Double Diamond
TL;DR
The Double Diamond — the mid-century design process the UK Design Council formalized in 2005 as diverge, converge, twice — runs work through four phases in two diamonds: Discover (diverge on the problem), Define (converge on a problem statement), Develop (diverge on solutions), Deliver (converge on one). Its governing insight for regulated work is that the first diamond is about the problem, and skipping it means solving the wrong question well. As a graph, the two diamonds are separated by a framing gate: the problem statement must be carried by evidence and must not presuppose a solution, and no option may be generated until it passes. A selection gate closes the second diamond, binding the chosen option to criteria written before the options were ranked. Reach for it whenever the presenting complaint is not yet the defect — which, after an examination finding, is most of the time.
The remediation that answered the finding
The finding is two sentences long: callback verification was not evidenced on eleven of forty sampled high-risk payment instructions. The response is due in thirty days, and by the end of the first week the remediation plan is drafted — an attestation checkbox on the payment screen, a refresher module for the operations team, and a monthly quality sample to show the fix holding. It is a competent plan, produced quickly by people who know the process. It is also a plan written entirely inside the second diamond: the finding named a symptom, and the team began generating solutions to the symptom before anyone established what it was a symptom of.
An analyst who spends three days in the first diamond comes back with something different. The eleven unverified payments are not scattered; they cluster after 3 p.m., and they arrive through a same-day channel whose cutoff falls outside the callback desk’s coverage window. Nobody skipped the control — for a predictable slice of volume, the control was not reachable. That reframing changes what a fix even is. An attestation added to an unreachable control produces attestations rather than verifications, and the monthly sample would have shown the fix holding right up until the next examination. “Staff did not follow the procedure” is a problem statement with its remedy already inside it; “the control is unavailable on one channel after cutoff” is a problem statement that still has to be answered.
An unguided model asked for a remediation plan returns a remediation plan — the prompt named a solution shape and generation obliges. The Double Diamond graph cannot reach options at all until a problem statement has passed a gate that checks it is carried by evidence and does not smuggle in its own answer.
How the graph works, step by step
Discover — diverge on the problem. A generation node explores the problem space rather than the solution space: what else is true around the presenting complaint, who else touches it, when it does and does not occur, where in the data it clusters. The output is not a list of fixes — it is a bounded set of candidate accounts of what is failing, each tied to retrieved evidence. The node runs against a declared budget of time, retrievals, or candidates, because divergence without a bound is the failure mode this phase is known for.
Define — converge on the problem statement. An evaluation node narrows the candidate accounts to one statement, and the framing gate sits here, between the diamonds. It checks two things: that the statement is carried by specific evidence, and that it does not presuppose a solution. The second condition is the strategy’s sharpest instrument. “Staff need training” and “the queue routes around the control” are not two descriptions of one problem; they are two different problems, and only one of them survives contact with the evidence.
Develop — diverge on options. Only now does the graph generate solutions, and it generates several against the defined problem, with the criteria for choosing among them written down before anything is ranked. Options that answer the presenting complaint but not the defined problem become visible immediately — the attestation checkbox is still on the list, and it is now plainly a fix for a different problem. That legibility is the practical return on the first diamond.
Deliver — converge on one. A second evaluation node scores the option set against the stated criteria, and the selection gate admits a single choice with its reasoning attached: which option, against which criteria, answering which problem statement, and what the runners-up were. The discarded options stay in the record. A reviewer who disagrees with the choice can see what was declined and why, which is a materially different conversation from asking why nothing else was considered.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Framing gate | The problem statement is carried by specific retrieved evidence and does not contain its own remedy | The statement restates the presenting complaint, or names a fix in the grammar of a diagnosis — “staff need training,” “the policy needs updating” |
| Breadth gate | Each diamond explored within its declared budget and the set it produced spans materially different accounts or mechanisms | Every candidate is a variant of the first one, or divergence ran past its budget without a recorded reason — the two shapes of paralysis |
| Selection gate | The chosen option is traced to the defined problem and scored against criteria recorded before the options were ranked | Criteria were written after a favorite emerged, or the winner answers a narrower problem than the one defined |
Where it fits — three use cases
1. Remediation design after an examination finding
A finding names what an examiner observed, not what failed. The first diamond converts an observation into a diagnosis and does it on the record, so the response carries an account of the underlying defect and the evidence for it — the part that makes a remediation credible rather than merely responsive. It also reduces and interrupts the most expensive outcome in this work: a fix that closes the finding while leaving the mechanism that produced it in place, to be rediscovered at the next examination.
2. Policy and procedure design where the presenting complaint is not the defect
Complaint volume on a fee disclosure, repeated escalations at one adjudication step, a procedure everyone quietly works around: each arrives as a request to change a document. The first diamond asks what the complaints have in common before the second writes new language, because a procedure rewritten against the wrong problem statement is a change-control exercise that buys nothing and costs a training cycle.
3. Control design — choosing what to control before choosing the control
Control selection is where the two diamonds are collapsed most often: the risk is named and a control is named in the same breath, usually the control the team already knows how to build. Holding them apart forces the intervening question — what precisely is the thing to be controlled, and at which point in the process is it observable? — and the answer frequently moves the control upstream of where the first instinct put it.
When to reach for it
Reach for the Double Diamond when the problem arrives already framed by someone else — an examiner, a complaint, an incident report, an executive request — and the framing is the thing most likely to be wrong. It is the wrong shape when the problem is genuinely settled and only the option set is open; running a first diamond over a well-understood problem manufactures ceremony, and decomposition or constraint satisfaction will serve better.
Two limits are worth stating plainly. Most tooling and most organizational habit live in the second diamond — options, scoring, plans — while most of the value is lost in the first, which means the graph spends its governance budget defending a phase that feels like delay to everyone waiting for the plan. And a diamond that diverges without a bound becomes analysis paralysis: the gates are time-boxed by design, and a first diamond that exhausts its budget without a defensible statement should escalate rather than run longer. The graph reduces and interrupts the wrong-problem failure. It does not settle whether the problem statement is right — the trace exists so a reviewer can read the framing, see the evidence behind it, and disagree.
- Problem-space divergence
- Exploration deliberately scoped to what is failing rather than to what might fix it; the first diamond’s only output is candidate accounts of the problem, each tied to evidence.
- Solution-presupposing statement
- A problem statement whose wording already contains its remedy — the framing gate’s principal target, because it turns the second diamond into a formality.
- Divergence budget
- A declared bound on exploration — time, retrievals, candidates — recorded with the phase, so breadth is a choice on the record rather than an open-ended search.