Task statement 4.3 · Workflow integration and solution design
4.3 · Supporting Solution Design
In Wrenfield Housing, the course example, the operations team has its requirements for the repairs-request redesign and a pilot plan for the Millbrook estate. Now it must build the solution: a guide that call handlers use to sort each incoming repair request. If the team takes Claude's first idea, drafts it once and never tests it on real requests, a tenant with black mould can wait in the routine queue.
share
The tempting request is "write us a triage guide", followed by rolling out whatever comes back, because the draft looks finished. Claude then offers one plausible design, so the choice is made before anyone compares options. The draft has also never met a real request. The opposite mistake is to plan something the team cannot build, such as a rule inside the repairs system. That work needs code or a system connection, so it goes to the developers as a written requirement.
KnowledgeThree stages, and who decides at each
Here a solution is a process and its working materials, such as a triage guide, a template or a checklist, not software. Solution design is choosing the approach: Claude proposes several options and states the trade-off of each, meaning what you give up to get its benefit. Development is making the materials, which Claude drafts and the team checks. Iteration is improving them after real use: the team tests them on real cases, collects what failed, and Claude helps revise. At every stage Claude proposes and drafts, while the team chooses, tests and owns the result. This rules out adopting Claude's first proposal untested, and any answer in which the team builds an integration itself.
- Give the requirements: paste what the solution must do and the limits it must respect.
- Ask for options: several different approaches, each checked against every requirement, with what each one gives up.
- Choose, and split off: pick the option that meets the requirements; send anything needing code or a system connection to developers.
- Develop the materials: have Claude draft them in a form the team can edit and share.
- Iterate on real cases: test the materials on real work, collect every failure with the right answer, revise, and test again.
Interactive · building the triage guide, stage by stage
Step through the four stages and see what each one hands to the next.
Design: the team picks the decision guide and sends automatic tagging to IT.
Develop: Claude drafts guide version 1 and a pilot log; the operations lead approves them.
Test: the pilot log shows which requests the guide sorted wrongly. Those failures feed the next stage.
Revise: one change per cause of failure, then the guide meets real requests again.
KnowledgeDesign: options against the requirements
Claude should see the requirements before it proposes anything, because an option can only be judged against what the solution must do. Anthropic's guide to comparing options says to tell Claude what matters for your decision and to ask what you give up by not choosing the others. One proposal leaves nothing to compare. Wrenfield's team pasted the guide's three requirements, drawn from 4.1 and 4.2. Call handlers, who have no surveying training, sort every request when it arrives. No damp and mould request is ever sorted as routine. The method works from the pilot's first day without a new IT system. The categories were urgent (attend within 24 hours), damp and mould, and routine. Claude returned three options:
A one-page flow of three steps that staff work through in order.
Gives upSpeed: staff ask each step's question. Meets all three requirements, so the team chose it.
Staff tick words such as "leak" or "mould" that appear in the request.
Gives upEvery request that never uses the word, so damp can land in routine. Fails the second requirement.
The repairs system tags each online request by category before staff see it.
Gives upThe pilot start: it needs a change to the repairs system. Fails the third requirement, so it goes to IT.
The third option was worth keeping, but not for Wrenfield's staff to build. The operations lead asked Claude to draft a hand-over brief, a short written requirement that says what is needed and why and leaves how to the developers, then checked it and sent it to IT:
# requirement for the IT team (drafted with Claude)
Need: tag each new repair request as urgent, damp and mould, or routine.
Why: the pilot sorts by hand; online requests could arrive already sorted.
Must: never tag a damp and mould request as routine; staff can change any tag.
Test: the pilot requests, with the categories the surveyor confirmed.
SkillWorked example: development, then the first revision
For development, Claude drafted the guide as an artifact, the document Claude builds beside the chat for content that will be edited and shared (2.6). Version 1 had three steps. Step 1: is there danger now, such as no heating or hot water, water coming in, or a door that will not lock? Then it is urgent. Step 2: does the tenant mention damp or mould? Then it is damp and mould. Step 3: otherwise, routine. Claude also drafted a pilot log with three columns: request reference, category given, and category the surveyor confirmed. The operations lead adjusted lines with "Edit with Claude", then exported version 1 to Word. After the first pilot week on Millbrook, the log showed 9 of 96 requests sorted wrongly, and the first revision request was vague:
# revision request, before The guide got some requests wrong in the pilot. Make it more accurate. → two pages, five new steps, nothing about the actual misses # revision request, after These 9 requests were sorted wrongly in week 1 (request text only, no names or addresses), each with the category given and the one confirmed. Group them by cause. Change only the steps a group shows are wrong. Keep the guide to one page and keep everything else as it is. → step 2 adds tenants' words: black marks, musty smell, windows always wet → step 1 now asks: is water coming through a ceiling or near electrics?
The bad round lacked evidence: "more accurate" told Claude that something failed but not what, so it guessed and made the guide longer. The real failures had two causes. Six damp and mould requests went to routine because tenants wrote "black marks behind the wardrobe" or "windows always wet" and never used the words damp or mould. Three dripping taps went to urgent because staff read "leak" as water coming in. The pasted cases, without names or addresses as Wrenfield's AI policy requires, showed Claude what to fix, and the "keep" line left the rest unchanged, as in 1.3's one-change loop. Anthropic's testing advice is to mirror the real work, edge cases included. In week two, version 2 sorted 104 requests with 2 wrong, and both joined the test set.
- In design, ask for several options and the trade-off of each against your requirements; Claude proposes, and the team chooses.
- Iterate from evidence: collect the real cases the materials got wrong, with the right answer, and revise only what those cases show.
- A hand-over brief to developers says what is needed and why, and leaves how to them.