was compiled → the stale page is refused and the step runs with its original prompt.
• The reader's clearance does not reach the page → the panel shows nothing.
• The page is classified above the agent step's clearance ceiling → the step runs with its original prompt.

Related

• Company Context — the organization-wide facts a briefing is drawn from.
• Action types — including metadata.companyContextBriefing, the per-step retrieval question that asks your organization something.
• Knowledge-base clearance — how page-level classification is enforced on reads.