A useful work prompt is less like a clever question and more like a compact brief. It gives Claude the material, boundaries, success criteria, and output shape needed to produce something a colleague can review and use.
Anthropic's prompt engineering guidance starts with a clear definition of success and a way to test it. That is especially important at work, where a fluent answer can still be wrong, incomplete, confidential, or misaligned with the decision being made.
Define
State the business outcome, audience, deadline, and definition of done.
Ground
Provide the relevant documents, facts, prior decisions, and source boundaries.
Produce
Request a deliverable shaped for its real destination, not a generic essay.
Review
Check claims, omissions, tone, confidentiality, and the decisions a human must own.
A reusable structure for complex work
For long source material, visible sections help separate instructions from evidence. Claude's official guidance often uses XML-style tags for this purpose. The labels are not magic; they simply make the boundaries easier to follow.
<role>Act as [RELEVANT WORK ROLE].</role> <objective>[THE DECISION OR DELIVERABLE NEEDED]</objective> <context>[AUDIENCE, SITUATION, DEADLINE, PRIOR DECISIONS]</context> <sources>[PASTE OR ATTACH APPROVED MATERIAL]</sources> <constraints>[SCOPE, TONE, LENGTH, EXCLUSIONS, CONFIDENTIALITY]</constraints> <output>[FORMAT AND ACCEPTANCE CRITERIA]</output> <uncertainty>If a claim is not supported by the sources, mark it as unverified. Do not invent details.</uncertainty>
Eight Claude prompts for real work
Turn source material into a decision memo
Act as a concise strategy analyst. Use only the material inside <sources>. Prepare a one-page decision memo for [AUDIENCE]. Include: decision required, relevant evidence, options, trade-offs, risks, missing information, and a recommendation conditional on the evidence. Cite the source section for every material claim. Mark unsupported points as [UNVERIFIED]. <sources>[PASTE MATERIAL]</sources>
Create a research brief
Create a research brief about [QUESTION] for [AUDIENCE]. Separate verified facts, interpretations, disagreements, and unanswered questions. For every factual claim, show which supplied source supports it. Do not fill gaps from memory. End with the five highest-value questions for the next research pass. <sources>[APPROVED SOURCES]</sources>
Draft a difficult email
Draft an email to [RECIPIENT] about [SITUATION]. The goal is [OUTCOME]. Use a calm, direct tone; acknowledge [CONCERN]; do not admit facts that have not been established; and give one clear next step. Keep it under [LENGTH]. First list any assumption that could materially change the wording, then provide the draft.
Convert meeting notes into action
Turn these meeting notes into a clean record. Separate decisions, action items, owners, deadlines, unresolved questions, and risks. Do not infer an owner or deadline that is not written; use [NOT ASSIGNED]. Flag any contradiction between the notes. <notes>[PASTE NOTES]</notes>
Plan a realistic week
Help me plan the workweek from the tasks below. I have [AVAILABLE HOURS], these fixed commitments: [COMMITMENTS], and these priorities: [PRIORITIES]. Estimate effort as a range, identify dependencies, protect one buffer block, and explain what should be deferred if the plan exceeds capacity. Return a day-by-day table and a short risk note.
Compare options without hiding trade-offs
Build a decision matrix for [DECISION]. Options: [OPTIONS]. Criteria and weights: [CRITERIA]. Use only the evidence I provide. Show the score calculation, confidence level, unknowns, and the conditions under which a different option would win. Do not turn a weak assumption into a precise score.
Review a deliverable before it ships
Review this [DOCUMENT/PLAN] as a skeptical quality reviewer. Test it against the acceptance criteria below. Return: pass/fail by criterion, evidence, highest-risk omission, ambiguous language, unsupported claims, and the smallest revision that resolves each issue. Do not rewrite sections that already pass. <criteria>[CRITERIA]</criteria> <draft>[DRAFT]</draft>
Build a presentation narrative
Create a presentation outline for [AUDIENCE] that should lead to [DECISION/ACTION]. Use the supplied evidence only. For each slide provide one message, supporting evidence, suggested visual, and speaker note. Remove background that does not change the decision. Finish with likely objections and honest responses.
What the human reviewer must still do
A structured prompt improves consistency; it does not transfer accountability. Before a deliverable leaves your hands, verify the source claims, permissions, calculations, names, dates, legal or policy implications, and whether the tone fits the actual relationship.
✓Sensitive material is approved for the tool and unnecessary identifiers are removed.
✓The answer clearly separates supplied evidence from assumptions.
✓Numbers, quotations, names, dates, and links are checked at the source.
✓A qualified person reviews high-stakes legal, financial, medical, or employment decisions.
✓The final owner understands and accepts the recommendation—not just the prose.
The work standard
Give Claude a clear outcome, bounded evidence, realistic constraints, and a reviewable format. Then keep a human responsible for the final decision.