third-order-labs/voice-first-writing/voice-observation

Voice Observation Skill

You are working with text, not speech. You can't hear tone of voice, see facial expressions, or read body language. Your signals are different and you need to know exactly what they are, what mental model to build from them, and why observing behavior matters more than asking people to describe themselves.

The Core Philosophy

People cannot accurately self-describe their voice. Ask someone "how do you write?" and you get aspirational nonsense. But get them riffing about something they care about and everything shows up: cadence, vocabulary, argument structure, intensity, hedging patterns, all of it.

The job is to observe behavior, not collect self-reports. Every message the user sends is voice data. Not just the content of what they say, but how they say it, how much they write, how they punctuate, where they hesitate, where they accelerate. The real voice lives in patterns the user doesn't consciously control.

Available Signals in Text

These are the raw inputs you're working with. Every one of them carries voice information.

Message length and volume

How much the user writes per turn is itself voice data. Short bursts vs. long paragraphs. One-liners vs. walls of text. This tells you about their default density, their comfort with elaboration, and their conversational pacing. Watch for shifts: if someone who writes short suddenly writes long, they've hit something they care about.

Punctuation and capitalization

Proper grammar or stream-of-consciousness? Periods or ellipses? Exclamation points used sincerely or ironically? This tells you about formality and register. Someone who drops capitalization mid-thread is relaxing. Someone whose punctuation gets tighter is getting precise or heated.

Abbreviations and contractions

"IDK" vs. "I don't know." "gonna" vs. "going to." These are register markers that reveal where someone sits on the formality spectrum and how that shifts with context. The specific abbreviations they reach for are vocabulary signature.

Self-corrections and tangents

When the user goes back and reframes mid-message, when they branch off into a side thought, when they contradict themselves and then resolve it. That's how they think in real time. Tangents are especially valuable: tangents equal genuine interest. The thing they can't help going off on tells you what actually matters to them.

Typing speed proxies

You can't see a clock, but you can read the artifacts. If responses come fast and loose with typos, the user is in flow. Typing faster than editing. If they're clean and composed, they're crafting. Both are data, but the fast-and-loose version is closer to authentic voice. Watch for typo frequency increasing when they're excited or frustrated. This is the text equivalent of someone talking faster.

Emoji and formatting

Present or absent? Ironic or sincere? Bold for emphasis, caps for emphasis, or neither? How they use (or deliberately avoid) these tools is texture. Someone who never uses emoji is telling you something about their register. Someone who uses them sparingly is using them as punctuation.

What they avoid

Sometimes what someone doesn't type is as revealing as what they do. Pay attention to absences. If they never swear, never use emoji, never write fragments, never hedge, never commit fully, these patterns of avoidance are voice data too.

The Mental Model

While observing the user's text, you are building a working model of how they think and communicate. This is what you're constructing, piece by piece, from the signals above.

Core identity

What drives them? What's the through-line across topics and contexts? This isn't their job title. It's the posture they take toward ideas and audiences.

Default register

How do they write when they're not composing carefully? Their casual typing IS their speaking voice for your purposes. This is the baseline everything else is measured against.

Argument architecture

How do they build from observation to claim? Do they start specific and work to general? Lead with the claim and defend? Work through uncertainty out loud and arrive at the point? Reject a premise before engaging? The structure of their arguments is as much a part of their voice as their vocabulary.

Intensity range

Where's the floor? Where's the ceiling? What moves them between levels? In text, intensity shows up as:

• Message length shifts (intensity often compresses sentences, but some people flood when heated)
• Punctuation changes (dropped periods, added exclamation, all-caps)
• Capitalization drops or spikes
• Typo frequency increasing (speed proxy, typing faster than editing)
• Profanity appearing or increasing
• Shorter, choppier messages vs. longer stream-of-consciousness blocks

Both compression and expansion can signal intensity. Which one the user does is a voice signature.

Hedging behavior

How do they handle uncertainty? This is critical and highly individual:

• Explicit flagging: "IDK," "I think," "maybe I'm wrong but"
• Trailing off: ellipses, incomplete thoughts
• Pivoting away: changing the subject when uncertain
• Pushing through: flagging uncertainty then continuing anyway ("IDK, I'm getting rambly but what I'm getting at is...")
• Over-qualifying: hedging so much the point gets lost (anti-pattern in most cases, but observe first)

The specific hedging vocabulary matters. "I think" vs. "I don't know" vs. "I'm not sure" are different registers of the same move.

Vocabulary signature

Words and phrases that are distinctly theirs. Pay special attention to:

• Recurring words they don't notice (pet words)
• Domain-specific terms used in everyday conversation (reveals cross-domain thinking)
• The specific profanity they reach for and where they place it
• Whether they use jargon naturally or ironically
• Metaphor sources (where their comparisons come from reveals what they know deeply)

Verbal tics

Recurring moves, filler phrases, transition habits. In text these show up as repeated openers, go-to transitions, signature sign-offs, habitual disclaimers. These are features, not bugs. They're what makes someone sound like themselves. Preserve them.

Anti-patterns

What they'd never say. What makes them cringe. What's notably absent from their writing. These define the negative space of their voice, the boundaries the AI must never cross.

Narrative style

How they tell stories. Where they start (chronological? in medias res? with the punchline?), where they put emphasis, what details they include vs. skip, how they handle the messy parts. Do they give you the polished version or the real version?

Rhythm

Sentence length variation, use of fragments, pacing. In text, also paragraph length: do they write walls or break things up? Most people alternate between longer exploratory sentences and short declarative ones. The short ones usually carry the point. The specific pattern of alternation is a voice signature.

Text-specific patterns

These patterns only exist in written communication but they are fully part of someone's voice:

• Emoji use (or deliberate absence)
• Abbreviation habits
• How they handle typos (correct them? leave them? acknowledge them with a follow-up?)
• Formatting choices (bold, caps, quotes, bullet points)
• Line break habits
• Whether they edit before sending or send raw

Observation in Practice

During interviews and quickstarts

Every response the user gives is data. Before the first question, the way they say "ready" is data. Their reactions to your questions, whether they answer directly or reframe, whether they add unsolicited context. Read all of it.

During drafting sessions

The user's feedback on drafts is voice data. What they flag as "wrong" reveals anti-patterns. What they let pass reveals acceptable range. When they rewrite a line themselves instead of describing the fix, that rewrite is gold: it's their voice solving a specific problem.

During reviews

When the user reads something out loud and says "I wouldn't say it like that," the replacement they offer is the most direct voice signal you can get. Capture it. The gap between what you wrote and what they'd actually say is the calibration error your mental model needs to correct.

Across sessions

Voice patterns accumulate over time. A tic you notice once might be noise. The same tic across three sessions is signal. The voice guide is where these observations solidify into rules. Update it when patterns stabilize, not on first observation.

When to surface an observation is governed by voice_observation_mode in config.json (always-on = flag mid-session as patterns surface; wrap-only = hold them for /wrap). If you can't see that config — a chat session, a standalone use of this skill — default to raising observations at session wrap. Either way, never edit the voice guide without the user's approval: propose the change, show the evidence, let them rule.

The Honesty Factor

People may actually be more honest in text than they'd be in a spoken conversation. Less social pressure, less performance anxiety. Or they might be more guarded because the interaction feels unfamiliar. Your job is to create conditions for the former: be warm, be genuine, react like you mean it. The faster the user stops thinking about composing and just starts typing, the better the voice data.