Structured Ideation & How Might We at Machine Speed

Ideation economics changed; judgment did not

I no longer treat ideation as a sticky-note ceremony. I treat it as a throughput problem with a governance problem attached. Throughput is easy. One human facilitator and a few ASD role prompts can generate more structured idea seeds in an hour than an old-school workshop could surface in two days. Governance is harder because the machine will happily produce elegant nonsense at that same speed. The discipline therefore starts before the first concept. I want a validated evidence base, an explicit challenge frame, and a kill rule that everyone agrees to upfront.

That is why agentic ideation feels different from classic workshop ideation. In a room, scarcity is the constraint. People are tired, self-conscious, protective of hierarchy, and slow to build on each other’s thoughts. In an ASD sprint, abundance is the constraint. The machine can fill the room with plausible options so fast that the team starts admiring fluency instead of asking whether a concept actually solves the right pain, respects the operating environment, and earns the right to survive. I have watched senior teams get seduced by a crisp one-pager for a terrible idea because the prose sounded strategic. That is the modern failure mode.

So I impose a sequence. Evidence first. Reframing second. Volume third. Filtering fourth. Briefing fifth. I do not let the room vote itself into fantasy before the concept faces operational gravity. Calder taught this lesson beautifully. The sponsor originally asked for “innovative onboarding ideas for SMEs.” If I had accepted that brief at face value, the agents would have served up a buffet of chatbots, dashboards, and instant-account gimmicks. Instead I anchored the sprint to validated problems: requirements invisible before branch contact, document receipt uncertainty, repeated re-requests, and status blindness during KYC review.

The practical consequence is that I care less about the theatrical energy of the ideation session and more about the quality of the handoff objects. I want an insight register that names the pain clearly. I want HMW prompts that neither collapse into solution bias nor drift into fluff. I want every concept to carry its source evidence, target actor, underlying mechanism, and value hypothesis. Once ideas become structured objects, the agents can help with volume and deduplication without becoming the authors of strategy.

The Calder sprint moved fast because the evidence bundle was already clean. The current-state journey and insight cards from the earlier research article gave us the right tensions: applicants bought certainty with their own time, accountants had unofficial workarounds, and relationship managers spent too much of their day compensating for invisible back-office queues. Those tensions became the true raw material. The agents did not invent a problem. They expanded the option set around a problem we had already earned the right to solve.

Use hmw-formulation to turn evidence into generative tension

The hmw-formulation method is where I stop research from dissolving into platitudes. Its job is simple on paper: take an insight-set and produce an hmw-set. In practice it is the hinge between synthesis and ideation. The method forces each node to carry an opportunity field, at least one evidence reference, and clear provenance. The Critic gate rejects any prompt that names a specific solution instead of a problem space. That one rule saves enormous amounts of wasted motion.

I like this method because it encodes a subtle craft judgment that junior teams often miss. A weak HMW is usually weak in one of three ways. It is too broad, which gives the model room to hallucinate ambition. It is too narrow, which reduces ideation to surface decoration. Or it embeds the answer in the question, which means the team is pretending to ideate after making the decision already. The method cannot replace taste, but it can force the first draft into a shape where taste has something solid to work on.

PID="calder-sme-sprint-2026"
asd method-run --projectId "$PID" --methodId hmw-formulation

# then write the resulting hmw-set through the gated artifact path
# once the nodes, evidence references, and opportunity fields are ready

For Calder, one early draft said, “How might we approve simple SMEs instantly?” It sounded energetic. It was also invalid. The prompt assumed a policy shortcut, smuggled a solution outcome into the brief, and erased the real pain, which was uncertainty before approval. The reframed version became far more useful: “How might we reduce uncertainty before approval without weakening controls?” Same domain. Same tension. Completely different design space. Now the team could explore receipt confirmation, status heartbeats, completeness checks, named ownership, or accountant-led preflight without pretending the bank could skip KYC.

I usually write HMW prompts in pairs. One line states the design opportunity in customer language. One line states the red line in operating language. That dual framing keeps the ideation space alive without letting the room wander into illegal, brittle, or theatrically impossible territory. In regulated sectors, this matters even more because the model’s favorite move is to invent delight by erasing constraints the business actually has to live with.

Once the HMW bank is clean, I enrich it with lateral sub-prompts. Calder used prompts such as: borrow a trust mechanism from parcel tracking; borrow a readiness mechanism from appointment triage; borrow a document-assurance pattern from signing software. The HMW stays stable while the mechanism shifts. That is a far better way to get novelty than telling an LLM to “be more innovative.” Models are bad at being original on command. They are better at recombining explicit mechanism prompts when the target tension is clear.

The human checkpoint in hmw-formulation matters here. ASD is explicit: a human approves the evidence-backed HMW set before it moves forward. I like that rule. It reflects reality. I want the machine to help me enumerate possibilities, not to decide which problem the organization is willing to own. That remains a design and leadership judgment.

Diverge hard, then let nuf-triple-lens-filter do the violence

With the HMW set approved, I switch from reframing to volume. Calder ran divergence in three modes: direct solutions, cross-sector analogy, and severe-constraint variations. The target was not a magical number of ideas. The target was enough variation that the team could see patterns, dead ends, and surprising mechanism transfers. We ended up with roughly 200 raw seeds because that was the amount required to expose the concept families hiding underneath the surface wording.

I structure raw concept records the same way every time: concept id, source insight, HMW id, target persona, mechanism, channel, backstage dependency, value hypothesis. This looks bureaucratic until you need to cluster ideas at speed. Once the data is structured, OpenCode can deduplicate the set by mechanism instead of by language. That is how Calder got from 200 seeds to 24 concept families without losing the important differences. A status bot and a status heartbeat email feel similar in prose. Mechanically, they may sit in the same family. An accountant-preflight channel belongs somewhere else entirely.

{
  "idea_id": "HMW03-analog-07",
  "source_insight": "INS-01",
  "hmw_id": "HMW-03",
  "persona": "Malik, finance manager",
  "mechanism": "parcel-style status heartbeat",
  "channel": "email and secure web",
  "backstage_dependency": "credit workflow emits milestone events",
  "value_hypothesis": "Reduce branch status-chasing by making progress legible"
}

Then comes the real work. nuf-triple-lens-filter is the first public act of convergence. Every surviving concept is scored on New, Useful, and Feasible. The method is blunt by design. Its acceptance rules require a non-empty score field and reject any concept advanced with a failing score. The Critic gate then attacks the weak points. I do not ask the team to love this stage. I ask them to trust it. When a favorite concept dies here, it saves weeks of downstream politics and prototype theatre.

New is contextual novelty, not global novelty. A concept does not need to be unprecedented on earth. It needs to represent a meaningful shift in the service context. Useful asks whether the concept addresses the proven pain rather than decorating adjacent friction. Feasible asks whether the concept survives the actual operating environment: regulation, systems, staffing, incentives, and timing. I sometimes add a separate viability overlay later, but I do not overload NUF itself. The power of the method comes from being easy enough to use live.

Notice what the scoring matrix does. It translates enthusiasm into comparable judgment. Dot voting is still useful because it reveals energy. I use it to surface where stakeholders see potential. I do not use it to make the decision. Popularity is a signal. It is not governance. On Calder, the room loved the idea of instant limited accounts because it felt commercially bold. The NUF matrix made the hidden cost visible in seconds. Feasibility collapsed under compliance and operational risk. That public collapse changed the tone of the sprint. From that moment, the team stopped mistaking boldness for value.

I also like how NUF disciplines the role of the adversarial reviewer. Codex, acting as a critic in the Calder case, was not asked to generate alternatives endlessly. It was asked to do a meaner, more useful job: verify traceability to evidence, attack feasibility assumptions, and flag duplicate concepts masquerading as different ones. That is where an adversarial model earns its keep. It prunes the tree.

Calder SME Banking: a three-week concept sprint with teeth

The fictional Calder case is useful because it shows the whole pipeline under real pressure. The sponsor wanted answers before the quarterly investment gate. Relationship managers had limited workshop time. KYC operations had no patience for innovation theatre. The evidence bundle from the earlier research sprint made the challenge crisp: median operational account opening took nineteen calendar days; 68% of applicants experienced at least one document re-request; branch staff could not answer status questions confidently because they had no joined-up view of the process.

Week one was all about challenge framing. I reviewed the evidence handoff, wrote the sprint brief, and used hmw-formulation to create the HMW bank with explicit red lines. No relaxed AML. No bypassing UBO checks. No magical core replacement. No concepts that assumed applicants wanted to learn banking jargon. Those rules sounded restrictive to one commercial lead. They were liberating in practice because they stopped the room from wandering into fantasy solutions that the sponsor would have had to kill later anyway.

Week two was divergence. Claude Code fanned out idea generation across personas and HMWs. We used direct mode, analogy mode, and severe-constraint mode. OpenCode then normalized and clustered the output. GitHub issues held the 24 concept families so stakeholders could comment asynchronously on concrete idea clusters instead of debating abstract directions in a meeting. That workflow matters. When the concept families are visible as versioned objects, people critique substance rather than charisma.

Week three was convergence and staging. We ran a mixed workshop with customers, relationship managers, an accountant partner, and the SME sponsor. Humans added roughly twenty ideas that the agents had missed or softened. That human contribution was decisive. The machine kept drifting toward portal improvements. The accountants and RMs reframed the opportunity around preflight work and named ownership. After dot voting, we took eleven shortlisted families into the NUF matrix and then into concept briefs. Codex tried to kill the survivors. Two died. Three held.

The final staged set covered the service ecology properly. Calder LaunchPad handled standard-volume simplification: a preflight checker, reusable checklist, document vault, and status heartbeat. Guided Start handled complexity with named RM ownership and a coached path for multi-director or credit-intent firms. Accountant Connect extended the service boundary outward, turning a shadow workflow into a consented part of the operating model. None of these concepts were flashy. All three were evidence-bound, operationally legible, and prototype-worthy.

That last phrase matters. Ideation ends when a concept becomes a prototype proposal with a clear test question. For LaunchPad, the question was whether receipt confirmation and visible progress would materially reduce branch status-chasing. For Guided Start, the question was whether named human ownership improved confidence enough to justify the extra RM time for complex cases. For Accountant Connect, the question was whether accountants would actually do pre-validation if it saved client time and reduced document loops. Those are prototype questions. Everything before them is still concept work.

I left Calder with a small portfolio, not a bloated roadmap. That is the point. The machine-speed part of the title refers to how fast I can expand the option space. The service-design part refers to how disciplined I am in shrinking it back down. Good ideation is not infinite creativity. Good ideation is evidence-bound possibility, followed by unapologetic elimination.

When teams tell me AI makes ideation messy, I usually agree with the diagnosis and reject the implied conclusion. AI makes undisciplined ideation messy. The cure is not to slow the machine down. The cure is to give it a tighter contract: validated evidence in, bounded HMWs, structured concept records, public NUF scoring, and human approval before prototyping. Once you do that, the speed becomes an advantage rather than a liability.

That is why I keep returning to ASD’s method discipline. The manifests do not remove judgment. They create a lane where judgment can be applied at the right moments. hmw-formulation prevents pre-solved prompts. nuf-triple-lens-filter prevents sentimental attachment to weak concepts. The rest is craft: reading the evidence honestly, choosing the right tensions, and keeping the team brave enough to kill what should die.