Task statement 1.4 · Prompting and task execution
1.4 · Prompting by Task Type
In Wrenfield Housing, the course example, the operations team has 640 repair jobs waiting. It asks Claude four things about that one backlog: what its figures show, what other housing providers have done, how to brief the board, and how to clear it. Each is a different task type, so one style of prompt gets most of them wrong.
share
The tempting habit is to write every prompt the same way: describe the result and wait for it. That works for a letter, because the reader's needs are the test. On your own figures, the same prompt returns a confident conclusion with nothing to show where it came from, so you cannot check it. On outside facts, it gives no sources or dates. For ideas, it returns a few safe options, because asking for "the best ideas" makes Claude judge before it has explored.
KnowledgeWhat changes from one task type to the next
The parts of a prompt from 1.1 stay the same for every task: the task, the purpose and audience, the facts, the limits and the format. What changes is where the result's quality comes from, so each task type puts its weight on a different part. The exam guide names four task types. Analysis is work on material you give Claude, such as a spreadsheet. Research is gathering information that is not in the conversation yet. Drafting is writing for a reader. Brainstorming is generating options to choose from. This rules out any answer that treats one well-built prompt as right for every task.
So name the task type first, then build the prompt around it:
- Give what the type works from: data, a scope and dates, a reader and an example, or a problem and its limits.
- Ask for what lets you check it: evidence, sources, a match to the example, or a wide spread of options.
- Plan the next turn: check the figures, open the sources, edit the draft, or choose from the list.
Interactive · one backlog, four task types
Choose the backlog's task type.
Analysis: Claude works only from the export, so the prompt asks for the counts behind each finding, and you check them.
Research: the answer is outside the conversation, so the prompt sets scope and dates, and you open the cited sources.
Drafting: the board is the reader, so the prompt gives tone, format and an example, and a person checks the draft.
Brainstorming: range comes first, so the prompt asks for 20 unranked ideas, and a second turn chooses.
Analysis is where a confident answer does the most harm, because it sounds as if it came from your data. Anthropic's prompting guide gives two moves for long documents and data: put the material near the top, above the question, and ask Claude to quote the relevant parts before it carries out the task. The quotes show which rows each finding rests on, so you can check a finding instead of trusting it. Anthropic's prompting advice also suggests telling Claude to say so when the data is insufficient, rather than speculating.
Research differs because the material is not in the conversation yet, so the prompt must say where to look and what counts as a good answer. Anthropic's guide asks for clear success criteria and for checking information across several sources. For an associate, the scope is usually a place, a kind of source and a date range. Web search, a quick lookup, suits a fact that one or two searches can answer. Research is a mode for in-depth investigation. It runs searches that build on each other for one to three minutes, can use connected tools such as Google Workspace, and returns a longer report with citations.
Drafting and brainstorming pull in opposite directions. A draft has a reader, so its quality comes from the audience, the tone and the format, and an example of the style is often the fastest way to show all three. Anthropic's guide recommends wrapping an example in <example> tags so Claude can tell it from your instructions. Brainstorming needs range instead, so the prompt asks for many options and names the kinds to cover. Anthropic's blog quotes the designer behind Claude Design: ask for ten options, then remix the best, because most will not be good. Judging belongs in a second turn, because asking for the single best idea makes Claude filter before it has explored.
KnowledgeOne backlog, four prompts
Here is Wrenfield's backlog as each task type:
The export first, in <data> tags, then: "Quote the counts behind each finding. If the data cannot show a cause, say so."
Comes backFindings the team can check against the export.
With Research on: "How have other housing providers cut repair waiting times? Sources from 2024 on, with a link for each claim."
Comes backA cited report. The team opens its sources before quoting it.
"Draft a one-page briefing for the October board meeting: neutral tone, three headings, in the style of the example." Last quarter's briefing goes in <example> tags.
Comes backA draft the operations director checks before the board sees it.
"Give 20 ideas to cut the backlog, across scheduling, contractors, tenant reporting and staffing. Do not rank them yet."
Comes backA wide list. A second turn picks three that need no new IT system.
SkillWorked example: the backlog analysis that read like an opinion
Wrenfield's backlog export lists 640 open jobs by job type, estate, date reported and status, with no tenant names or addresses, as the AI policy requires. The first prompt asked for a verdict, and the verdict blamed something the export does not contain:
# before [backlog export attached] What's going wrong with our repairs backlog? → "Delays are mainly caused by contractor shortages. Consider hiring." # after <data>[backlog export: 640 open jobs]</data> Which three job types have the longest median wait, in days open? For each one, quote the job count and the median from the data. If the data cannot show why a job type waits, say so. → "Damp and mould: 112 jobs, median 41 days … the export has no contractor field, so it cannot show the cause."
The cause was a drafting-style prompt on an analysis task. "What's going wrong" asks for a conclusion. The export holds no contractor data, so Claude supplied a plausible cause from general knowledge and stated it as a finding. The new prompt asks questions the export can answer, asks for the figures behind each finding, and lets Claude say where the data stops.
More precise than the guide. The exam guide says "research mode"; the Claude apps call the feature Research. On paid plans, including Team, you turn it on from the + menu in a chat, with web search on (checked 2026-10-15). Exam answer: "research mode" means Research, chosen for an investigation across many sources rather than a quick fact.
| Your own data or documents | Analysis: put the material first and ask for the evidence behind each finding. |
| Information outside the conversation | Research: set the scope and dates, use Research, and ask for a source for each claim. |
| What a reader needs | Drafting: give the audience, the tone and the format, with an example. |
| A range of options | Brainstorming: ask for many varied ideas, then choose in a second turn. |
- The parts of a prompt stay the same; the task type decides which part carries the weight.
- Analysis and research need support you can check: quotes from your data, or a dated source for each outside claim.
- Drafting needs a reader and an example; brainstorming needs range first and judging later.