How to build an AI character: make the model interview you instead of asking it to create one?

Why One-Shot Personas Fall Flat

You type "You are a skeptical CFO reviewing a SaaS pitch" and hope the model conjures a convincing character from those nine words. Most of the time, it doesn't — and the research explains why.

A PromptHub analysis of role-prompting studies found that simple "You are an X" personas deliver no accuracy improvement and sometimes degrade performance. One paper, "Role-Playing is Not All You Need," tested four model families across thousands of fact-based questions and found system-prompt personas produced no gains — occasionally negative effects. Gains from more sophisticated techniques hovered around 0.5–2%.

Filling in a template doesn't fix the problem either. A vendor taxonomy of AI persona generation describes template-filled output as "a starting hypothesis, not a finding" — and critiques research-synthesizing tools with a sharper sting: "the AI is telling you what you already told it, formatted more neatly."

The deeper issue is what the default approach asks the model to do:

• Invent rather than elicit — the model fabricates character details from training-data patterns instead of drawing out what you actually know
• Produce one-shot output, when peer-reviewed research identifies incremental prompting — gradually expanding persona information — as the more established technique
• Skip primary data, despite quality guidelines stating personas should be grounded in primary data about the population they represent

Academic analysis of 83 persona prompts across 27 studies confirms the pattern: one-directional generation dominates practice, and interview-based construction barely registers. As one paper puts it, the design space for prompts is "close-to-infinite" — yet almost everyone defaults to the weakest option.

The contrast with detailed personas is stark. PromptHub's analysis found that exhaustive, comprehensive personas can produce real accuracy gains, and LLM-generated personas "almost always performed better" than human-written ones. The character isn't the problem — the construction method is.

This is where the interview pattern comes in. An AI interview coach documented in Lenny's Newsletter flips the dynamic: it asks questions one at a time, using reflective prompts like "When were you at your best at work?" to surface material users "didn't know they had." The same principle applies to character-building — the model should interrogate you, not improvise for you. SkillzBench is building toward exactly this kind of workflow: verified, benchmarked agent skills that can turn a chat model into a structured interviewer rather than a one-shot generator.

Flip the Workflow: Let the Model Interview You

The best AI characters rarely come from asking a model to invent one. They come from the model asking you questions until the material surfaces — details you didn't know you had, framed in a way no one-shot prompt could extract.

This inverse-prompting pattern is already working in practice. An AI interview coach built by Noam Segal and documented in Lenny's Newsletter doesn't generate answers upfront. Instead, it walks users through the process "one question at a time," using reflective prompts like "When were you at your best at work? What made it different?" to surface stories users didn't realize they had. Its design principles are explicit: self-reflection before feedback, powerful questions over prescriptions, co-creation over telling.

The same structure maps directly onto character building. Rather than typing "You are a skeptical CFO" and hoping for depth, you instruct the model to interview you about the character's background, motivations, and voice — then synthesize what you gave it.

The evidence behind this approach is strong:

• An academic analysis of 83 persona prompts across 27 research articles found that interview-based, model-asks-user persona construction was almost entirely absent from common practice — meaning most people are still using the weaker one-shot method.
• The same paper cites peer-reviewed findings that personas generated in collaboration between LLMs and humans evoke more empathy than those generated by either party independently.
• Incremental prompting — gradually expanding character information through successive prompts — is documented as an established technique for richer output.

Why does the interview beat the generation? Because simple "You are an X" prompts don't improve accuracy and can actively hurt it, while detailed, exhaustive personas do. As PromptHub's analysis puts it, a basic persona "is definitely not going to increase accuracy" — but a comprehensive one can produce real gains. An interview is simply the most efficient way to produce that detail, because it pulls from your actual knowledge instead of the model's training-data averages.

Two practical details make the pattern work. First, keep the questions sequential — one at a time, not a wall of twenty. Second, maintain a persistent state file (the interview coach uses a coaching_state.md memory file in a Claude Code setup) so the character survives across sessions and the model remembers what it already learned about you.

For agent builders, this is worth noting: the interview coach wasn't a standalone app — it was implemented as an agent skill, using instruction files and persistent memory. That's exactly the kind of skill SkillzBench is building to discover, verify, and benchmark, so you can find interview-driven patterns that have already been tested rather than reinventing the workflow yourself.

The Interview-Driven Character Build, Step by Step

Here's the workflow that puts the model in the interviewer's chair — and why each step matters.

Step 1: Scaffold before you start. Don't just say "interview me." Give the model a role and an output template upfront. A peer-reviewed SBES 2025 case study found that this kind of structured scaffolding — assigning a role, feeding structured artifacts, and providing a predefined output template — reduced hallucination, cut time and effort, and improved persona quality and reusability. The template constrains the model to synthesize what you actually said, not what its training data suggests.

Step 2: Run the interview sequentially. Ask the model to pose one question at a time and use reflective prompts — the same pattern documented in an AI interview coach that surfaces stories people "didn't know they had." Your job is to feed it primary material:

• Sales call summaries and customer conversation transcripts
• Support tickets, reviews, and complaint language
• Real quotes, objections, and decision moments
• Analytics or behavioral data the model has never seen

This matters because practitioner guidance is blunt: the AI "hasn't talked to your actual customers." Eliciting real material beats letting the model invent — template-filled output is "a starting hypothesis, not a finding."

Step 3: Synthesize exhaustively. After the interview, have the model draft the full character profile from your answers. Research analyzed by PromptHub shows thin "You are an X" personas don't help — and can hurt — while detailed, exhaustive ones produce real gains (one study showed roughly a 10-point accuracy improvement, from 53% to 63% on the AQUA dataset). LLM-generated personas almost always outperformed human-written ones, so let the model do the drafting.

Step 4: Validate before deploying. Experts warn: "Check these with your customers before using this." Treat the profile as a draft until it's been tested against real interactions.

A ready-to-use starter prompt:

You are an expert interviewer building a character profile. Interview me one question at a time — never more than one at a time, never generate content on my behalf. Use reflective follow-ups to surface specifics: real quotes, objections, decisions, and moments I may not think to mention. Ask about primary material I can share (calls, transcripts, reviews). When I say "synthesize," produce an exhaustive profile in this template: [background / motivations / pain points / language patterns / decision process / representative quotes]. Flag anything you inferred rather than elicited.

One more note: this pattern works as an agent skill. The interview coach above ran as a SKILL.md instruction package with a persistent coaching_state.md memory file — meaning the character you build survives across sessions. SkillzBench is building a free directory and MCP connector for exactly these kinds of skills, so you can discover, verify, and benchmark interview-style skills before loading them into your agent.

Skills, proven.

Validate Before You Deploy

The interview ends, the model synthesizes a polished character file, and everything about it looks finished. That polish is precisely the trap: a character built from your answers is still a hypothesis about your audience, not a finding, and treating it as done is where most AI characters fail in production.

B2B persona expert Ardath Albee didn't mince words when reviewing AI-generated personas: "You didn't validate this. Check these with your customers before using this. I don't trust this." Her reaction reflects a broader consensus in practitioner guidance — AI hasn't talked to your actual customers, hasn't sat in on your sales calls, and hasn't seen your campaign analytics. The interview pattern narrows that gap by eliciting real material from you, but it doesn't close it. Andy Crestodina's warning applies to the output: if your AI persona is inaccurate, everything the model says through it will be inaccurate too.

Peer-reviewed persona research sets the quality bar explicitly. Good personas should be based on primary data about the population they represent, and their development process should be logically justifiable and empirically scrutinizable. An interview-derived character can meet both criteria — but only if you can trace every trait in the file back to something you actually said or supplied during the session. Any detail the model invented to fill a silence is a liability, not a flourish.

There's also an empathy trade-off worth respecting. The SBES 2025 case study found that while cognitive empathy was strongly supported in AI-assisted persona creation, affective and behavioral empathy varied significantly across participants. The authors cite earlier research warning that excessive automation risks compromising contextual awareness and empathy. A structured validation pass is how you restore that human context before the character goes to work.

A practical pre-deployment checklist looks like this:

• Trace every trait back to a source — an interview answer, a transcript, a review, a sales call summary. Flag and delete anything the model invented.
• Check the character against real interactions: read it alongside actual customer emails, call notes, or support tickets and note where it rings false.
• Show it to one real customer or colleague who knows the audience, and ask specifically what's missing or exaggerated.
• Deploy it on a low-stakes task first — a single email draft or support reply — and evaluate the output before rolling it into daily workflows.

One more reason to validate: the interview pattern is easy to package as an agent skill, with instruction files and a persistent state file that carry the character across sessions. Once a character is embedded in an agent's setup, errors compound silently with every run. SkillzBench is being built to discover, verify, and benchmark free agent skills like these — pinning versions, security-scanning, and measuring each skill's lift in a sandbox — so you can see which ones actually help before they touch your workflow. If you want interview-driven character building packaged and tested for you, join the waitlist and be first in line when the directory and free MCP connector go live.

From Character to Capability: Skills That Do This for You

The interview-coach pattern isn't just a prompting trick — it was shipped as a working agent skill. The documented implementation runs as a Claude Code project built on SKILL.md instruction files, with a persistent coaching_state.md memory file that tracks progress across sessions. Based on interviews with 30+ tech professionals, it asks questions one at a time, scores answers on five dimensions from substance to differentiation, and holds a story bank of at least 8 stories before mock interviews begin.

That matters because it shows the technique generalizes. The same structure — sequential questioning, reflective prompts, state tracking — works for building buyer personas, brand voices, product characters, or any agent that needs to extract what's in your head instead of inventing from training-data patterns. And the economics are hard to ignore: human career coaches charge $300/hour, while the AI coach is essentially free and takes about five minutes to set up.

If you want to build an interview-driven skill of your own, the anatomy is simple:

• Instruction files (SKILL.md) that define the model's role and its interviewing behavior — powerful questions over prescriptions, co-creation over telling.
• A persistent state file so the skill remembers what was asked, what was answered, and where the session left off.
• An explicit output template the model fills only after the interview, keeping responses consistent and in-domain.
• A validation step, because experts warn AI-drafted personas must be checked against real customers before use.

The challenge for builders isn't writing these skills — it's knowing which ones are worth running. Agent skills are everywhere, but published research found roughly 26% carry at least one vulnerability. Nobody can tell which packages actually help and which are risky, and a skill with persistent memory files deserves extra scrutiny before you let it near your agent.

This is the gap SkillzBench is building to close. As the platform launches, it will continuously crawl and dedupe skills from GitHub, pin every version to a commit SHA, run security scans and license checks, and benchmark each skill in a sandbox — testing with and without the skill, using the same model, and publishing the Lift in percentage points. A trust score from 0 to 100 will combine security, quality, license, maintenance, and community signals, with Lift always shown separately and money never touching scores. Everything — the directory, the MCP connector, benchmarks, alerts — will be free.

Skills, proven. If you're building interview-driven skills or want to equip your agents with ones others have already built and verified, join the SkillzBench waitlist to browse, verify, and benchmark free agent skills the moment the platform goes live. The waitlist is open now — early subscribers will get launch alerts and first access to the free directory and MCP tool.