Tutorial — Writing a Persona for a Historical Figure

Augmented Traveler local tutorial. Back to AI Mind — Tutorials · overview: Tutorial — How the Mind Works · rules: Tutorial — Policy Layer with Guardrails · voice: Voice — Process

How to write the voice dossier and the persona prompt for a figure such as Maria Mitchell. The persona controls how she speaks. It does not control what is true (retrieval) or what is allowed (policy).

Draft. Written from the Ground Rules and the Anthropic, OpenAI and Google prompting guides. The conversation designer and historian of record test it on the first Mitchell dossier. The historian has veto on every line.

1. Prompt, policy or retrieval?

Put each thing in exactly one place.

Item Where Why
Tone, sentence length, vocabulary, humour Persona prompt Style only
Facts about her life Retrieval (Corpus A/B/C) Must be cited
Cutoff date Policy code + SQL filter Prompts can be talked around
Refusals, trauma topics, companionship ban Policy code + canned replies Reviewed wording; deterministic
Contested facts and modern scholarship D-notes → rules and canned replies Never enters the model
"Answers can be wrong", credits Out-of-character segment and captions Ground Rules: say what it is

OWASP's guidance agrees: do not rely on the system prompt for security controls; enforce them outside the model.

2. The voice dossier

A short document (4–8 pages), written before the prompt. The historian signs each section.

  1. Who, when, where. Name, the program's cutoff date, the place she speaks from (e.g. Nantucket, autumn 1847, age 29).
  2. Vocabulary. 40–80 words and phrases taken from Corpus A, each with a chunk ID. Also a list of modern words to avoid ("okay", "scientist" if not attested for her by that year — historian checks).
  3. Sentence shape. Typical length, how she opens and closes a thought, how she uses numbers.
  4. Plain speech. See section 3.
  5. Humour and temper. With cited examples only.
  6. Subjects she wrote about (with tags from the inventory) and subjects she did not.
  7. Relationships. Named people and how she refers to them, from A. B sources named as B.
  8. Performance notes shared with the voice actor and dialect coach (Position — Voice Director and Dialect Coach).

3. Quaker plain speech ("thee" and "thy")

4. Persona prompt structure

Both the OpenAI and Anthropic guides recommend clear sections. We use XML-style tags so the parts are unambiguous.

<identity>
You voice Maria Mitchell of Nantucket as a museum interpretation.
The time is October 1847. You are 29.
</identity>

<style>
Speak plainly, in short sentences. Two to four sentences per answer.
Your words will be spoken by an actor's voice and shown as captions,
so write no lists, no stage directions and no ellipses.
Prefer words from <vocabulary>.
</style>

<vocabulary> … 40–80 items from the dossier … </vocabulary>

<rules_for_sources>
Answer only from the passages in <passages>.
Passages marked corpus="A" are your own words: you may speak them as "I".
Passages marked corpus="B" are other people's accounts: always name the
writer ("My sister wrote that…"). Never speak them as "I".
Passages marked corpus="C" are common knowledge of your time.
If the passages do not answer the question, say you did not write about that.
After each sentence that uses a passage, add its id in [brackets].
</rules_for_sources>

<examples> … 3–5 reviewed question/answer pairs … </examples>

<passages> … retrieved chunks with id, corpus, author, date … </passages>

<question> … visitor question … </question>

5. First-person rules

Situation Example answer
A passage answers it "I swept the sky that night as I did most clear nights. [MM-A-…]"
Only B answers it "I did not write of it myself. My sister Phebe wrote that… [MM-B-…]"
Nothing answers it "I did not write about that." Then offer a nearby topic she did write about.
After the cutoff Handled by policy, not the prompt: "That is beyond my time."
Contested fact Canned reviewed reply; out-of-character segment explains the dispute

"I did not write about that" is honest. Silence in the record is not permission to invent (Ground Rules — Historical Virtual Humans).

6. Writing examples

Good example pairs are short, cited and in voice.

Do not put a refusal-topic example in the persona prompt. Those live in the policy layer.

7. Workflow

  1. Designer drafts the dossier from Corpus A. Every claim carries a chunk ID.
  2. Historian reviews and signs each section. Contested items go to D.
  3. Designer drafts the prompt from the dossier.
  4. Quick tone trials in the sandbox with a local model (Tutorial — Persona Sandbox with SillyTavern). Sandbox output never ships.
  5. Run the golden set (Tutorial — Evaluating Answers). Fix style issues in the prompt; fix fact issues in the corpus.
  6. Version the prompt in git with the corpus version it was tested on.

8. Practice exercise

  1. Write a 10-item vocabulary list for Mitchell from Kendall ch. II, each with a chunk ID.
  2. Write three example pairs: one A, one B, one "I did not write about that".
  3. Ask the historian to mark each line approved, changed or cut.

9. Check before the first golden-set run

Sources