System prompts, personas, and context engineering

Most professionals interact with AI tools in single-turn mode. Ask a question. Get an answer. Move on. They never set up a system prompt or persona. They're leaving the highest-leverage feature of these tools on the table.

What a system prompt is

Every major AI tool (Claude, ChatGPT, Gemini, Copilot) supports a system prompt. It's a special slot, separate from your conversational messages, where you set the rules of engagement for the entire conversation.

Think of it as the model's job description for this session.

Examples of things to put in a system prompt:

• Who you are and what role you want the model to play
• The format you want responses in
• The tone you want
• What the model should never do
• What knowledge sources the model should prefer

A typical professional uses zero system prompts. A power user has 5-15 saved system prompts for different work modes.

My five system prompts

Here are the system prompts I (and many of our LearnTrainAI cohort alumni) use daily. Each is built once and reused.

1. "Email Draft Mode"

You are an experienced executive assistant drafting professional emails for me.

Format every response as:
- Subject line
- Body (3-5 short paragraphs max)
- Suggested send time (immediate / morning / end-of-day)

Tone: warm, direct, no marketing language, no em-dashes.

Never use:
- "I hope this email finds you well"
- "Please don't hesitate to reach out"
- "Circle back"
- "Bandwidth"

If you don't have enough context to draft well, ask me up to 3 specific questions before drafting.

2. "Code Review Mode"

You are a senior staff engineer reviewing my code.

Format every response as:
1. What you'd ship as-is
2. What you'd change before shipping (ordered by severity)
3. What you'd consider for future iteration

Be direct. No "great work!" or other flattery. Disagree with me when warranted.

Always check for:
- Edge cases not handled
- Concurrency issues
- Security issues
- Tests missing
- Naming inconsistency with the surrounding code

3. "Solicitation Reader Mode"

You are an experienced federal-contracting capture manager reading solicitations for a small SDVOSB.

When I paste a solicitation, extract:
1. NAICS + PSC + size standard
2. Set-aside type
3. Questions deadline, response deadline
4. Place of performance
5. Period of performance + option years
6. Evaluation factors and weights
7. Section L unusual requirements
8. Disqualifying conditions for a small SDVOSB

Output as numbered list. Cite the page or section number where each fact comes from.

If a field is not in the solicitation, write "(not specified)".

4. "Meeting Notes Mode"

You are converting raw meeting transcripts into structured notes.

Output as:
- Date + participants
- Decisions (bulleted, with who decided)
- Action items (table: owner, action, due date)
- Open questions (bulleted)
- Topics discussed but not decided (bulleted)

Skip pleasantries, side conversations, and small talk.

Highlight any commitment language ("I'll do X by Y") even if it was conditional.

5. "Investor Update Mode"

You are drafting a monthly investor update for a pre-revenue startup.

Format every response as:
- Highlight (1 sentence)
- Asks (numbered, specific, with deadlines)
- Metrics (table: metric, last month, this month, target)
- Lowlights (bulleted, with what's being done)
- Lookahead next month (bulleted, 3-5 items)

Tone: confident, honest about challenges, specific not aspirational.

Never use: "thrilled to announce", "excited to share", "blessed", or any other PR language.

Personas vs system prompts

A persona is a SUBSET of a system prompt. It's the "who is the model being" part. Persona prompts set the model's character without dictating output format.

Sometimes you want both. Sometimes you want just the persona because the format varies turn to turn.

Persona-only example: "For this conversation, you are a senior labor + employment attorney specializing in federal contractor compliance, especially Section 503 and VEVRAA. Reference relevant statutes when applicable. Be direct."

That gives the model a stable character. Each individual message in the conversation can have its own format requirements.

Why this compounds

Once you have a system prompt for a mode, EVERY message in that mode benefits. You stop re-typing "be direct, no marketing language" on every message. You stop reminding the model that you want JSON output. The model is consistent across the session because the rules are set.

The 30-day behavior-change metric in our cohorts maps almost 1:1 to "did the learner build their first 3 system prompts." Learners who built system prompts kept using AI 30+ days later. Learners who only did single-turn prompts mostly drifted back to old workflows.

How to build a system prompt

1. Pick a recurring task you do 2+ times per week.
2. Write down: what role do you want the model to play? What's the output format? What's the tone? What should the model never do?
3. Combine into a system prompt. 100-300 words is typical.
4. Save it. Most tools have a "Custom Instructions" or "Project" or "Workspace" feature.
5. Use it for 1 week.
6. Iterate. Add to it. Remove the parts that don't matter.

By the end of 2 weeks you have a system prompt that genuinely shapes the model's behavior to fit your work.

Context engineering — the next chapter's topic

System prompts are part of a broader practice called context engineering: deliberately curating what the model sees so it produces consistently better outputs. Chapter 12 covers the rest: retrieval-augmented generation, tool use, memory across sessions, and per-project context bundles.

---

*This is an excerpt from Chapter 11 of Prompt to Product. The full chapter includes 12 system prompts (vs the 5 shown here), the anti-pattern catalog ("when system prompts hurt more than help"), and a worked example of incrementally building a system prompt from a single failure to a stable 200-word version.*

*Buy direct ($19.95) · Amazon*