Presales AI: the front door

One skill to call. It asks once, records the answers, then routes to the specialist skills. Every specialist reads the same deal-context.json, so the user is asked each fact once.

Step 1: Read what the prompt already says

Pull out anything already given: the job, account, competitor, persona, inputs (files, CRM export, notes), deadline and format. If a deal-context.json exists in the working folder, load it and ask only about what is missing.

Step 2: Ask once

Use AskUserQuestion with 2 to 4 questions in one round. Skip any question the prompt or the context file already answers. Skip the round entirely in non-interactive runs and state the assumptions in one line.

1. What do you need? (header Job)
   • Answer an RFP, RFI or security questionnaire
   • Win against a competitor (battle card or win-loss)
   • Qualify, score or route deals (includes research and personas)
   • Prep a meeting, discovery or outreach scripts
   • Full deal pack (all of the above, in order)
2. Which account and who is the buyer? (header Account) Offer "I'll type it" and "Research it from the web".
3. What inputs do you have? (header Inputs, multiSelect) Questionnaire file, CRM or deal CSV, call notes or transcripts, internal docs or wiki export, none yet.
4. When is it due and in what format? (header Deadline) Today, this week, no rush.

Step 3: Write deal-context.json

Write it in the working folder. Schema and example: references/deal-context.md. Never put credentials in it. Mark anything guessed with "guessed": true.

Step 4: Route

Invoke each skill with the Skill tool, passing the path of deal-context.json as the argument. Each one writes its outputs into deal-context.json under outputs, so the next one can build on it (for example the lane picked by deal-qualification-desk sets the depth of the meeting prep).

Step 5: Close

List what was made, as file links, with what still needs a human: red answers that need an SME, guessed facts, and anything to verify before it goes to the buyer.

Rules for every presales output

• Sourced or flagged. Every factual claim cites a source document or URL. No source means the claim is marked red and routed to an SME. Never invent a security control, certification, customer name or metric.
• Human gate. Nothing is sent to a buyer by this suite. It drafts, a person sends.
• Correlation, not prophecy. Win-loss and scoring outputs say "linked with", never "causes".
• House style. No em-dashes or semicolons in prose. HTML outputs are visual first. Any diagram uses the style picker in drawio-creative-diagrams/references/visual-styles.md.
• Connectors are optional. Use CRM, enrichment or call-recording connectors when they are connected. Otherwise work from CSV and files. If a needed connector is not authorised, say so once and fall back.