Skip to content

Effective Prompting

How to ask an AI assistant for a TongueToQuill document and get a useful first draft. The same advice applies whether the AI is calling the MCP server or you're pasting Quillmark into the editor by hand.

The minimum useful prompt

A prompt that lands a clean draft contains four things:

  1. The document type. "USAF memo," "MFR," "DAF 1206 nomination," "pass request."
  2. The recipient(s). Office symbol on the chain (20 FW/CC, AFROTC DET 730/CC). Multiple recipients are fine.
  3. The sender. Your office symbol or rank/name if you're signing.
  4. The substance. What the memo is for, in the user's words. The AI fills in the structure; you provide the content.

A complete example:

Draft a USAF memo from 20 OSS/CC to 20 OG/CC. Subject: Range 7 closure 22 Apr. Body should note that the closure is for scheduled depot maintenance, the alternate range assignment for affected sorties is Range 12, and that the POC is Capt Smith at 555-1234. Sign it as Lt Col Jane A. Public, Director of Operations.

That's enough for the AI to produce a near-correct first draft.

What to leave out

  • Letterhead formatting. The Quill handles letterhead. Don't dictate fonts, margins, or spacing.
  • Paragraph numbering. The Quill handles the 1., 2., 3. convention. Just enumerate your points in prose.
  • Signature block format. Name and rank are enough; the renderer applies the hanging indent and line breaks.
  • The date. It's stamped at export.
  • AFI citations you don't actually need. Don't ask the AI to add references unless your memo genuinely cites them — the AI will fabricate plausible-looking ones if pressed.

What to be specific about

  • Direct quotes from regulation. If your memo must cite specific regulatory language, paste it in and tell the AI to keep it verbatim. AIs paraphrase by default.
  • Names and ranks. Confirm spelling. The AI cannot verify these.
  • Dates. Give exact dates. The AI defaults to "today" or to plausible-looking guesses.
  • Office symbols. Verify against your unit's symbology. Format is ORG/SYM, all-caps; the renderer uppercases automatically but the AI may use mixed case.

Verifying AI output

Before you sign, scan for these:

  • Names, ranks, dates. AIs hallucinate these confidently. Check every one.
  • Regulatory citations. Verify against the actual reg. A wrong DAFI number sails through every spell-check.
  • Office symbols. Routing only works if these are right.
  • Numeric facts. Costs, counts, percentages — verify against your source.
  • Tone. AIs tend toward florid. Strip filler. AFH 33-337 is plain.

The render itself can't help with any of these. The Quill guarantees the layout; you guarantee the content.

Prompts that consistently fail

  • "Make it sound more authoritative." The AI adds adjectives and weakens the writing.
  • "Add some boilerplate." Boilerplate is template-owned, not content-owned. If you find yourself wanting boilerplate, the template is wrong for the job.
  • "Write it like a colonel would." This produces parody. Specify substance instead.
  • "Include all relevant AFIs." Without knowing your unit's reg corpus, the AI fabricates.

A prompt template

Draft a {document type} from {sender ORG/SYM} to {recipient ORG/SYM}.

Subject: {your subject, title case}

Body:
- {point 1}
- {point 2}
- {point 3}

POC: {name, rank, phone or email}

Signed by: {name, rank, USAF/USSF}, {duty title if any}

The AI fills in formatting; you've supplied every content decision.

Working with get_specs

get_specs (see Tool Reference) returns a blueprint of the Quill — every field with its type, default, and description — and the server requires it before every create and update, so the AI is always composing against the real schema rather than from memory. If your prompt skipped a field the blueprint marks !must_fill, the AI either:

  • Fills it from context (e.g., infers the letterhead caption from the office symbol), or
  • Returns a clarifying question.

Neither is a guess you should accept unread. An inferred letterhead caption is exactly the kind of plausible-and-wrong output that survives to signature.

Sensitive content

The MCP path sends your draft to whatever AI client you've connected. Treat that as a third-party data flow:

  • For UNCLASSIFIED routine business, this is usually fine.
  • For CUI / PII / classified content, do not use third-party AI unless your deployment has explicitly approved it.

When in doubt, draft by hand.

The deployment's classification ceiling is enforced on this path too: get_specs states the allowed values up front, and create_document rejects content that exceeds them. That is a guard on the marking, not on the content — it cannot tell that an unmarked paragraph is CUI.