Working with your squad
Recruiting testers
A squad member can represent a real person, be deliberately synthetic, or combine permitted source material. Recruiting turns that material into a governed persona your AI can play consistently — round after round.
Why an interview?
Anyone can write “pretend to be a busy small-business owner.” That usually produces a stereotype. SQUAT's interview makes the construction explicit: what the persona claims, what may motivate it, how it tends to behave in particular contexts, where its traits conflict, and what is genuinely unknown. For a synthetic persona, those details are designed and labeled as synthetic; they are never disguised as human evidence.
Real and synthetic are both supported
The actor and moderator do not distinguish between a represented-human persona and a synthetic persona while performing. Workspace owners and admins do keep a source-governance record for diagnostic and record-keeping purposes:
- First-person human — material supplied by the human being represented.
- Documentary — operator recollection or observation, support records, research notes, or another third-party source.
- Synthetic — model-authored or deliberately invented material.
- Unclassified — legacy or genuinely unknown material.
Combined is a derived dossier label when two or more classified bases contributed; it is not a source type. This is separate from origin, which records how the persona entered SQUAT: interview, template, or import.
“First-person human” means the represented human supplied the material. An operator's recollection about someone else is documentary, even though the operator is human.
How the interview works
Say “recruit a tester.” With the plugin, /recruit starts the interview; a bare-MCP client can invoke the recruit prompt or use the same words. SQUAT asks questions one at a time, reflecting each answer so you can correct it. It first asks what you intend to construct and what material will support it. It does not push you toward a real named person when a synthetic persona is what you need.
Over the conversation you'll cover five areas, in whatever order your answers lead:
- Who they are — a display name or chosen archetype, practice, context, fluency, and limits to what they know.
- What they're trying to get done — their goals, in their own terms, never your feature names.
- How they behave under friction — patience, workarounds, and their rage-quit threshold. Do they retry, search for help, or just leave?
- Communication range — vocabulary and avoided terms, formality, directness, brevity, uncertainty, question/correction behavior, and contextual variation.
- Churn vs. delight — what would make them abandon your product for good, and what would make them a fan who tells other people.
Short answers are fine. Three sentences per question is plenty to work with — and if you genuinely don't know something, say so. SQUAT records uncertainty rather than inventing detail, which keeps the persona accurate.
When there's enough to work with, SQUAT stops — a tight interview beats an exhaustive one. You should end wanting to recruit another, not relieved it's over.
What gets compiled
The interview produces governed sources plus derived artifacts, all validated and protected automatically by the hosted service:
- Sources — basis, allowed uses, runtime eligibility, active/withdrawn state, and protected material/permission evidence. Operator attestation is never presented as verified consent by the represented human.
- The portrait — goals and context plus separate claims, inferred motivations, conditional behavior patterns, and tensions that preserve contradictions. Self-description and observation are evidence, not ground truth.
- The communication range — an approved range of communication behavior, not a memorized imitation.
- The actor briefing — the provenance-excluding briefing an AI uses to perform. It excludes provenance, source basis, permission records, raw samples, real/synthetic labels, and operator commentary.
Derived artifacts pin their persona, communication-range, and compiler versions, so a completed round always identifies what actually performed.
Validation is automatic. As each piece is saved, SQUAT checks it server-side — the communication range for caricature and protected-trait inference, the actor briefing for provenance leakage — and records its own validation receipt. There is no validation step for you or your AI to perform, and nothing a client could skip: material that fails the check is refused with the reason, and everything that saves has already passed.
Writing samples and synthetic auditions
You can import authorized content supplied by a represented human and use it to calibrate communication behavior. You can also ask SQUAT to generate a short synthetic audition from permitted material. The audition uses neutral prompts and is reviewed for fidelity, overstatement, and stereotype risk before approval.
Every audition is synthetic — even one based on human-provided writing. It is calibration, never evidence, and it can never feed back into the persona’s sources. SQUAT does not infer ethnicity, nationality, age, disability, class, education, intelligence, or other protected or sensitive traits from prose. It does not render phonetic accents, racialized slang, or demographic caricatures. Specifically authorized vocabulary or functional accessibility behavior can be represented when relevant, but never exaggerated into a catchphrase.
The persona record keeps one compact acquisition/source-basis note. Canonical dialogue records separately note whether their transcript was generated in the SQUAT round or imported from an archive. Cards, reports, issues, ledger summaries, and ordinary responses do not repeat those notes.
The dossier
When you ask to see a squad member, you get a dossier: name, archetype, and practice; goals, behavior under friction, tech fluency, habits, “leaves if / delighted if” triggers, an alter-ego when recorded, and version/status details. It's how a squad member is introduced, and the quickest way to sanity-check that the interview captured the persona you intended.
A hosted rich card receives operational metadata only. SQUAT retrieves the full dossier through the same authorized workspace read used by its Markdown view; the card renderer does not receive the profile body. Missing rich UI never means missing information.
Already have a persona document?
If you bring an existing document — research notes, a support-ticket export, a marketing avatar sheet — SQUAT won't force you through a full interview. It asks one question first: import it verbatim, or use it as a baseline?
| Import verbatim | Use as a baseline | |
|---|---|---|
| What happens | The document becomes the complete current source basis; SQUAT derives a separate behavioral portrait. | A short adaptation interview — a handful of questions on what's genuinely different about your users versus what the document describes. |
| What's stored | The complete source document as its own protected source, plus a separate derived portrait. | The complete source document and interview answers as distinct protected sources, plus a derived portrait with traceable dependencies. |
| Best when | The document already reflects your users well. | The document is good raw material but wasn't written about your users specifically. |
| Afterward | SQUAT offers a validation interview, then requires an approved synthetic audition and an active actor briefing before the first round. | The same audition and activation gates apply before the persona can run. |
Preserve, separate, then derive. The complete source document is kept as its own dedicated protected source, never compressed or copied into the runtime profile. The profile is a derived behavioral portrait; the actor briefing is a further provenance-excluding derivative. This preserves source fidelity without exposing raw human or documentary material to actors, cards, or round briefs.
The depth check
The test of a good profile is that it is specific to this persona, not that it resembles a real person. SQUAT checks anchored conditional behavior, useful knowledge limits, distinctive but non-caricatured communication range, real gaps, and preserved tensions. Synthetic specificity is valid; its governance simply remains synthetic.
For documentary material, the fix for a thin persona is targeted operator questioning, not invented human evidence. For a synthetic persona, the fix can be explicit synthetic design. In both cases the basis stays accurate.
Validation interviews
An imported persona is good raw material that hasn't yet been confirmed against your own read of it. A validation interview is a short pass where SQUAT walks the imported profile with you — does this match the persona you intended? — before the persona runs its first round. SQUAT offers one automatically for verbatim imports; you can ask for one anytime.
Learnings and recompilation
Hosted squad members accumulate learnings from rounds — short behavioral observations worth remembering (“found capture via the shortcut this time”). Learnings are what turn a first-time visitor into a returning user. You stay in control of them:
- View — a persona's learnings appear with its dossier, newest first, each tied to the round it came from.
- Pin a learning to always include it, prune one that was a fluke (pruned learnings leave the compilation but stay in history), or accept a batch at once.
- Recompile — when learnings accumulate, SQUAT offers to recompile the actor briefing into a new persona version. It's always offered, never forced; rounds record exactly which version ran either way.
If learnings contradict the interviewed profile — the rounds show more patience than you described, say — the profile wins until you say otherwise. SQUAT raises the conflict as a validation question (“rounds suggest Maya is more patient than you told me; want to update her?”) instead of silently rewriting your persona.
Version history
Every save of a squad member — a fresh interview, an enrichment, a recompile, even a status flip — creates a new version, and the prior version is archived, never overwritten. Scenarios version the same way. Ask “show me all versions of Maya” and SQUAT lists the full timeline (version number, when it was archived, origin, status); ask for a specific version and SQUAT retrieves its complete content; ask “what changed between version 2 and 3?” and SQUAT compares the real archived documents rather than recalling from memory. Rounds pin the exact persona and scenario versions they ran, so historical results always point back to the version that actually performed.
The learning slot on Starter
On Starter, the free plan, your workspace holds up to three squad members. All three are hosted and synced across your devices; one of them holds the learning slot, accumulating what rounds reveal. The others keep their full content, without that carried memory. To change which persona is synced, just say so — SQUAT steers you through the swap, and the displaced persona keeps its content locally.
Managing the roster
- List the squad: “list my testers” — optionally including retired members for the full history.
- Retire a member: “retire Dana” — a status flip, never a content edit. Retired personas keep their full profile and history and can be revived.
- Name a line-up: “create a mobile squad with Dana and Marcus” — then “run a round with the mobile squad” resolves the whole line-up in one phrase.
The starter library may omit generic persona templates, but SQUAT explicitly supports synthetic persona creation and authorized persona imports.