Task statement 4.1 · Workflow integration and solution design

4.1 · Analyzing Requirements and Use Cases

In Wrenfield Housing, the course example, the operations team is starting to redesign how a repair request moves from report to booked visit. First it uses Claude to work out the requirements and the use cases, the steps where Claude could do or support the work. If either list rests on Claude's guesses rather than on what staff said, the redesign starts from the wrong brief.

Why it's tested

The tempting habit is to hand Claude the notes, ask for "the requirements" and accept a long, confident list. Where the notes are silent, Claude fills in what similar processes usually have, so staff needs and features nobody asked for sit side by side, unmarked. Use cases fail the same way: telling tenants where their repair stands sounds ideal for Claude, but it needs live data from a system Claude cannot reach.

KnowledgeRequirements: what the process must do

A requirement is a statement of what a process or change must do, for whom, and under which constraints. A constraint is a limit the design must respect, such as a policy rule or a channel that must stay open. The evidence for requirements is in the words of the people who run the process: interview notes, complaints and the current process description. Claude groups many notes quickly, but it knows only what the conversation gives it (1.1), so where the notes say nothing it fills the gap from what similar processes usually need. Judge a requirements list by where each line came from, not by how complete it looks.

Anthropic's guidance on reducing made-up claims gives the two moves that make the list checkable. First, ask Claude to support each claim with a direct quote and to drop any claim it cannot support. Second, give Claude permission to say the material lacks the information, which for requirements becomes a list of open questions:

  1. Give the material: the anonymised notes and the current process, each in its own tags.
  2. Ask for requirements with evidence: who needs each one, and a quote from the note that supports it.
  3. Set aside the extras: anything no note supports goes in a separate list of ideas, not requirements.
  4. Ask for the open questions: where the notes are unclear, give no figures, or disagree.
  5. Close the gaps with people: put the questions to the staff, then revise.

KnowledgeUse cases: where Claude fits

Choosing which steps suit Claude is what Anthropic's AI Fluency course calls Delegation: setting goals and deciding whether, when and how to engage with AI. Its three parts fit the job. Problem Awareness means understanding your goals before involving AI, which here means having the requirements first. Platform Awareness means knowing what an AI system can and cannot do. Task Delegation means sharing the work so that people and AI each do what they are good at. So judge each candidate step in three checks, in order: does it meet a requirement, does it need data Claude cannot see, and does it decide something about a person?

The data question is the easiest to miss, because a step can sound perfect for Claude while its data sits out of reach. In a chat, Claude sees the conversation, attached files and what any connector brings in. A connector is a link that lets Claude read another service such as Google Drive. Wrenfield's Team plan connects Google Drive and Gmail, but nothing connects its repairs system. A step that needs live data from that system needs an integration, a built connection between Claude and the system. Anthropic's help pages describe a custom connector as a remote MCP server, a small service that lets Claude talk to another system. Developers build and host these, and on a Team plan only Owners can add one, so such a step is escalated to a developer or architect.

Interactive · four candidate steps, three checks

A candidate step for the repairs process runs top to bottom through three checks: value, does it meet a requirement in the notes; data, does it need data Claude cannot see; risk, does it decide something about a tenant. A NO at the first question ends at drop it. A YES at the second ends at developer. A YES at the third ends at people. Passing all three ends at Claude, with a person checking. The chosen candidate's path is solid and its outcome is labelled RESULT; everything off its path is dashed and faint. CANDIDATE STEP each step Claude might take on draft the damp and mould questions that call handlers ask tell tenants where their reported job stands decide how urgent each reported repair is build a tenant mobile app CHECK 1 · VALUE Does it meet a requirement in the notes? YES: R2 YES: R1 YES: R4 NO NO DROP IT an idea, not a requirement RESULT YES 2 · DATA Does it need data Claude cannot see? NO YES YES DEVELOPER needs an integration first RESULT NO 3 · RISK Does it decide something about a tenant? NO YES YES PEOPLE a person decides RESULT NO CLAUDE Claude drafts, a person checks RESULT Each candidate stops at the first question that rules it out.

Choose a candidate step for the repairs process and follow it through the three checks.

Damp questions: it meets R2 (each damp report records the room, the size and any spread), uses only the notes and the handbook, and decides nothing about a tenant. Claude drafts; a person checks.

Job status: it meets R1, but the status lives in the repairs system, which no connector reaches, so the step goes to the IT team's developers.

Urgency calls: it meets R4 (one set of urgency rules), but each answer decides how fast one tenant's home is repaired, so a person decides. Claude can draft the rules.

Tenant app: no handler's note asks for it, so it is an idea, not a requirement, and it stops at the first question.

Solid, ending at RESULT: the path this candidate takes. Dashed and faint: questions and outcomes it never reaches.

KnowledgeTwo lists, two kinds of evidence

The two lists are checked against different evidence, and requirements come first, because a use case is only worth judging if it serves one:

What the process must do

Requirement

Says what is needed, by whom, under which constraints. Checked against what staff and tenants said.

Wrenfield: "Tenants can find out where a reported job stands without calling."

Where Claude fits

Use case

Names one step Claude could do or support. Checked against what Claude can see and who must decide.

Wrenfield: "Claude drafts the damp and mould questions for call handlers."

SkillWorked example: the requirements list nobody asked for

Wrenfield's operations team interviewed five call handlers, A to E, about how repair requests arrive today: by the repairs line, the online form or a letter. The notes were anonymised, with no tenant names or addresses, as the AI policy requires. The first prompt asked for the requirements in one line, and the list came back long and polished:

# before
<notes>[handlers A–E]</notes> <process>[phone, form, letter]</process>
List the requirements for our new repairs process.
→ 14 requirements, incl. "a tenant mobile app", "a 24/7 chatbot",
  "automatic contractor booking": in none of the notes
# after
For each requirement: who needs it, and a quote from the note behind it.
Put anything no note supports under "Ideas, not requirements".
Then list the questions the notes leave open.
→ 6 requirements, each quoted · 3 ideas set aside · 4 gaps to ask about

The prompt asked for a list but not for its evidence, so Claude topped up the handlers' words with features common in repairs services. The new prompt makes every line carry a quote. R1, that tenants can find out where a job stands without calling, now rests on handler B's note: "Most of my afternoon is people chasing jobs they reported last week." The unsupported features sit in their own list, where nobody mistakes them for staff needs. One open question asks how many letters arrive each month, because handler C said letters wait up to a week to be typed in but gave no count. The team took the questions back to the handlers; designing the process from the list is 4.3.

Ask one question of each candidate step: where does it stop?
No requirement behind itSet it aside as an idea, not a requirement.
Needs data Claude cannot seeEscalate to a developer or architect for an integration.
Decides something about a personA person decides; Claude may draft the rules or wording.
None of theseA use case for Claude, with a person checking the result.
Takeaway
  • A requirements list is only as good as its evidence: ask Claude to quote the note behind each line.
  • The gaps in the notes are a result too, so ask Claude for the open questions instead of letting it guess.
  • A step that sounds right for Claude still fails if its data sits in a system that no connector reaches.

check yourself

0 / 3 answered

A clinic's practice manager gives Claude notes from interviews with six receptionists and asks for the requirements for a new appointment-booking process. Claude returns 15 polished requirements, including a patient app and automatic text reminders, which no receptionist mentioned. The partners will see the list next week and will ask where each requirement came from.What is the most effective next prompt?

Select one.

A council licensing team wants Claude's help with licence renewals. It uses Claude in chat with no connectors. Licence records sit in the council's own database, the law requires a licensing officer to make every renewal decision, and applicants' personal details may not go into prompts.Which TWO steps are the best use cases for Claude in the team's current setup?

Select two.

An HR coordinator asks Claude where it could help with onboarding, pasting in only the onboarding checklist. Claude's top suggestion is to send each new starter a personal first-week schedule built from the shift rota. The rota lives in a scheduling system, and IT says the suggestion cannot work with the team's Claude setup.What is the most likely root cause?

Select one.

sources

  1. Claude Platform Docs, "Reduce hallucinations": allow Claude to say "I don't know" when the material lacks the information; verify with citations, finding a supporting quote for each claim and retracting any claim without one. Checked 2026-10-15.
  2. Anthropic, "The AI Fluency Framework" (Dakan, Feller and Anthropic): Delegation as setting goals and deciding whether, when and how to engage with AI. Checked 2026-10-15.
  3. Anthropic, "AI Fluency: Key Terminology Cheat Sheet": Problem Awareness, Platform Awareness and Task Delegation as the parts of Delegation. Checked 2026-10-15.
  4. Claude Help Center, "Get started with custom connectors using remote MCP": developers build and host remote MCP servers; only Owners and Primary Owners add custom connectors on Team plans. Checked 2026-10-15.
  5. Wrenfield's call handler notes, the repairs requirements and the four candidate steps are this course's example.