The problem was not a shortage of ideas
The founder had years of stories, strong opinions, and a distinct way of explaining the work. None of that appeared reliably when they sat down to "make content." A blank document produced safe summaries. Generic AI prompts produced polished language that could have belonged to anyone.
The useful source material showed up in conversation: the small physical details, the corrections made halfway through a sentence, the phrases repeated without noticing, and the stories that connected a business decision to a deeply held belief.
Why the interview was a voice conversation
The interview agent was designed as a patient, resumable conversation rather than a form. The founder could speak naturally, stop after one session, and return for the next part without rebuilding the context.
The questions moved from origin stories into decisions, standards, difficult moments, customer beliefs, language, and future vision. Follow-up questions asked for scenes and specifics instead of accepting an abstract answer.
The completed source set contained two interview sessions and roughly 7,500 words. That number matters less than the quality of the material: it captured the founder thinking aloud rather than performing finished marketing copy.
Evidence artifact
From an answer to reusable source material
An anonymized reconstruction based on the real interview structure. Details have been changed to protect the founder while preserving the mechanism.
| Interview layer | Anonymized example | What the system keeps |
|---|---|---|
| Story | “The first version worked, but I knew immediately what I would never compromise on again.” | A concrete origin scene and a repeatable standard |
| Language | “The part people cannot see is usually the part I care about most.” | A signature phrase and a point of view |
| Decision | The founder chose the slower option because it protected the customer experience. | A decision rule, not just a chronology |
| Rhythm | The founder softened the opening, corrected themself, then landed on one short image. | Sentence shape that makes future drafts sound recognizable |
The thinking and voice rapport
The transcript was not treated as a bag of quotes. It was read as evidence for a detailed rapport describing how this founder tells stories and makes meaning. The rapport included:
- recurring stories and the situations that call them forward;
- decision patterns and standards the founder applies automatically;
- signature language, metaphors, and sensory details;
- tensions and contradictions that make the point of view interesting;
- sentence rhythm, hedges, repetition, self-correction, and quiet endings;
- content themes that could support more than one isolated post;
- boundaries around personal material, factual claims, and language that does not sound like the founder.
The rapport is descriptive, not diagnostic. It does not claim to measure a person or reduce their thinking to a score. Its job is to make the source material usable and reviewable.
Turning the rapport into a working skill
The final system does more than store the interview. It gives an AI collaborator specific instructions for writing with the founder's source material:
- Load the approved voice guide and interview excerpts.
- Pull real phrasing and story beats before drafting.
- Match the requested format without forcing every idea into the same hook.
- Check claims against approved factual references.
- Remove writing habits the founder would never use.
- Return a draft for the founder to review, change, or reject.
Format-specific prompts handle short-form video scripts and social captions. The same source can also support newsletters, an About page, or founder-led sales copy without inventing a new voice for each channel.
Evidence artifact
The voice check
A genericized before-and-after example showing the kind of correction encoded in the real voice skill.
Generic AI draft
Building a business taught me that perfection is the enemy of progress. Every version brings you closer to your vision, so keep going.
Source-grounded draft
I think the first version mostly taught me what was missing. I could feel it before I could explain it. So I made another one. That was the work: paying attention closely enough to know what had to change.
The second version keeps the founder's softer entry, self-observation, concrete sequence, and quiet ending. It does not manufacture a motivational lesson the founder never gave.
What remains human
The agent can make source material easier to retrieve and reuse. It cannot decide which personal story should become public, whether a vulnerable detail still feels safe to share, or whether a draft represents the founder today.
Raw interviews remain private. Public examples are anonymized reconstructions, not lightly redacted screenshots. Personal details, company identifiers, call IDs, and distinctive quotations are removed unless the speaker explicitly approves them. Every generated piece remains a draft until the founder chooses to publish it.
That boundary is what makes the system useful: the AI carries the memory and the consistency. The founder keeps authorship.