Deep Design Research and Competitive Teardowns

Module 1 established the research packet as the unit of agentic research: the question, the sources, the method, the findings, and the evidence trail in one auditable structure. This module fills two of the most common packets a design team builds — deep desk research and competitive teardowns — and the failure mode they share: breadth without judgment. An agent will happily fetch forty sources or walk eight competitor products. Whether that produces insight or a museum of screenshots depends on decisions the human makes before and after the run.

Be explicit about the two workflows that anchor the module, because they are different jobs that often get conflated. Deep research reads what has been written about a space — articles, documentation, reviews, standards — and synthesises it with citations. A teardown walks the products themselves: signs up, clicks through onboarding, screenshots the pricing page, counts the steps. Teams get into trouble when they substitute one for the other.

The through-line for the whole module is the title's claim: coverage is cheap, judgment about it is not. The agent supplies coverage. The researcher supplies the question, the rubric, the reading of the evidence, and the call about what it means. Every slide that follows is a technique for keeping those two contributions from blurring into each other.

This slide carries the scoping objective for the module. The deep research workflow's quality is set before the run starts, in the question. A good research question for agentic work has four parts: the question itself with named products and a named moment in the experience, the decision it informs, what counts as evidence, and what would change the team's mind. That last part is the honesty test — if no possible finding could change the decision, the run is confirmation-seeking, and the agent will happily oblige.

The angles are the second scoping decision. Deep research runs fan out across angles — typically one per named competitor or platform, one for the primary standard or guideline fetched from its source, one for published criticism and counter-examples, and one for recency. Five to eight angles is the practical range: fewer than four and the cross-checking has nothing to push against; more than ten and the review time goes to duplication. The point of naming them yourself is that the angles are where research breadth comes from, and breadth chosen by the agent reflects what is easy to find, not what the decision needs.

Connect this back to the packet from Module 1: the question, the decision, and the evidence standard are the first fields of the packet, written before any source is fetched. If they are vague, everything downstream inherits the vagueness — and no amount of cross-checking repairs a question that did not need answering.

This comparison sets up the rest of the module, so make the distinction concrete. Deep desk research is a literature search at agent scale: the question fans out into angles, each angle searches and fetches sources, claims are cross-checked against each other, and what survives becomes a cited report. It is the right tool for standards questions, pattern research, and anything where the answer already exists in public writing. A teardown is the other kind: it walks the live products, captures ordered onboarding screenshots, full pricing pages, verbatim positioning claims, and core flows, and treats marketing copy as a claim to verify rather than a fact to record.

The trouble comes from substitution. A team that runs only the literature search inherits the category's self-description — what gets written about products is shaped by marketing, launch coverage, and whoever bothered to publish a teardown. A team that runs only walkthroughs misses the standards, the published evidence, and the criticism that would have warned them off a pattern. The two pair well in practice: the desk research frames what to look for, the teardown checks what is actually there.

The time figures are honest as of June 2026 and worth keeping on the slide: a deep research question is twenty to forty-five minutes of mostly unattended run time; a teardown is two to four hours including the human review, because creating trial accounts and verifying evidence does not compress.

The dimensions file is the rubric the whole teardown is normalised against, and writing it is design work, not admin. The dimensions come from the design decision: an onboarding redesign needs ordered screenshots and step counts; pricing work needs the page, the tiers, and what each tier gates; a positioning refresh needs the verbatim claims and whether the product backs them. The discipline is that every competitor folder must contain evidence for every dimension, or an explicit note explaining why it could not be collected — paywalled, sales-gated, region-locked. That note is itself a finding about the competitor's go-to-market, not a gap to fill by guessing.

Empty states deserve a specific mention because they are where category research pays off most directly and where memory is least reliable. What a product shows before any data exists — a guided checklist, sample data, a template picker, an illustration with one call to action — is exactly the kind of detail that nobody remembers accurately and that a dated screenshot settles in seconds.

The last bullet is the cross-check rule that does the most work downstream. Positioning claims get recorded verbatim with their URLs and labelled as claims; the walkthrough evidence then marks each one supported, contradicted, or unverified. That single habit is what stops marketing copy laundering itself into the comparison matrix as observed fact.

Walk the pipeline left to right and name the owner of each stage. Defining the set is human: the design decision, the five to eight competitors, the dimensions file, and the access and rights rules. Capture is agent-run, one collection agent per competitor, each filling its evidence folder against the same dimensions and writing a dated manifest of every file. Normalisation is where the rubric earns its keep: every finding cites an evidence file, marketing claims are checked against what the walkthrough showed, and cells the evidence does not cover stay empty rather than getting filled. Synthesis assembles the comparison matrix, the named patterns, and the opportunity gaps. The review is human again: sample the cells against their evidence, separate observation from inference, and decide which gaps are opportunities and which exist for good reasons.

The one-agent-per-competitor structure matters more than it looks. Run by hand, teardowns produce detailed notes on the first three competitors and a shrug for the rest, because attention decays. The fan-out gives the eighth competitor the same systematic treatment as the first, which is the entire point of doing this with agents.

Point at the dashed return line. Teardown evidence decays — pricing pages change without notice, onboarding gets rebuilt — so a teardown older than a quarter is a historical document. The archived evidence folders make the next pass a diff rather than a restart, which is also when the matrix becomes genuinely interesting: what changed tells you where the category is moving.

Screenshots are what make a teardown a teardown: a dated capture of what the product showed settles in seconds the arguments that memory and marketing copy keep alive for weeks. But evidence collection on competitor products has rules, and the time to set them is in the dimensions file, before any agent or person starts capturing. The baseline as practised in the source workflow: trial accounts and public pages only, stay within terms of service, and never misrepresent who you are to get access. If a competitor gates the product behind a sales call, that is a finding about their go-to-market — record it as such rather than treating it as a hole to fill by other means.

The second rule is about use, not collection. Competitor screenshots are internal evidence: they sit in the evidence folders, they back the matrix cells, and team members open them to verify findings. Republishing them — in pitch decks, public articles, or marketing — is a different act with different legal and reputational exposure. The agency case in the source workflow is the model: the pitch's central argument was built on a gap the teardown found, illustrated with the agency's own concept work, with the matrix kept as an appendix demonstrating rigour.

Dating matters more than it seems. Pricing pages and onboarding flows change without notice; an undated screenshot is unverifiable within a quarter. The manifest — every file, what it shows, where it came from, when it was collected — is what keeps the evidence auditable and makes the next teardown a diff.

The pattern survey is the synthesis layer of the teardown and the part most worth presenting to the team, because it answers the question stakeholders actually ask: where does the category set expectations, and where is there room to be different? Table stakes are the patterns that have become the category's grammar — in the project-tool example, guided setup checklists and template galleries; in the pricing example, three tiers with the middle anchored and an annual toggle. Users do not praise these patterns; they notice their absence. Differentiators are the patterns one or two products own, and they are only useful in the matrix when the evidence shows how the execution differs, not just that a feature exists.

Counting discipline is what separates a survey from an impression. The useful sentence has a numerator and a denominator with citations behind it: four of the five researched products show a setup checklist on first run, with the evidence files listed. The vague version — most products do this — cannot be checked, ages silently, and gets repeated long after it has stopped being true.

Keep the limit visible: prevalence is not proof of effectiveness. A pattern shared by the whole category may be a well-tested best practice or a herd error nobody has questioned; usage data, conversion, and retention are invisible from the outside. The matrix tells you what exists. Whether to follow the convention or break it is a design judgment, made with your own users in view — which is also where the gaps list comes in, since the most interesting output is often the thing nobody in the category does well.

This is the discipline named in the module's learning objectives, and it is worth treating as a formatting rule rather than an aspiration: the packet keeps observations, inferences, and recommendations in separate, labelled layers. An observation cites an evidence file and can be verified by anyone who opens it. An inference is a reading of what the observations mean — it can be reasonable, but it is a step away from the evidence and has to be labelled as that step. A recommendation is a design decision, and it belongs to the designer who will be accountable for it, informed by the evidence but not dictated by it.

The failure modes are all blurrings of these layers. Marketing claims laundered into observed facts — the cross-check exists to catch exactly that. Inferences written in the confident voice of observations — the report says users prefer when the evidence only supports four of five products show. And recommendations smuggled inside findings, so the team experiences the packet as a decision already made rather than evidence to decide with. Agents are particularly prone to the second failure: synthesis models reach for fluent, confident phrasing, so the layer labels need to be demanded in the output format, not hoped for.

A practical convention: in the matrix and the findings document, observations carry file citations, inferences carry an explicit 'inference' tag and name the observations they rest on, and recommendations live in a separate one-page document the designer writes by hand. That last part is deliberate — writing the recommendation page yourself is what keeps the judgment human.

The cross-checking step in the deep research workflow — a claim must be supported by two independent sources or one primary source, or it gets dropped and listed separately — is the mechanical defence against repeated misinformation. The pricing-page case from the source workflow is the cautionary tale: of fourteen widely repeated claims about pricing pages, only six survived cross-checking, and the famous precise-percentage conversion claim traced back through a circle of articles to a single vendor case study with no published data. Reading the dropped-claims appendix before the findings is a habit worth building; it is often the most useful page in the report.

But cross-checking has limits worth stating plainly. It cannot detect a claim that is wrong everywhere it appears — a myth the whole industry repeats survives any vote among sources. It inherits the biases of what gets published in English on the public web: paywalled research, internal data, recently changed product surfaces, and non-English sources are systematically underrepresented. And it says nothing about your users; web research describes the category, not your product.

The per-source guidance on the slide is the practical layer. Marketing copy is a claim to verify. Docs and changelogs are good primary sources for what exists, silent on quality. Reviews surface real friction but skew angry and recent. Published teardowns date quickly and cite each other. Standards should be fetched from the primary text and cited by criterion number — that precision is what later lets an accessibility lead verify the load-bearing claims in minutes rather than redoing the research.

Each row of this table is a real failure pattern from the source workflow's quality comparison, and they are worth reading as a pair exercise: what makes the left cell unusable is not that it is wrong, it is that it cannot be checked. 'Best onboarding in the category' is a judgment dressed as a finding; the good version gives the step count, the deferral detail, and the dated manifest, and lets the reader form the judgment. The AI-setup row is the cross-check in action — the claim is recorded verbatim with its URL, the walkthrough evidence contradicts it, and both sit in the matrix so the team's planning discussion stays grounded in what users will actually meet.

The transparent-pricing row shows the value of walking flows rather than reading pages: the published per-seat price and the sales-form redirect above twenty-five seats are both true, and only the walkthrough exposes the gap between them. The final row is the one that makes reviewers uncomfortable and is the most important: an estimated price typed into a cell is fabricated evidence, however plausible. 'Not collected', with the reason, is more useful — it is accurate, and the reason is itself a finding about how that competitor sells.

When reviewing an agent-built matrix, sample cells against their evidence files rather than reading the matrix as prose. Three to five cells per competitor is usually enough to know whether the run can be trusted, and it is the same spot-checking habit as opening citations in a deep research report.

This is a condensed view of the teardown traced in the source workflow: five project-management tools torn down in an afternoon ahead of an onboarding redesign, with one collection agent per competitor and the human time spent on verification and on arguing about what the matrix meant. The slide shows three of the competitors and four dimensions so the structure is readable in a 16:9 frame; the full matrix had every competitor as a column and every dimension from the rubric as a row, with each cell citing its evidence files.

Walk the rows. Step counts ranged from four to eleven across the full set, and the two shortest paths both opened into a populated template rather than an empty state — an observation, dated and screenshotted. The cross-check earned its keep on positioning: two competitors claimed AI-powered setup, and the walkthroughs showed one generated a real project structure from a text prompt while the other had relabelled its template picker. Both the claims and the observations went into the matrix verbatim, which kept the planning discussion grounded.

The finding that shaped the redesign came from the gaps file, not from any single cell: every product designed onboarding for the account creator and dropped invited teammates into a bare board with no orientation. That is the layered discipline working as intended — the observation is the bare board across the set, the inference is that the category has not designed for the second teammate, and the recommendation to make that the focus of the redesign was the team's call, made knowing it could equally have been a graveyard nobody had bothered to fix. The archived evidence became the baseline for a re-run two quarters later.

The exercise deliberately shrinks the workflow to a size one person can finish in about an hour without any orchestration: one competitor, one flow, evidence captured by hand. The point is not the matrix — with a single competitor there is nothing to compare — it is the habits: fixing the dimensions before capturing anything, recording the positioning claim verbatim and then checking it against the walkthrough, and writing the output in three labelled layers. Most people discover that the observation layer is easy and the discipline breaks at the inference layer, where the temptation to write conclusions in the voice of facts is strong.

Steer participants towards a flow connected to a live decision, because the exercise is also a test of the scoping slide: if they cannot name the decision the teardown serves, the flow they pick will be whichever one is easiest to screenshot. The 'one thing you could not see from outside' prompt matters too — it seeds the limits conversation, since usage data, retention, and the reasons behind a competitor's choices are invisible from a walkthrough.

If running this with a team, have two people tear down the same flow independently and compare folders. The differences are instructive: where the evidence agrees, the rubric worked; where the inferences differ from the same evidence, that is the judgment layer showing itself — which is exactly the part that should not be delegated to an agent.

Recap by connecting the bullets back to the module's three objectives. Scoping was the first: the question, the decision, the evidence standard, and self-chosen angles are what stop agent breadth becoming unfocused accumulation. The teardown pipeline and the rubric were the second: evidence captured per product against fixed dimensions, dated manifests, marketing claims cross-checked against walkthroughs. The layered output — observation, inference, recommendation — was the third, and it is the discipline that travels furthest beyond this module, because it applies to every research artifact an agent helps produce.

Restate the limits once more, because they are easy to lose in the enthusiasm of a good matrix: desk research and teardowns describe the category and the public web, not your users and not performance. They cannot say why a competitor built something or whether it works for them, and the evidence decays within a quarter. The packet informs the decision; it is not the decision.

Preview Module 3 concretely. The next module keeps the same packet structure and the same evidence discipline but changes the corpus: interviews, support tickets, reviews, and open-text feedback — your own data rather than the public web. The new problems are different in character: consent and privacy in the corpus, themes that must trace back to specific quotes, counting prevalence rather than loudness, and the interpretation pass that stays human. The habit built here — every claim points at evidence someone can open — is the one that carries over directly.