Model a Real Person as a Persona Expert

Time: ~12 minutes
Learning outcome: Build a persona expert that reasons like a specific person, train it from their documents, inspect the learned profile, use it to anticipate a decision, and progressively sharpen the resemblance by feeding it more material.

What you’ll learn

By the end of this tutorial, you will:

• Create a trainable persona expert that represents a specific person
• Choose what documents, decisions, and data to feed it
• Train it with council expert train and read the learned profile
• Use the persona to anticipate how that person would react to a proposal
• Feed it more documents over time so it becomes more like the real person
• Understand how the profile and document retrieval shape its responses

Prerequisites

• Completed Tutorial 7: Ground a Debate in Documents
• Completed Tutorial 9: Author High-Quality Expert Personas
• This tutorial uses --engine mock so the training output is deterministic and reproducible offline. Swap in your real provider (for example --engine copilot) for genuine results.

Consent and privacy come first

A persona expert is only as ethical as the documents behind it. Model yourself, or someone who has given you permission, using material you are allowed to use. Don’t model public figures or colleagues without consent. Everything you feed a persona is processed by your configured AI provider during training and retrieval — see Security and Privacy.

Pedro Fuentes in this tutorial is a fictional, illustrative VP of Engineering. The documents below are invented for the example.

The scenario

You’re about to bring a proposal to Pedro Fuentes, your VP of Engineering. Before the meeting, you want to anticipate how he’ll react: which questions he’ll ask, what he’ll push back on, and what would earn a “yes.” You have a folder of his past decisions, proposal reviews, and meeting notes. Let’s turn that into a persona expert you can consult.

How a persona expert differs from a generic one

A generic expert (Tutorial 5) reasons from a role template — “a VP of Engineering” in the abstract. A persona expert reasons from documents about a specific person: Council distills their communication style, decision patterns, biases, and vocabulary into a learned profile, and retrieves relevant snippets from their documents during deliberation.

The result reflects this person’s priorities, not a generic role’s.

Step 1: Create the persona expert

Create a trainable persona with the --persona flag. Unlike a generic expert, a persona gets its own documents folder and can be trained.

council expert create --persona \
  --slug pedro-fuentes \
  --name "Pedro Fuentes (VP of Engineering)" \
  --role "VP of Engineering" \
  --expertise "Past decisions and their outcomes, Written rationale in proposals and reviews, Team and org constraints" \
  --stance "Pragmatic, outcome-oriented: prefers reversible bets and clear measures of success." \
  --persona-description "VP of Engineering whose decisions I want to anticipate before bringing a proposal"

✓ Expert "pedro-fuentes" created at ~/Council/experts/pedro-fuentes.yaml
  Place documents about this person in ~/Council/experts/pedro-fuentes/docs

The --persona-description is the relationship — who this person is to you. It anchors the persona so it knows the context in which you’re consulting it.

Persona vs. generic

Only experts created with --persona can be trained on documents. If you created an expert without --persona and now want to train it, re-create it with the flag (or see Author High-Quality Expert Personas to decide which kind you need).

Step 2: Gather the person’s documents

The persona is only as good as the material you feed it. Good sources capture how the person actually decides:

• Decision logs — choices they made and the reasons they gave
• Proposal reviews and feedback — what they approved, declined, or pushed back on
• Meeting notes — their questions, concerns, and recurring themes
• Written rationale — strategy memos, design-review comments, postmortems
• Communications — emails, Slack/Teams exports (with consent)

For this tutorial, create two small sample documents. In a real project you’d point Council at the person’s actual files.

Create ~/decision-log-2026-q2.md:

# Decision Log — Q2 2026

- Approved the payments re-platform, but only as a phased rollout behind a
  feature flag. Rationale: a reversible bet; we can roll back per-region if
  error rates climb. Success = checkout conversion holds within 1% during ramp.
- Declined the "big bang" data migration. Asked for a dual-write window and a
  measurable cutover plan first. I don't approve one-way doors without an exit.
- Pushed back on adding a third queue technology. Operational surface area is a
  cost; show me the on-call burden before we commit.

Create ~/proposal-review-notes.md:

# Notes on the Search Revamp Proposal

Good problem framing. Three asks before I sign off:
1. What is the measurable outcome? "Better relevance" is not a target.
2. What is the reversible first step? Ship to 5% and compare, don't rewrite.
3. Who carries the operational load? Name the on-call owner.

I weight the post-launch numbers over the design doc. Ship small, measure, iterate.

Step 3: Train the persona

Feed the documents in with --file (repeatable) and let Council process them. The --file flag copies each document into the persona’s docs folder, then trains.

council expert train pedro-fuentes \
  --file ~/decision-log-2026-q2.md \
  --file ~/proposal-review-notes.md \
  --engine mock

Copying decision-log-2026-q2.md to expert docs...
Copying proposal-review-notes.md to expert docs...
Training "pedro-fuentes" from ~/Council/experts/pedro-fuentes/docs...
  decision-log-2026-q2.md: 88 words
  proposal-review-notes.md: 61 words
✓ Processed 2 document(s) (0 failed, 0 needs review, 0 removed, 149 total words); 0 already up to date.
✓ Persona profile updated.

Council extracted the text, indexed it for retrieval, and asked the engine to distill a persona profile.

Fetch documents from a URL

You can also pull a document straight from the web with --url. The URL’s path must end in a supported extension (.md, .txt, .html, .pdf, .csv, .docx, …):

council expert train pedro-fuentes --url https://example.com/pedro-strategy-memo.md --engine mock

Step 4: Inspect the learned profile

Look at what Council learned. council expert inspect shows both the authored fields and the learned profile distilled from the documents.

council expert inspect pedro-fuentes

Expert: pedro-fuentes
Name:   Pedro Fuentes (VP of Engineering)
Role:   VP of Engineering
Kind:   persona
File:   ~/Council/experts/pedro-fuentes.yaml

Panels: (none)

Expertise:
  Weighted Evidence: Past decisions and their outcomes, Written rationale in proposals and reviews, Team and org constraints

Epistemic Stance: Pragmatic, outcome-oriented: prefers reversible bets and clear measures of success.
Persona:          VP of Engineering whose decisions I want to anticipate before bringing a proposal

Learned Profile (from training):
  Communication Style: Direct and structured, favoring concise declarative sentences with concrete examples over abstract framing.
  Decision Patterns:   Weighs trade-offs explicitly before committing, Prefers reversible decisions when evidence is thin, Anchors choices to measurable outcomes
  Biases:              Optimism about iterative improvement, Skepticism of unvalidated claims
  Vocabulary:          trade-off, iterate, signal, tension, concretely
  Epistemic Stance:    Forms beliefs by triangulating evidence from multiple sources and updates readily when new data contradicts prior assumptions.
  Documents:           2 (149 words)
  Last Updated:        2026-06-21T20:54:45.399Z

Each field of the learned profile feeds the persona’s system prompt and shapes how it responds:

Profile field	What it captures	How it shows up
Communication Style	Tone and structure of how they write	The persona answers in their voice
Decision Patterns	The moves they make when deciding	It reproduces their reasoning shape
Biases	What they lean toward or distrust	It weights options the way they would
Vocabulary	Words and phrases they reach for	It echoes their language
Epistemic Stance	How they form and update beliefs	It updates (or holds firm) like they do

Reproducible numbers

Because you used --engine mock, the profile text above is deterministic. With a real provider, the wording will differ and will reflect your documents — but the structure (style, patterns, biases, vocabulary, stance) is the same.

Step 5: Use the persona to anticipate a decision

Now consult the persona the way you’d consult the real person — a 1:1 conversation:

council chat pedro-fuentes

Ask it to react to your proposal:

You> I want to replace our notification service with a new event-driven
system this quarter. How will you react?

[Pedro Fuentes — VP of Engineering]
Before I say yes, three things. What's the measurable outcome — not
"cleaner architecture," a number I can read during rollout? What's the
reversible first step; can we run it behind a flag for one notification
type and compare, instead of a full cutover? And who owns the on-call
burden for a second system while both run? Show me that and I'm inclined
to support a phased pilot.

Notice how the response reproduces Pedro’s pattern from his documents: a measurable outcome, a reversible first step, and the operational/on-call question. The persona combined its learned profile with retrieved snippets from his decision log and review notes.

Bring the persona into a panel

A persona expert can also sit on a multi-expert panel so it deliberates alongside other roles (a Security Auditor, a PM, another persona). Add pedro-fuentes to a panel and convene it — see Build a Custom Panel and Design and Lint a Panel.

Step 6: Make it more like the real person

A persona sharpens as you feed it more material. Suppose you find a fresh set of staff-meeting notes. Create ~/staff-meeting-notes.md:

# Staff Meeting Notes

Pedro on the reorg proposal: "Show me the on-call burden and the blast radius
before we split the team. I'd rather ship a reversible pilot in one squad and
read the signal than reorganize three teams on a hypothesis."

Train again — Council picks up only what’s new and merges it into the existing profile:

council expert train pedro-fuentes --file ~/staff-meeting-notes.md --engine mock

Copying staff-meeting-notes.md to expert docs...
Training "pedro-fuentes" from ~/Council/experts/pedro-fuentes/docs...
  staff-meeting-notes.md: 42 words
✓ Processed 1 document(s) (0 failed, 0 needs review, 0 removed, 42 total words); 2 already up to date.
✓ Persona profile updated.

The 2 already up to date confirms Council only reprocessed the new file. The profile is updated, not rebuilt — each run folds the new material into what it already knew, and recent documents are weighted more heavily so the persona reflects the person’s current thinking.

See the full corpus that backs retrieval with council expert docs:

council expert docs pedro-fuentes

filename                  words  processed                 status
------------------------  -----  ------------------------  ---------
decision-log-2026-q2.md   88     2026-06-21T20:54:45.395Z  processed
proposal-review-notes.md  61     2026-06-21T20:54:45.398Z  processed
staff-meeting-notes.md    42     2026-06-21T20:55:02.978Z  processed

Profile counts vs. the full corpus

After an incremental run, the Documents: line in council expert inspect reflects the most recent batch the analyzer saw (here, 1 (42 words)). The complete set of indexed documents — what retrieval actually draws from — is what council expert docs lists. To rebuild the profile from the entire corpus in one pass, use --retrain (below).

Rebuild from scratch when you’ve changed a lot

If you’ve replaced or heavily revised the source documents, clear the profile and rebuild from the whole corpus:

council expert train pedro-fuentes --retrain --engine mock

↻ Retrain: cleared profile and tracking for "pedro-fuentes" (2 cleared).
Training "pedro-fuentes" from ~/Council/experts/pedro-fuentes/docs...
  decision-log-2026-q2.md: 88 words
  proposal-review-notes.md: 61 words
  staff-meeting-notes.md: 42 words
✓ Processed 3 document(s) (0 failed, 0 needs review, 0 removed, 191 total words); 0 already up to date.
✓ Persona profile updated.

After a full retrain, council expert inspect reports the whole corpus — Documents: 3 (191 words).

How it works

Two mechanisms combine to make the persona reason like the person:

1. Profile analysis — at training time, Council sends the documents to the engine and distills the structured profile you saw in Step 4. It’s cached and only refreshed when you train.
2. Document retrieval (RAG) — at conversation time, Council retrieves the most relevant snippets from the person’s documents and gives them to the persona as evidence, weighting recent material higher.

For the full mechanics, see Persona Experts and Document RAG.

What you accomplished

• Created a trainable persona expert representing a specific person
• Trained it from documents and read the learned profile
• Used it to anticipate how that person would react to a proposal
• Fed it more material and saw how the profile updates incrementally
• Learned when to retrain from scratch versus train incrementally

Next steps

• Make it a teammate: add the persona to a panel in Build a Custom Panel so it deliberates with other experts
• Sharpen disagreement: revisit Author High-Quality Expert Personas to pair the persona with strong generic experts
• Manage the corpus: un-index stale files with council expert docs <slug> --remove <file> and retrain

Key concepts introduced

Concept	Definition
Persona expert	An expert that reasons like a specific person, trained on their documents
Learned profile	The communication style, decision patterns, biases, vocabulary, and stance Council distills from the documents
Persona description	The relationship between you and the person, used to anchor the persona
Incremental training	Re-running train folds new documents into the existing profile
Retrain	--retrain clears the profile and rebuilds it from the full corpus
Recency weighting	Recent documents count more, so the persona reflects current thinking

Commands used

Command	Purpose
council expert create --persona	Create a trainable persona expert
council expert train <slug> --file <path>	Train (or re-train) from documents
council expert train <slug> --url <url>	Fetch a document from the web and train
council expert train <slug> --retrain	Clear the profile and rebuild from scratch
council expert inspect <slug>	View the authored fields and learned profile
council expert docs <slug>	List the indexed documents backing retrieval
council chat <slug>	Consult the persona in a 1:1 conversation