The Conversation Is the Specification
🎙️ Listen to this article
The Conversation Is the Specification
Your best AI brief isn't a brief at all. It's the unfiltered planning call where you actually understood the problem — the one with the wrong client name, the casual aside about CRM exports, the constraint nobody wrote down.
Here's the evidence.
Why Clean Briefs Fail AI Build Tools
The instinct when briefing an AI build tool is to write a clean document. User stories, acceptance criteria, technical constraints, wireframes. The kind of artefact that looks professional and feels thorough.
But the constraints that matter most are the ones people mention casually, not the ones they write down.
A strategy portal we built for Husqvarna Australia proved this. The project involved 19 CSV files, 360,000 rows of data, 55 meeting transcripts, and a $1.03M media budget across six channels. The kind of scattered mess that normally takes weeks to assemble into a coherent argument.
The best input the AI got wasn't any of the data. It was a 31-minute recording of two people working through the actual problem — downloading files, explaining quirks, surfacing constraints the client hadn't formally stated. That conversation carried three pieces of intelligence that a polished specification would have lost entirely.
Three Things a Requirements Doc Would Have Missed
Columns That Shift Between Exports
Nick, the account lead, explained how his team exports CRM data from the client's system. The export format changes every time.
"Generally, he's told me that the export is somewhat different every single time... the columns, like, might be slightly off."
That casual comment became a hard rule in the build: parse by header name, never by column position. A requirements document might have said "CRM data import" as a feature. The conversation encoded how CRM data actually behaves in the wild — and the AI built the parser accordingly.
The principle: Your team's workarounds and frustrations contain the real technical specification. The planning conversation is where those surface. A clean brief strips them out.
Constraints Nobody Documents
Brand work "basically just stops from March" through August. A Vision IQ product launch in June would spike the budget, but nobody had documented exactly when or by how much. The client hadn't formally communicated either of these.
Yet both became real pages in the portal — a calendar with campaign bars showing exactly when spending shifts, because two people mentioned it in passing during a planning call.
The principle: Every organisation runs on undocumented constraints. They live in the conversations between the people doing the work. If you clean up the brief, you clean these out.
The Gap Nobody Had Quantified
Nick had tried mapping demand data to budget allocation, but the lead-to-revenue model didn't exist.
"We don't really have like a set goal apart from like this Nesret sales, which like at the end of the day, we don't even know what a lead is like worth to them realistically."
The AI built the model from scratch — CRM conversion rates, Shopify order values, budget data — because it heard the gap described as a real frustration, not as a tidy line item on a feature list.
The principle: When someone says "we don't even know," that's not a gap to skip. It's the most valuable feature to build. A requirements doc rarely captures what's missing. A conversation almost always does.
What Happened When the Conversation Was the Brief
The planning call ended. A voice-dictated prompt went to Claude — messy, run-on sentences, and with the client called "SharkNinja" by mistake. (A meeting with SharkNinja's CEO the day before left the name stuck.) Claude identified the actual client from the attached transcript.
Three clarifying questions. Three answers. Out came a 15-file configuration package:
- A CLAUDE.md with the full business context
- Settings with a PreToolUse hook blocking writes to the /data directory
- Six slash commands — including one for the shifting CRM exports and one for revenue forecasting
- Review agents for data accuracy and brand compliance
- Skills for strategic context
The voice prompt got the client's name wrong, and it didn't matter. The transcript carried the real context. No user stories. No wireframes. The conversation was the specification.
The Build as Evidence
The portal went from planning call to feature-complete in under twelve hours. Not because the AI was fast — because the brief was right.
The first commit landed at 00:35 on February 20: 350,632 insertions across 69 files. 359,588 rows of CSV data organised into structured parsers. Four of six portal pages wired to real data. The CRM parser used header-based column mapping — Nick's warning, honoured automatically.
Then a midnight discovery: the entire session had been running on the wrong model. Sonnet 4.5, not the Opus that was supposed to be running. Rather than kill the session and lose everything, the model wrote a 607-line handoff document briefing its successor. An improvised relay that included warnings like "This is NOT a lifecycle-first strategy — it's paid media enhanced by lifecycle" and "Parse CRM CSV by header name, NEVER by column index. Columns shift between exports!" — constraints that trace directly back to the planning call.
Opus 4.6 picked up forty-nine minutes later. By morning: interactive components, a Gantt-style campaign calendar rendering the undocumented constraints as visual timelines, a creative intelligence page, and a 543-line refactor that transformed the portal from spreadsheet replacement to decision tool.
Four commits. Midnight to morning. But the timeline is not the point. The point is that every page, every parser, every constraint in the build traces back to something someone said casually on a 31-minute call — not to something someone wrote in a requirements document.
Six Pages, One Argument
The portal makes a single strategic case across six pages.
Residential Strategy is the centrepiece — a Growth Model showing the integrated approach delivering 143 qualified leads above target where paid media alone falls 221 short. Audience Intelligence processes 4,172 CRM leads with customer voice themes surfaced from real enquiries. Investment Plan breaks the $1.03M budget by channel, audience, and month. Calendar & Milestones maps the full 2026 timeline. Creative Intelligence and Professional Robotics round it out.
Every metric traces to a source CSV. Every page has a data source footnote. The Growth Model makes the case for the retainer — lifecycle nurture lifting qualification from 28.6% to 38%.
The client couldn't assemble this picture before. Not because the data didn't exist — 360,000 rows of it existed. Because the argument connecting those rows lived in conversations, not in documents.
What This Means for How You Brief AI
Three things worth stealing from this:
Record the planning call. Not for transcription's sake — because the casual asides, the workarounds, the frustrations, the half-stated constraints are the actual specification. The AI needs to hear what you heard.
Stop cleaning up the brief. The messy voice prompt with the wrong client name produced a better result than a polished requirements document would have. The mess carries signal. The cleanup strips it.
Let the AI hear the problem, not the solution. Nick didn't say "build a header-based CRM parser." He said the columns shift every time. The AI figured out the right solution because it understood the real problem — in his words, not in a translated specification.
The conversation where you understand the problem is the best brief you'll ever write. Stop translating it into something cleaner. That translation is where the most valuable information gets lost.
