In early 2023, Morgan Stanley Wealth Management deployed an internal Claude-powered assistant to help financial advisors retrieve information from the firm's 100,000-page research library. The tool worked β but only after the team spent weeks refining a single element of their prompts: the role definition. When advisors asked general questions without context, Claude returned accurate but generic answers appropriate for a lay audience. When the system prompt established that the user was a licensed Series 7 financial advisor serving high-net-worth clients, the same questions produced precise, regulation-aware, jargon-fluent responses. The knowledge Claude drew on didn't change. The frame did.
The Morgan Stanley team called this the "advisor identity anchor" β a three-sentence preamble that specified role, audience, and stakes. It reduced follow-up clarification requests by an estimated 40% in their pilot cohort.
When you assign Claude a role β "You are a senior product manager at a B2B SaaS company" β you are not playing a game. You are activating a coherent cluster of assumptions: vocabulary level, what counts as a relevant consideration, how formal the tone should be, what tradeoffs matter, and what background knowledge to take for granted.
Claude's training exposed it to enormous volumes of domain-specific writing. Role framing tells it which portion of that knowledge to weight heavily. A prompt that begins "As a first-year medical student, explainβ¦" will produce a different answer than "As a hospitalist with 15 years of internal medicine experience, explainβ¦" β even if the topic sentence is identical.
The Morgan Stanley pattern generalizes cleanly. A strong role frame answers three questions in one or two sentences:
You don't need all three in every prompt, but the more expensive the decision or the more specific the domain, the more each element pays off.
Weak: "Explain cloud cost optimization."
Strong: "I'm a DevOps lead at a 60-person startup. Our CTO wants a plain-English explanation of cloud cost optimization strategies for a board meeting next week. Focus on AWS. Keep it under 300 words and avoid acronyms."
Many people specify their own role but forget to define the audience. Claude then has to guess β and it often guesses "general educated reader," which is rarely what you need. In 2022, research teams at Anthropic documented internally that audience specification was among the highest-leverage prompt modifications for improving output relevance in professional contexts.
Audience specification changes: vocabulary selection, assumed prior knowledge, depth of explanation, use of examples, and the presence or absence of caveats. A legal memo written for a partner differs radically from one written for a first-year associate β same facts, same law, completely different document.
The audience frame isn't about dumbing things down. It's about calibrating precision. Telling Claude your audience is "three senior engineers reviewing our API architecture" gives it permission to be technical, dense, and assumption-heavy β which is exactly what you want.
These four patterns cover the majority of professional prompting scenarios:
In this lab, you'll practice the three-part identity anchor with Claude. Start by describing a real work scenario β your role, who you're communicating with, and what decision is at stake. Then ask Claude to draft a prompt opener using the identity anchor pattern. Refine it together until it precisely fits your professional context.
When Notion integrated AI writing assistance into its product in 2023, the engineering team made a counterintuitive discovery during user testing: the feature that drove the most user satisfaction wasn't smarter generation β it was structured output. Users who received responses in bullet points, numbered steps, or clear header-delimited sections rated the AI dramatically higher than users who received equivalent information in unbroken prose.
The Notion team traced this to what they called the "clipboard problem." Users weren't reading Claude's output β they were copying it into documents, Slack messages, and slide decks. Prose required reformatting. Structured output was already done. The lesson cascaded into their product: Notion AI now defaults to structured formats for almost all professional writing tasks, and users can request prose explicitly.
Without explicit format instructions, Claude defaults to the format most common in its training data for the given topic β which is usually flowing prose. This is appropriate for essays, narratives, and explanations, but it's the wrong format for most professional deliverables. Memos, action items, project updates, product specs, and executive summaries all have established structural conventions that make them faster to scan and act on.
The good news: Claude follows format instructions with high fidelity. The challenge is that you have to give them explicitly and specifically. "Use bullet points" is better than nothing. "Use three-word bullet points" is better still. "Use a table with columns for Feature, Priority, and Owner" is best of all.
Format instructions range from vague to precise. More precise instructions consistently produce more usable output:
Claude's default length is calibrated to completeness β it tries to say everything relevant. For many professional uses, this is too long. An executive summary isn't improved by being comprehensive; it's defined by what it leaves out.
Word counts work. Token counts work. Structural limits work better. Instead of "keep it under 200 words," try "one paragraph of 3-4 sentences." Instead of "be brief," try "no more than five bullet points." Structural constraints are clearer than word counts because they force decisions about what to include rather than just truncating what's already there.
"Keep it brief."
"Don't be too long."
"Summarize this."
"Give me a short version."
"Three bullet points maximum."
"One paragraph, 4 sentences."
"Executive summary: 100 words."
"TLDR in 2 sentences, then details."
These patterns produce clipboard-ready professional output without revision:
Before finalizing a format instruction, ask: "Could I paste Claude's output directly into the final destination without reformatting?" If not, your format instruction isn't specific enough yet.
Pick a real deliverable you produce regularly β a status update, a summary email, a meeting agenda, a decision memo. Describe it to Claude, then work together to build a format instruction so precise that Claude's output could paste directly into your final destination without any editing. Aim for at least three rounds of refinement.
In late 2023, HubSpot's content marketing team published an internal post-mortem on their AI writing experiment. The team had spent six months using Claude to accelerate blog content production. Early results were disappointing: articles sounded "corporate" and "generic" β readers could tell. The team nearly abandoned the experiment.
Instead, a writer named Connor Cirillo proposed a different approach. Rather than asking Claude to "write in HubSpot's voice," he spent a week building what he called a "voice document" β 400 words describing the brand's tone using concrete examples: "We use second-person address. We write short sentences. We never use the word 'leverage' as a verb. Here are three paragraphs that exemplify our voice: [pastes examples]."
The difference was immediate and measurable. Editors' revision time dropped from an average of 47 minutes per AI-assisted piece to 11 minutes. The voice document became a standard component of HubSpot's AI content process. Tone isn't taught by instruction β it's taught by example.
Vague tone instructions β "be professional," "keep it friendly," "sound authoritative" β are Claude's weakest inputs. These words describe thousands of different styles. "Professional" at Goldman Sachs looks nothing like "professional" at Patagonia. Claude has no way to know which version you mean unless you show it.
The three most reliable mechanisms for voice transfer are, in order of power: example text, anti-examples (what to avoid), and constraint lists (specific rules). Used together, they constitute a voice brief that can be reused across many prompts.
A voice brief is a reusable prompt component β typically 200β500 words β that establishes your writing style for Claude. It contains four elements:
Second-person address ("you"). Short sentences under 20 words when possible. Active verbs. No jargon without definition. No "leverage" as a verb. No passive voice. Opener must make a concrete claim, not a question.
Most professional prompts require multiple simultaneous constraints β on voice, format, length, scope, and forbidden content. The challenge is specifying all of them without making the prompt unworkable. The solution is layering: address constraints in a consistent order so Claude processes them sequentially without conflict.
When constraints conflict β for example, a short format limit and a requirement to include many items β Claude will usually flag the conflict rather than silently dropping constraints. This is useful: it tells you to either loosen one constraint or split the request into two prompts.
The same content requires different tones for different audiences. A product launch announcement for internal engineers, for customers, and for investors requires three different tone calibrations β even if the facts are identical. The fastest way to manage this is to maintain three short voice descriptions and swap them into a template prompt.
"Write a product launch announcement. Keep it professional and exciting."
"Write a product launch announcement for our enterprise sales team. Tone: direct, confident, numbers-first. Lead with the business impact. Skip the origin story. No exclamation points."
Spending 30 minutes building a voice brief pays for itself across hundreds of prompts. Once built, paste it as a standing component at the top of every writing prompt. The HubSpot team estimated a 4Γ speedup in editorial review time from this single change.
You'll build a voice brief for your own writing or your organization's brand. Start by sharing 2β3 sentences that exemplify your target voice, then work with Claude to identify the underlying style rules, anti-patterns, and persona anchor. By the end, you should have a 200β300 word voice brief you can reuse in future prompts.
In 2023, Stripe's developer documentation team piloted Claude for technical writing. Their senior writer Gina Trapani documented the learning curve in an internal retrospective: the team initially tried to produce final documentation in single prompts. The output was competent but required significant revision β approximately the same effort as writing from scratch.
The breakthrough came when the team adopted a four-stage chain: first, ask Claude to outline the document and identify all the technical claims that need verification; second, review and correct the outline; third, expand section by section with the corrected outline as context; fourth, ask Claude to review its own output for consistency, jargon, and structural completeness.
The result: revision time dropped by 60%. More importantly, the errors that remained were content errors β wrong technical facts β not structural or prose errors. Human reviewers could focus entirely on accuracy rather than writing quality. The chain separated concerns: Claude handled structure and prose; humans handled ground truth.
Every complex deliverable has a single-prompt ceiling β a level of quality that cannot be reliably exceeded by refining one prompt further. Beyond that ceiling, the only path to better output is breaking the task into sequential steps: outline, then draft, then review, then refine. This is how professional writers work. Claude works better the same way.
The signals that you've hit the ceiling: Claude's output is structurally correct but substantively thin; you're writing longer and longer prompts trying to specify every nuance; you keep getting the same weakness regardless of how you phrase the instruction.
The Stripe pattern generalizes to most complex writing tasks:
The most common error in prompt chaining is failing to carry context forward. In a long conversation, Claude can lose track of earlier decisions. In a new conversation, it has no access to previous sessions at all. The solution is explicit context summaries.
At each stage transition, summarize what has been decided: "We've agreed the document will be structured as [X]. The target audience is [Y]. The tone is [Z]. Now, expand section 2 following those constraints." This reinvests the decisions from Stage 1 at every subsequent stage.
"Context: [Role]. [Audience]. [Format decided]. [Voice rules agreed]. [What has been produced so far: brief description]. Next task: [specific request]."
Not every task needs a chain. Short deliverables β emails, summaries, single-section documents β usually improve faster through prompt refinement than chaining. The decision rule: if the deliverable has more than three distinct structural components, or if it will require human review of content (not just prose), chain it. If it's a single coherent piece under 500 words, refine the prompt.
Email drafts. Slack messages. Short summaries. Single-section explanations. Social media posts. Meeting agenda items.
Full reports. Technical docs. Marketing campaigns. Proposal decks. Policy documents. Multi-section analyses.
Design your chains so humans handle ground truth and Claude handles structure and prose. When a chain is working well, human review time is spent on content accuracy β not reformatting, not restructuring, not rewriting prose. That's the sign the chain is correctly dividing labor.
Choose a complex deliverable you genuinely need to produce β a report, a proposal, a multi-section analysis, a policy document. Work with Claude to design a four-stage prompt chain for it: structure, expansion, self-review, and targeted revision. Then execute Stage 1 together: produce and agree on the outline. You'll see immediately how the chain approach changes the quality conversation.