System Prompts and Personas

In Tier I, participants wrote one-off prompts for individual tasks. A system prompt changes the game. It defines Claude’s identity, voice, constraints, and standards before the first message is sent. Every response in the conversation inherits those settings. Think of it as a job description that Claude reads before clocking in. A system prompt that says “be professional” produces nothing useful. A system prompt that bans specific words, mandates specific formats, and defines the audience by name produces consistently excellent work across dozens of interactions.

System prompts live in different places depending on the interface. In Claude Projects on the web app, they are the “Custom Instructions” field. In Cowork, they live inside skills. In the API, they are the system parameter sent with every request. The concept is identical across all three. Only the plumbing changes.

A well-built system prompt has four sections. Identity: who Claude is and what expertise it brings (“You are a senior financial analyst specializing in manufacturing companies”). Voice: how it writes (“Direct, data-driven, no jargon, every statement backed by a number”). Constraints: what it avoids (“Never recommend a solution exceeding $50,000 without flagging the budget threshold. Do not use the words synergy, leverage, or alignment”). Output standards: what the finished work looks like (“End every response with a concrete next step. Use tables for comparisons. Keep responses under 500 words unless asked for more”).

The specificity of these sections determines the quality of every response that follows. Vague instructions produce vague behavior. Precise instructions produce consistent, high-quality output.

Ironclad Industrial: direct, no-nonsense Midwest manufacturing. Short sentences. Blue-collar vocabulary. Confidence without flash. The voice sounds like a plant manager who respects your time. Verdant Health: warm, knowledgeable, approachable wellness. Longer sentences. Empathetic tone. Scientific credibility with accessible language. The voice sounds like a trusted doctor who also happens to be a good listener. Vex Cybersecurity: sharp, technical, slightly aggressive. Punchy cadence. Industry-specific terminology used without apology. The voice sounds like a security analyst briefing a board after a near-miss.

Each company has three writing samples in the demo file. The samples are distinct enough that a room full of participants can identify which company produced which sample after reading two sentences. That distinctiveness is the standard the exercise aims for.

The demo file includes four system prompt templates. Internal Communications Writer: bans corporate jargon by name (synergy, alignment, leverage, circle back), defines sentence rhythm (short for facts, longer for explanation), mandates ending every piece with a clear action item. Sales Proposal Analyst: reads proposals through the buyer’s lens, flags vague ROI claims, insists on specific timelines and dollar amounts, formats output as a risk assessment. Customer Support Specialist: matches the customer’s emotional register, escalates instead of guessing on technical questions, keeps responses under 150 words, always includes a next step. Data Analyst: leads with the finding not the methodology, uses tables for comparisons, flags data quality issues before presenting conclusions, defines confidence level for each finding.

Walk through the Internal Communications Writer template in detail. It is the most universally relevant because every participant writes internal communications. Point out the specific, named prohibitions. “Do not use the word synergy” is more useful than “avoid jargon” because Claude can follow a specific ban. It cannot reliably follow a vague one.

The demo file includes a three-question rubric for evaluating the outputs. Does it sound like the same company? Would it fit on that company’s website without edits? Does it avoid generic AI voice? Score each question yes or no. Two or three yeses mean the system prompt is effective. Zero or one means it needs more specific constraints. The most common failure: the system prompt captures the content domain (manufacturing, health, security) but misses the voice (sentence rhythm, vocabulary level, emotional temperature).