AI Guidelines

AI Guidelines are account-wide instructions that shape how SiftHub writes. Set them once and every answer follows them — instead of your team making the same corrections on every questionnaire and every proposal.

If you find yourself repeatedly fixing the same thing — a product name that no longer exists, a link it should not use, answers that come back longer than you want — that belongs in AI Guidelines.

Where to find them

Go to Settings > AI Guidelines.

Note: Only Admins and Owners can edit AI Guidelines. They apply to the whole account, so a change affects everyone.

The three tabs

Guidelines are set separately for each of the three places SiftHub generates answers, so you can hold a different standard in a customer-facing proposal than in a quick Slack lookup.

Each tab has its own field with a 10,000 character limit, and its own Save and Discard buttons. Saving one tab does not change the others.

What to write

Write each guideline on its own line, as a plain instruction. Below are the kinds of rules that work well, with examples you can adapt.

Naming and branding

The most common fix. SiftHub draws on many sources, and older documents often use names you have since retired.

• Refer to the company as "Acme", never "Acme Inc." or "Acme Corporation".
• Our product is called "Acme Vault". It was previously "SecureStore" — never use the old name.
• Write "Acme Platform" in title case, not "acme platform".

Terminology to use and avoid

• Say "customer", not "client" or "user".
• Never describe the product as a "chatbot" — use "AI assistant".
• Use "subscription", not "licence".

Links and sources

• Only link to pages on help.acme.com.
• Do not include links to internal Confluence or Notion pages in any answer.
• Never cite a source document by its internal filename.

Tone and length

• Keep responses short and crisp — three sentences unless the question needs more.
• Avoid marketing language and superlatives such as "best-in-class", "seamless" or "world-leading".
• Write in British English.
• Use a formal, factual tone. No contractions.

Structure and format

• For yes/no questions, lead with a direct "Yes" or "No", then the supporting detail.
• Use bullet points when listing more than three items.
• Do not use headings inside an answer.

Facts to enforce or avoid

This is where guidelines stop being style and start preventing mistakes.

• Do not mention the legacy Reporting module — it was deprecated in 2025.
• Our data is encrypted in transit with TLS 1.3 and at rest with AES-256. Always state both when asked about encryption.
• Never commit to a roadmap date. If a capability is not live, say it is on the roadmap without giving a timeline.
• We hold SOC 2 Type II and ISO 27001. We do not hold FedRAMP — never imply otherwise.

Compliance and disclaimers

• When asked about data residency, note that EU data can be hosted in Frankfurt.
• Never state or imply that customer data is used to train models.
• For questions about contractual terms, say that terms are subject to the signed agreement.

Writing guidelines that work

• Be specific. "Keep responses short and crisp" works better than "be concise". "Never use the name SecureStore" works better than "use current product names".
• One instruction per line. A paragraph mixing four rules is easier for a person to misread and for SiftHub to apply unevenly.
• Say what to do, not only what to avoid. "Say customer, not client" is clearer than "don't say client".
• Only include rules that apply to every answer. Anything true of one deal or one questionnaire belongs elsewhere — see below.
• Put your hard rules first. Correctness rules — banned claims, certifications you do not hold, deprecated products — matter more than tone. Lead with them.
• Revisit them. When a product is renamed, a certification is achieved, or a claim becomes untrue, the guidelines need updating too.

What not to put here

AI Guidelines are for rules that hold across your whole account. Anything narrower has a better home:

• Deal or project-specific context — what this buyer cares about, which sections to skip, the name of the incumbent vendor. Use the project’s Instructions instead.
• One-off formatting for a single questionnaire — use the instructions on that Autofill run.
• Source content. Guidelines shape how answers are written; they are not a place to store facts. Facts belong in your Q&A Repository and documents.

Reusing guidelines across tabs

Most accounts want the same rules everywhere. On the AI Teammate and Bots tabs, tick Replicate from Response Generation and that tab uses whatever is set on the Response generation tab.

This is a live link, not a copy. Update the Response generation tab later and the tabs replicating it update at the same time. Untick the box to write something different for that surface.

How guidelines interact with other instructions

AI Guidelines are the baseline, not the last word. When something more specific is supplied for a particular answer — instructions on a project, or supporting documents attached to it — that more specific guidance takes precedence for that generation.

In practice: use AI Guidelines for the rules that should never change, and project instructions for what makes one response different from another.

Saving

Click Save to apply your changes, or Discard to drop them. Guidelines apply to answers generated after you save — answers already in a document are not rewritten.