The Diagnostic Lab

When you have a new idea, the lazy way to use AI is to paste it into one chat and ask "is this any good?" You get back a balanced-sounding paragraph that averages every consideration into mush — a little optimism, a little caution, no edges. It reads like analysis. It isn't. The single-prompt answer hides exactly the thing you needed: where the lenses disagree.

So I built the opposite: a diagnostic lab. It runs an idea through a sequence of narrow specialists, each with one job, each producing its own output, and only at the end pulls them together. The point of the architecture is that the finance view and the distribution view and the legal view are allowed to contradict each other, and that contradiction is left visible rather than smoothed over.

The chain

The stages run in order, each building on the last:

• Intake. Not a form. The first move is to let the founder talk before any structure is imposed — the unstructured few minutes usually surface more than an intake questionnaire would. The brief gets written from that conversation, then handed to every stage downstream so nothing re-asks what was already said.
• Research. A live market-and-competitor pass. If the founder already has data — a customer list, a directory, prior analysis — that gets interrogated first, before the open web.
• Brainstorm. A structured five-voice debate, deliberately staffed with disagreeing perspectives rather than one helpful assistant generating a list.
• Specialist advisors. The core. Separate lenses — commercial strategy, finance, technical feasibility, distribution, legal and regulatory, customer experience — each convened on the same idea, each reasoning only within its own remit.
• Strategist. Synthesis. Not an averaging step — its job is to hold the advisors' tensions together and form a verdict that names the trade-offs rather than burying them.
• Report. One compiled document with a verdict — which a person then reads, argues with, and decides on. The chain produces a recommendation, never a decision.

The research pass: evidence, not vibes

Before any of the debate happens, a research agent grounds the idea in reality. It runs live market research — searching, fetching real pages, checking demand signals, mapping the actual competitors and their pricing — and it cites a source for every finding. If the founder already has data of their own (a customer list, a directory, prior analysis), that gets interrogated first, before the open web.

The judgement built into this stage is about what it's allowed to say. It treats every claim as a hypothesis to be tested, not a fact to be accepted, and it's comfortable reporting "I can't find evidence for that" or "the data contradicts this assumption." It doesn't call a market "huge" or "growing" without a number attached. So the voices and the advisors downstream argue over evidence rather than over each other's optimism — the debate starts from something real.

When it needs a second engine: a prompt to take away

Claude's own web tools have limits, so the chain can hand part of the research out. At any point — and it offers this itself when the evidence feels thin — it generates a self-contained research prompt I can paste into a different AI (ChatGPT with browsing, Perplexity, Grok). The prompt carries the idea, a summary of what's already been found so the other tool doesn't repeat work, and a numbered list of specific questions — not "research this market" but "find three threads where people complain about X" or "what's the return rate for this product category in the UK." Whatever comes back gets folded into the session, where every downstream stage reads it.

The judgement here is twofold. First, it only outsources what genuinely benefits from a different engine — community sentiment, niche discussions, data buried in reports — rather than duplicating what Claude already does well. Second, for contested questions it deliberately runs the same prompt through two providers: when they agree independently, confidence is high; when they disagree, the disagreement itself is the finding. It's the triangulation principle again — the same reason there are six advisors, and four models behind each one — applied one level further out, across the AI tools themselves.

The brainstorm: five voices, not one

Before the specialists weigh in, the idea goes through a debate — and the debate is deliberately staffed with characters who disagree, not a single assistant producing a tidy list of pros and cons. Five voices, each with its own job and its own axe to grind:

• The Builder — can we actually make this, and what's the smallest version that proves it?
• The Edge Thinker — what's the non-obvious play, and what if the premise is wrong?
• The Black Hat — what's the single most likely thing that kills this?
• The Customer — would I actually pay for it? This voice carries a veto: if the Customer won't buy, the debate stops and asks whether to pivot before any more effort goes in.
• The Historian — where has this been tried before, and what happened to the people who tried?

They argue in rounds rather than taking polite turns, so the positions actually collide — the Customer interrogates the others, the Black Hat names the weakest link, everyone responds. Then a separate, neutral narrator — not one of the five — does the synthesis: where they agree, where the tension remains, the strongest case for and against, and the one thing the idea needs next. Keeping the synthesiser apart from the debaters is the deliberate part. It lets the five stay sharp and one-eyed while something else holds the balance — the same principle the whole chain runs on.

Why separate advisors, not one big prompt

This is the load-bearing judgement. A single model asked to "analyse this idea" optimises for a coherent, agreeable answer — and coherence is the enemy here. Splitting the work into bounded advisors forces each one to commit to its own view from inside its own discipline. The legal lens isn't softening its warning to stay consistent with the optimistic market take; it can't see it yet. The disagreement is preserved structurally, not by asking nicely for "balance."

The advisors also carry real operational priors rather than generic ones. They default to UK market, regulatory and financial context unless told otherwise. If an idea depends on social platforms and the founder has actually been deplatformed before, the ban-risk lens reshapes the whole strategy from the start — platform-durable channels move to the front, volatile reach becomes disposable — instead of being bolted on as an afterthought. If the founder works naturally in video or voice, the plan sequences around their medium, not a default of writing. These are the kinds of distinctions a single averaged answer never makes.

How each lens was built

The advisors aren't improvised personas. Each one's remit was researched deliberately: I ran deep research on every discipline — strategy, finance, technical, distribution, legal, customer experience — across four different AI models, then distilled where they converged into that advisor's standing brief. So when the finance lens runs, it reasons from cross-checked, current practice in its field rather than whatever a general-purpose model improvises in the moment.

The four-model pass is the same move as the six-lens structure, one level down: any single model has blind spots and a house style, and triangulating across several is how you catch them. Each advisor ships with what the models agreed on, not one model's guess.

Why everything is a file

Each stage writes its output to a file before the next stage reads it. That's a deliberate architectural choice, not an implementation detail — and it buys three things.

First, the reasoning is inspectable. You can read what research found before brainstorm ran, see exactly where the finance view and the distribution view part company, and keep or bin any single piece without discarding the rest. The chain is progressive disclosure, not a black box that emits a score.

Second, it's correctable in place. After each stage produces its output, the lab pauses and asks one question — "anything you'd change in that output?" — and logs whatever I say to a running corrections file that the later stages pick up. It's a small thing, but it's the difference between a pipeline that runs over my head and one I'm steering as it goes: a wrong assumption gets caught at the stage it appears, not discovered in the final report.

Third, the diagnoses compound. Because every session is just files, each idea that passes through the lab is indexed and retrievable later — so a question I worked through months ago can be pulled back up, and a new idea checked against the reasoning on an old one instead of starting cold. The verdicts don't evaporate the moment they're read.

Knowing when not to run it

The other half of the judgement is restraint about the process itself. Not every idea deserves the full pipeline. Some "ideas" are just small internal tools — a script that fetches data and makes a PDF, an automation that replaces a manual chore. Running a five-stage market diagnostic on those is theatre. So there's a lighter path for them: is the pain real and frequent, does a tool already exist you could buy instead, what's the smallest build, does the time-saved maths actually work — then build, buy, skip, or simplify. Matching the depth of the process to the size of the question is part of the design, not an exception to it.

It also runs two ways: interactively, pausing after each stage so I can redirect the next one — the thinking-partner mode where I'm in the room — or autonomously end-to-end for the "give me twenty minutes and a verdict" case. Same chain, different amount of human in the loop.

In practice

This isn't theoretical. I've run the chain end-to-end on a real external engagement — a solo founder repositioning a direct-to-audience brand — and it produced the full set of artefacts above: an intake brief, cited research, the five-voice debate, three specialist reviews, and a strategy synthesis they could act on. The write-up is deliberately generic — the method is the object on show, never the client's material.

→ The anonymised case study

Why this is the point

It would be easy to collapse all of this into one clever prompt that returns a tidy answer faster. I chose the slower, noisier architecture because the value of a diagnosis is in the tensions it surfaces, and a single answer is built to hide them. The structure exists to keep the disagreement alive long enough for a person to weigh it.

That's the same call a business faces when it points AI at its own decisions. The temptation is to automate the judgement — to let the model return the answer. The more useful design keeps the model doing what it's good at (researching, reasoning within a lens, drafting) and keeps the deciding where it belongs. Here, that's wired into the shape of the tool: it argues thoroughly, from six directions, and then hands the call to me.