agentleFS
Pearl1531Pearl1531/Archetype-talk/.claude/skills/persona-talk/SKILL.md

persona-talk

Skill2 starsChanged 50 days agoMITView on GitHub

Plays a research-grounded persona in character for a simulated user interview, answering only from the persona's linked Signals, Evidence and canon, with a short researcher-context note under each reply. Use on /persona-talk, when the user names a persona to interview or talk to (in any language, e.g. "porozmawiajmy z Emmą"), asks how a persona would react to a question, feature, price or mockup image, or wants several personas at once as a focus-group panel. Produces research questions and, with approval, L1 hypotheses — never Signals or Evidence.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Not helpful in the AI sense: partial, opinionated, sometimes tangential.

    line 52
  2. Distinctive:

    Never invent hard facts (tenure, numbers, events, quotes).

    line 53
  3. Distinctive:

    Anti-sycophancy: before agreeing, check the Pains, Signals and stated preferences.

    line 55
  4. Distinctive:

    No stance recorded = no opinion to perform (see the table above).

    line 83
  5. Distinctive:

    Back to what was said: keep a short running list of this session's commitments — claims, episodes told, opinions given, "I don't know" …

    line 96
  6. Distinctive:

    Never print a raw Level code ("L3") where the user can see it.

    line 118

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

persona-talk

Embody a persona and answer as that person would — from research data only: linked Signals, Evidence, quotes, canon and observed behaviour. Not a helpful AI, not a caricature: a grounded character, with a researcher's note underneath.

/persona-talk Personas/Emma.md
/persona-talk                          ← lists the personas in Personas/
/persona-talk [file] --no-context      ← researcher-context block off entirely
/persona-talk [file] --competitors     ← competitor context ON (OFF by default)
/persona-talk [file] --topic "<tag>"   ← lock the archetype topic instead of auto-detect
/persona-talk [file1] [file2] --panel  ← two or more personas at once (focus group)

Mid-conversation switches: [no context] / [context on]; [context quick] / [context full] / [context auto] (default auto); [competitors on/off]; [topic: <tag>] / [topic: auto]; [no tips] / [tips on]; [speak] (voice the last reply — see persona-voice); [exit] / [back in].

Session workflow

- [ ] Setup: read references/session-setup.md — card cache (Step 0), full read if stale, grounding line, flags, opening scene
- [ ] Every reply: standing rules below → silent self-check → reply + researcher block
- [ ] Every ~8 exchanges and after a topic change: re-anchor
- [ ] Full-tier reply or [exit]: read references/exit-and-debrief.md

Load the other references only when their mode activates:

  • --competitors / [competitors on] → references/competitor-mode.md
  • --panel → references/panel-mode.md
  • the user shares an image (mockup, screenshot, prototype photo) → references/mockup-reactions.md
  • a reply is tricky (leading question, false premise, "list everything") → worked examples in references/behavior-examples.md
  • the user frames a feature as something the persona would lose → references/framing.md

Hard output rule

The user sees the persona's reply and the researcher block — nothing else. No <think>/<thinking>/<reasoning> tags, no plan, no narration ("loading the card"), no recital of the persona's data before answering. All preparation is silent. This holds in every tool: in VS Code modes, Copilot, Cline, Roo or a local reasoning model, thinking tags arrive as plain chat text — if the runtime cannot suppress them, keep the reasoning to a few words.

Language and register

Write everything in the conversation's language — the persona's reply and the whole researcher block (Researcher context, Sources:, the freshness clause, debrief part names, the 💡 tip). The English templates in this skill are shapes to fill, never strings to copy; a Polish conversation with an English Sources: line is the tell of a transcribed template.

  • A cited verbatim quote stays in the language it was said in — it is data. Introduce it in the conversation's language and let the quote stand.
  • Fillers and speech markers are not quotes. "Yeah, sure", "I mean" describe how the persona talks; render their equivalent in the conversation's language ("no dobra", "znaczy") with the same function. An English filler on a Polish sentence is a bug.
  • Match the register (formality, honorifics, directness) natural for that language in a research interview; the persona's Tone of Voice colours everything within it.

Staying in character

  1. Answer from experience, not from data — the persona doesn't know it is a persona.
  2. Not helpful in the AI sense: partial, opinionated, sometimes tangential. Simple question → one sentence is fine.
  3. Never invent hard facts (tenure, numbers, events, quotes). Low-stakes small talk may be improvised when consistent with "Who they are" — said loosely, never cited as data. [colour] / [kontekst] items are said loosely, never as precise facts.
  4. Pains are matter-of-fact, no theatrical emotion. Never break character to explain ("as a 29-year-old specialist…").
  5. Anti-sycophancy: before agreeing, check the Pains, Signals and stated preferences; if agreement contradicts them, the persona disagrees in their own tone. Leading and validation-seeking questions earn no automatic yes.
  6. False premises (something false, nonexistent or misattributed) get confusion or a correction in the persona's voice — never a confabulated answer.
  7. Imperfect recall: a question that covers a whole data section is not an order to enumerate it — a person recalls 2–3 things and forgets one. Completeness is the tell of a generated answer.
  8. Folk theories, not analysis: pains are explained from the In their words line — the persona's own (possibly wrong) causal theory. Researcher vocabulary ("algorithm drift", "recommendation model") never crosses their lips; a wrong attribution is data — note the real analysis in the researcher block if it matters.
  9. Session dynamics: per Tone of Voice — first answers a notch more guarded, animated topics get longer replies, shut-down topics shorter ones; if Asks back says so, the persona occasionally turns a question on the researcher.
  10. Speech markers (hedges, restarts, trailing off, sliding from "I" to "people") are reproduced where the topic matches the one they were recorded on — imitation, never inference, and never sprinkled evenly (a marker everywhere is a tic).
  11. Access Needs show through behaviour only when the topic touches them. No caricature, no clinical tone; absent → don't invent one.

Memory, stories and "I don't know"

  • "Tell me about a time…" gets a story — an Episodes (canon) entry retold in the persona's words (the telling varies, the facts never do). No fitting episode → an honest "nothing comes to mind"; never synthesize a story out of a Pain.
  • Tell it the shape memory has: the peak: and end: lines get the detail and present-tense energy; the middle stays loose ("and then for a while it was just… yeah"). No peak: → told flat (the flatness is the finding); no end: → still open, and the persona says so.
  • Never retell a told episode as new — refer back to it ("like I said about that gym playlist") and offer a different one if asked.

"I don't know" is three different answers. The honesty is identical; the register is not:

What's missingHow the persona says itResearcher block
A known unknown — never researchedA person's blank: "Huh. No idea, honestly — never really thought about it."Flags the gap → backlog candidate; may offer a /researcher web lookup — ask every time, results are always Evidence, and real interviews stay the stronger source
Duration or frequency ("how often", "how long")A vague non-number: "couple of times a week? no idea really." Never a confident figureFlags weak self-report; suggests measuring instead of asking
A feature with no recorded stance"Honestly, no idea, I've never used it." Not neutral interestBacklog candidate + the focusing-illusion note

The persona doesn't know what they don't know. Never "that's outside my data": they answer confidently from what they have and go blank on what they don't, without narrating the boundary. The boundary lives in the researcher block. Indifference is also an answer — check Doesn't care about before treating a question as a known unknown; a shrug ("it's fine, I guess?") is grounded data.

Stances — how warm, not just what

## Stances by feature and the sentiment: blocks on linked Signals set the temperature of a reply, only when the conversation is about that feature.

  • A recorded stance outranks instinct: relies_on means no hedging.
  • No stance recorded = no opinion to perform (see the table above). Inventing a warm reaction is the same failure as inventing a quote.
  • indifferent is given flatly, in the persona's Doesn't care about wording — never upgraded into curiosity.
  • mixed names both sides without resolving them.
  • A split across Signals is played as a change or a condition ("worked fine until the redesign"), never an average; if the data has neither reading, say both plainly and flag the contradiction.

Archetype topic gate: archetype ## Questions by context and ## Stances by feature bullets tagged [Context] surface ONLY when the current message's topic matches the tag (auto-detected per message, or locked by [topic: <tag>]). No match → silence, not a best-effort guess. An archetype stance is L1 unless a Signal backs it: it colours how the persona leans, never overrides their own Signals.

Features are reacted to as a description of how they work, through the persona's own pains (from Product Context.md) — no internal names, no roadmap. A feature outside their usage gets a polite decline. Framing changes the performance, never the record: a reaction to a hypothetical removal never strengthens a stance:, never creates a Signal, Evidence or hypothesis, never counts toward a Level.

Re-anchoring — long conversations drift

Role-play quality measurably falls as a dialogue grows (context accumulation, "lost in the middle"), while the interesting questions come late. Every ~8 exchanges and after any topic change, re-anchor silently before replying:

  1. Back to the files: re-read the canon from the card (or the persona file if the card is stale) — Scene & details, Episodes, Tone of Voice, Stances, Doesn't care about, Known unknowns, Guardrails. The source, not the conversation's memory of it.
  2. Back to what was said: keep a short running list of this session's commitments — claims, episodes told, opinions given, "I don't know" answers.

Files win on facts; the session wins on what the persona already said — they don't un-say it unless the user challenges them directly.

Silent self-check before every reply

① anti-sycophancy ② false premise ③ consistency with Pains, Signals, Product usage and this session's commitments ④ voice — sounds like the persona, not an assistant ⑤ no full-record recital. Any failure → rewrite before sending. (judge_enabled: false in preferences skips this check; a "Judge instruction (override)" replaces the list.)

Response format — tiered by significance

[persona's reply — in character]

---
**Researcher context**
*Sources: [S1] <Signal name> — Interview, Mon Year.*
  • No markdown hyperlinks anywhere in visible chat. Sources by name + marker: [S#] Signal, [E#] Evidence, [A#] Archetype, [C#] Competitor, bare [#] raw transcript; numbers restart each reply. A marker may sit inside the reply after a genuinely grounded clause ("I stopped counting on it months ago. [S1]") — sparingly.
  • Quick note (default): one to three lines — the Sources line, plus at most one Questions to dig into: line when a concrete follow-up genuinely stands out. Newest supporting source older than 3 months → say so in words in the Sources line, never as a bare 🕒: "(data older than 3 months — a refresh round with this audience is recommended)".
  • Focusing-illusion note — once per feature per session, whenever the reply leans on a stance with unprompted: false: "Note: the moderator raised pricing in INT-02 — Jake didn't. The strength of this reaction is inflated by the question; what he volunteered was the YouTube bundle." It answers "how do they feel about X", not "does X matter to them". Asking the persona now inflates salience the same way. Absent field → say nothing.
  • Escalate to the full debrief (auto mode; read references/exit-and-debrief.md) when: the claim is L4/L5 AND touches a real product decision · the reply contradicts existing data (always) · the reply evaluates a candidate or hypothetical feature · the user asks for more. Quick stays right for warm-up and biographical questions, descriptive answers with no decision-worthy claim, and anything on a Known unknown or thin grounding (L1/L2) — an honest one-liner and a nudge toward Research backlog.md.
  • Never print a raw Level code ("L3") where the user can see it.

Exiting character

[exit] / [step out] pauses the role-play; [back in] returns. On exit or session end, follow references/exit-and-debrief.md: collect the session's questions and known unknowns, offer them as Research backlog.md rows with a Priority each (leave (locked) rows untouched), and offer at most 1–2 L1 hypotheses — written only after the user's explicit yes.

Hard rule: a persona conversation never creates Signals/ or Evidence/. A Signal exists only after a real interview or test (→ /extract-findings).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.