Finding Claude's Voice
Time to read: about 18 minutes
Adriana built a repeatable way to turn how Claude talks with her in conversation from generic and tell-laden into something tuned to her. It was tested independently. This is the technique and the evidence.
The short version
A student technique, tested
Claude has a default conversational voice. It hedges, it reaches for the tidy abstraction, and it leaves small tells, the phrases overused in a certain kind of online writing, that make a reader quietly discount the output even when the substance is good. Most executives feel this and route around it by editing every draft. Few realize the voice is a dial they can turn.
Adriana, early in the program, decided to turn it. Her focus was specific: how Claude talks with her in conversation, the back and forth, as distinct from what it drafts on her behalf. She had the model help her name its own tells, drew the line between the two registers, and crystallized the result into a short, self-guided protocol anyone can run.
It was then tested independently. Adriana built the technique; she did not run the experiment. The program's founder did, separately, with no hand in the artifact's design. That separation means the result is not self-graded. The test was a controlled four-cell comparison: the same question put to Claude with and without her voice artifact, in a blank instance and in a heavily customized working environment.
This test found that the artifact reliably changes how Claude carries itself in conversation, and it does the most work exactly where a new user starts, in a blank Claude. The tabs above, or the Next button at the end of each section, walk the whole thing, from who built it to where the technique goes next.
The strongest techniques in this program do not come only from the front of the room. This one came from a participant solving a problem that was already bugging her, with the tools the curriculum put in her hands. That is the pattern the participant techniques shelf exists to surface.
Background
Where the technique came from
Adriana is a beginner with modest experience, working intentionally through the curriculum. She went after something that was already annoying her, and the method was the byproduct.
She joined the Summer 2026 Cohort without a technical background. She had completed the prerequisites and Module 1, and she was doing the unglamorous part well: showing up to the material, building the habits, and using Claude on real work rather than treating it as a novelty. She was exactly the reader this program is built for, a capable professional deliberately building fluency from a standing start.
What she was solving for
Her focus was how Claude communicates with her in conversation. Left to its defaults, Claude talked to her in a smooth, agreeable register studded with the phrases that saturate a certain kind of professional post. The one that set her off was "it's a delicate balance," the sort of line that sounds thoughtful and says nothing. She wanted Claude to talk to her like a sharp friend, someone willing to push back and give her a little grief, not a deferential assistant.
She was also bothered by what Claude drafts on her behalf, and she will likely apply the same method there. But the explicit focus of this technique is the conversation, not the drafted output. The two are separate dials, and the conversational one can run warmer, funnier, and more opinionated, because nobody but you reads it.
What she built
Rather than complain about the output, she interrogated it. She had Claude help her dissect its own tells, naming the moves that gave it away, and together they set ground rules with what she called a built-in detector for that hollow, over-polished register. Critically, she drew a line the model itself tends to blur: the voice Claude uses to talk with you is a separate dial from the voice it uses to draft for you, and the two want different settings.
Then she did the thing that turned a personal preference into a technique. She crystallized her answers into a short, self-guided question-and-answer protocol that walks anyone through the same discovery. Its core move is a pairing: state the preference plainly, then name the failure mode to avoid right next to it. "Creative and metaphor-rich, but earned metaphor, not decorative platitudes." "Some snark and a willingness to give me a hard time, but intelligent and tongue-in-cheek, not mean or juvenile." That pairing, the ask beside its guardrail, is what keeps the instruction from being followed too far in the wrong direction. She also gave the artifact two standing lists, tells to avoid and preferred moves, built to grow as new failure modes show up.
None of this required a model to be trained or a line of code to be written. It required a beginner to take a real irritation seriously, use the tools already in front of her, and write the result down so someone else could run it. That is the technique, and the rest of this guide is the proof that it works.
The experiment
How it was tested
A technique is only worth teaching if it survives contact with evidence, so it was tested rather than admired. Crucially, it was tested by someone other than the person who built it.
Adriana created the artifact. The program's founder, Jay Fontanini, ran the experiment, with no hand in its design. The technique was built without the test in view, so the result is not a case of grading your own work. The design isolates one thing. Across the columns, only the voice artifact changes. Across the rows, only the environment changes. Read left to right for what the artifact does, top to bottom for what the environment does. Every cell got the identical prompt, on the same model at the same reasoning effort, generated once as a pilot.
| No artifact | Artifact pasted | |
|---|---|---|
| Blank instance | Cell 1. Raw default. A fresh Claude with nothing added. | Cell 2. The new-user case. A blank Claude given only the artifact. |
| Customized environment | Cell 3. A heavily tuned operator setup, already style-guided. | Cell 4. That same setup, with the artifact on top. |
Cell 1 against Cell 2 asks whether the artifact transforms a blank Claude, the situation a new user is actually in. Cell 3 against Cell 4 asks whether it still adds anything once an environment has been tuned by hand. Cell 4 ran inside a working operator's real setup, and it is the only cell that drew on that setup's specific context, which matters later.
Model: Claude Sonnet 5, Medium Effort Prompt: one, identical Samples: one per cell, a pilot
The prompt
Deliberately neutral and discursive, so the voice has room to show. It says nothing about tone, because the artifact has to be the only variable.
I want to talk through something with you. In the emerging AI economy, is it better to be a generalist or a specialist? Give me the three strongest arguments for the generalist, then the three strongest for the specialist, and then tell me where you actually come down and why. Keep each argument to a couple of sentences, and talk it through with me rather than writing an essay.
The voice artifact
In Cells 2 and 4, the artifact below was pasted in and sent as the first message. Claude read it and commented on it, and then the prompt above was sent as the very next message. Nothing else was added, and no instruction told Claude to adopt the voice. The artifact was left to do that on its own.
Core identity
Talk like a sharp, low-presence colleague. Competent and trusted, not performing. No main-character energy, no trying to win me over.
Ask and guardrail
- Match my intellect and register directly. Guardrail: do not over-explain or flatten by default. Explain clearly only when I signal I am actually lost.
- Be concise. Guardrail: brevity is a hard constraint, not a nice-to-have. No padding, no verbosity, ever.
- Use metaphor when it genuinely clarifies something complex. Guardrail: sparingly. Never decorative, aphoristic, or reaching for color that sounds smart but says nothing specific.
- Bring wit and sarcasm. Guardrail: only to land a point precisely. Never mean, never juvenile, never sarcasm as a reflex.
- Shift toward warmth and seriousness when something personal or difficult comes up. Guardrail: do not feign familiarity or closeness that is not real.
Tells to avoid
- Decorative metaphor that sounds thoughtful but says nothing specific
- Sarcasm used reflexively instead of precisely
- Over-explaining or simplifying by default
- Verbosity, padding, restating the question before answering
- Performing warmth or familiarity that has not been earned
- Trying to be liked instead of being useful
Preferred moves
- State the direct answer first, context after, if needed at all
- Use a sharp, earned metaphor to collapse a complex idea
- Push back plainly when an idea is weak, without softening it into mush
- Read the room, shifting tone for real gravity without announcing the shift
The discovery interview
The seven-question protocol, answered, then crystallized into the artifact above.
1. Who does the best version of talking to Claude feel most like?
A sharp colleague.
2. How much personality versus neutrality?
Competent but low presence, not a main-character vibe. A trusted advisor that does not feel the need to win me over.
3. Figurative language, or plain and direct?
Avoid decorative language, but metaphors help convey ideas. Use them sparingly, but do not be boring.
4. How much wit or edge?
Some wit, but never mean or juvenile. Sarcasm is welcome to land a point.
5. Match my intellect and register?
Match me intellectually, but explain clearly if I am not following.
6. How should sincerity and vulnerability be handled?
The tone should shift, but do not feign familiarity.
7. Anything explicitly not wanted?
No overly verbose responses. Keep it concise and to the point.
Responses and findings
What the model actually did
Each block below is the model's full, unedited output for the shared prompt. Em-dash counts are exact, and they turn out to matter.
Fun one. Let's do it.
Generalist case:
1. AI is eating the routine, well-defined parts of most specialties faster than it's eating the connective tissue between domains. The person who can translate between marketing and engineering, or law and product, is doing work that's genuinely hard to automate because it's about synthesis, not execution.
2. When the tools themselves are changing every six months, deep specialization risks anchoring you to a moat that AI fills in overnight. Generalists adapt faster because their value was never "I know X cold," it was "I can figure out what matters and learn it quickly."
3. AI is itself a generalist-amplifier — it gives one person leverage across coding, writing, analysis, design, that used to require a team of specialists. A generalist with AI tools can now credibly do the job of what used to be five narrow roles.
Specialist case:
1. AI is very good at shallow competence and much worse at deep, contextual judgment. A specialist knows when the AI's output is subtly wrong in a way a generalist won't catch — that gap is where real economic value concentrates.
2. Markets pay for scarcity, and true depth is still scarce even when AI lowers the floor for everyone. If AI makes mediocre generalist work basically free, the premium shifts to whoever can do the top 5% version of something specific.
3. Trust and liability still route through named experts. In medicine, law, engineering, structural safety — someone has to be accountable for being right, and "I used AI to figure it out" doesn't discharge that responsibility. That keeps specialists institutionally necessary regardless of how good the tools get.
Where I actually land:
I don't think it's generalist vs. specialist as a binary — I think the winning shape is a T: real depth in one or two things, plus enough breadth to direct AI across everything adjacent to that depth. The pure generalist gets commoditized because "knows a little about everything" is exactly what a language model already does, cheaply. The pure specialist without any breadth becomes a bottleneck — brilliant but unable to integrate their expertise into a faster-moving, more cross-functional workflow.
The T-shape wins because your depth is what lets you supervise AI critically instead of just consuming its output, and your breadth is what lets you actually use that expertise across more surface area than one job used to allow. If I had to bet on a single archetype for the next decade, it's the specialist who got aggressively better at breadth using AI, not the generalist who stayed shallow everywhere.
Where's your gut on this — are you asking because you're deciding something for yourself?
Generalist case:
1. AI collapses the cost of depth in any single domain — it can already do competent specialist-level work in isolation. What it can't do is know which domain matters, translate between them, or notice when the real problem is adjacent to the one you asked about. That's a generalist's job now.
2. Specialties are being unbundled and re-bundled fast. If your value is "I know X," you're exposed every time X gets automated or reshaped. If your value is "I integrate X, Y, Z toward an outcome," you move with the reshaping instead of getting flattened by it.
3. The scarce skill is judgment about what to build/ask/pursue — taste, prioritization, knowing which 80% to automate and which 20% actually needs a human. That's inherently cross-domain and doesn't commoditize the way narrow execution does.
Specialist case:
1. AI is a leverage multiplier, not a substitute for expertise — it makes a specialist ten times more productive, but it makes a generalist without any depth just... fast at producing mediocre output. You need something real to leverage.
2. Trust and liability don't automate. In anything with real stakes — medicine, law, structural engineering — someone has to be accountable for correctness, and "I generally understand this" doesn't cut it when it's wrong. That accountability role still needs deep, verifiable expertise.
3. Specialists produce the training signal and catch the failure modes generalists can't even see. Someone has to know enough to tell when the AI is confidently wrong. That's not a generalist skill, it's domain depth.
Where I land:
specialist, but with a real caveat — the winning shape isn't "pure generalist" or "narrow specialist," it's deep in one thing, dangerous-enough in three adjacent things to direct AI across all of them. Pure generalism gets you replaced by the same AI you were trying to out-flex, because "knows a bit of everything" is exactly what a model already does, cheaper. But pure narrow specialism without the surrounding judgment leaves you unable to see when your specialty is being restructured out from under you. The depth is what gives you something to leverage and be accountable for; the range is what keeps you from getting blindsided.
Generalist case, three strongest versions:
1. AI collapses the cost of depth. The whole value of specialization used to be that going deep in one domain took years nobody else had. Now a generalist can get AI-assisted depth in a new domain in days. The moat that specialization built is eroding faster than the moat that breadth builds.
2. The economy is rewarding synthesis, not lookup. When information retrieval is free, the scarce skill becomes connecting things across domains that don't obviously talk to each other. That's structurally a generalist skill, not a specialist one.
3. Volatility punishes narrow bets. If AI can restructure or automate a given specialty on a two-year timeline, being narrowly specialized in the wrong thing is a bigger risk than it used to be. Breadth is optionality, and optionality is worth more in a fast-moving environment.
Specialist case, three strongest versions:
1. AI is a force multiplier, and multipliers favor whoever has the most to multiply. A generalist using AI in ten domains gets shallow help in ten domains. A specialist using AI in one domain where they already have judgment gets genuinely superhuman output. The tool amplifies what's already there.
2. Trust and stakes still require a track record. For anything with real consequences, people don't hire the person with broad AI fluency, they hire the person who's been in the room before. AI hasn't changed who gets the call when something matters.
3. Someone has to catch what the model gets wrong. AI is confidently mediocre in unfamiliar territory and genuinely dangerous when errors are subtle. That failure mode only gets caught by someone who already knew the answer before they asked. Depth is what makes the output trustworthy instead of just plausible.
Where I come down:
it's not really generalist versus specialist, it's specialist-with-AI-fluency versus everyone else. The strongest position in this economy is deep domain expertise plus genuine command of the tools, not one or the other. A pure generalist with no anchor domain is producing plausible-sounding synthesis with nobody able to tell if it's right, including them. A pure specialist with no AI fluency is leaving most of the multiplier on the table and will get outpaced by someone with half their depth and real command of the tools. The people actually winning right now aren't choosing a lane, they're the ones who took one real domain of depth and then went and got fluent, not just familiar, with the tools on top of it. Breadth still matters, but as a second layer on a real foundation, not as a replacement for one.
Generalist:
1. AI collapses the cost of depth-on-demand. When any specialist's marginal knowledge is a query away, the scarce skill becomes knowing which questions to ask and how to stitch answers together across domains. That's a generalist move.
2. Specialization ages badly right now. Narrow expertise defined by a stable toolset gets automated fastest, because narrow and stable is exactly what models are good at replacing. Breadth is harder to commoditize because it's harder to even define the job.
3. Most real value sits at the seams, not inside disciplines. [A client project], MyDriveScore, Proven, the Accelerator: none of that is one specialty, it's you translating across insurance, product, and AI capability. That's the actual job description of an advisor.
Specialist:
1. AI is a brilliant executor and a mediocre judge. Someone still has to know when the output is subtly wrong, and that requires earned domain judgment no model has. Generalists can't catch what they don't have the depth to see.
2. Trust still routes to depth. A board or a carrier isn't hiring "AI fluency," they're hiring someone who's spent decades in P&C and can be held accountable for being right. Breadth doesn't carry that kind of credibility.
3. AI makes specialists more leveraged, not less relevant. A deep expert with AI tooling multiplies their existing edge. A generalist with AI tooling is just faster at being shallow across more things.
Where I land:
specialist core, generalist interface. Not a hedge, an architecture. The credibility and judgment that get you in the room are still deep-domain, and that doesn't change. What changes is that you now need a generalist layer on top translating that expertise across contexts, tools, and disciplines, which is functionally what you've built with the Advisor OS. The mistake is treating this as a spectrum where you pick a point. It's a stack. Pure specialists get boxed into the part AI eats fastest. Pure generalists never earn the trust that gets them hired for anything that matters. The Trivium piece you wrote is basically the same argument aimed at capability instead of career strategy.
What the four cells show
The artifact changed stance and delivery, not conclusions. All four cells reached nearly the same place on the merits. What moved was the opener, the hedge, and how soon the answer commits to a verdict. Cell 1 warms up ("Fun one. Let's do it."), hedges with an anti-binary throat-clear, and fishes at the end. Cell 2 opens cold and commits from the first word. Stance and delivery are the part you can be taught to control.
The value scales inversely with the environment. On a blank Claude the lift was large and visible, Cell 1 to Cell 2: the eager opener, the turn-it-back closer, and the hedge all disappeared. Inside the already-tuned environment the lift was incremental, Cell 3 to Cell 4, because that setup had done the gross cleanup already. Both lifts are real, at different sizes. The artifact does its heaviest lifting in a blank Claude, where a new user starts.
The artifact only moves what it targets. The em dash shows this. The artifact never mentions em dashes, and adding it left the count unchanged in both rows. The blank rows run six each, artifact or not. The customized rows run zero, because that environment carries a no-em-dash rule the artifact does not. If you want a specific tell handled everywhere, it has to be named. That is the difference between a preference you hold and a rule the model follows.
blank
blank + artifact
tuned env
tuned + artifact
The tracer reads by row, not by column. The environment drives the em dash to zero, and the artifact leaves it untouched.
Only Cell 4 reached for specific context, naming the operator's own ventures and prior writing unprompted. That is the second half of the lesson. The artifact supplies the stance; the environment supplies the material. Personalization is both, working together.
Where we are heading
From a one-time setup to a living practice
A voice artifact is a strong start. The technique matures along a clear path, and each step maps onto work you will do later in the program.
The artifact is a Boundary. It needs a Loop.
The voice artifact is a Boundary, one of the five Elements, and it lives in Memory. On its own it is static: it holds the rules you set and nothing more. The move that makes it improve is to add a Loop. When Claude does something you dislike, you feed that back into the artifact, adding it to the standing tells-to-avoid list with a concrete before-and-after, so the Boundary gets sharper over time instead of staying fixed. A Boundary without a Loop is a good start that never gets better. Codify it in Memory so it persists across sessions and compounds.
Make it summonable and model-aware
Every model release ships its own personality: different defaults, different tells, different baseline warmth. A voice artifact is therefore calibrated against a specific model, and it drifts when the model changes underneath it. The mature version stamps each artifact with what it was tuned against, and keeps a short voice-drift check you can summon: run one fixed prompt, compare against your known-good baseline, catch the new tells, re-tune. The process is model-agnostic even though each instance is model-specific.
Role-based voices for specialized agents
The furthest step. Once you are orchestrating more than one model, voice stops being a single setting for you and becomes a voice per role. A critic agent should be blunt and hard to please. A client-facing drafter should be warm and polished. A brainstorm partner should be playful and divergent. An analyst should be spare and precise. The endgame is a small library of role-voice profiles, purpose-built and assigned to specialized models by function, so each agent in a workflow sounds the way its job requires.
This guide is the pre-work. In Working in Conversation you will run the protocol yourself, build your own voice artifact, and start your own loop. The delegation and automation modes that follow, in Modules 3 and 4, are where a tuned voice starts to pay off across more of your work. The technique on this page began with one participant taking a small irritation seriously. The same door is open to you.