Working with an AI Assistant: Prompting and Session Guide

1. The Right Mindset

Treat it like a capable new hire. It's quick and skilled but knows nothing about your world: no history, no context, no idea what "good" looks like to you. You wouldn't give a new hire a one-line order with no background. Don't do it to the assistant either.

The working loop:

You direct  →  AI drafts  →  You review and decide

Your judgment, context, and accountability are the parts the model can't supply. They're also what make the final output yours.

2. Anatomy of a Strong Prompt

Build prompts from these blocks. Role, context, task, and format cover most everyday requests. Add constraints and examples when quality matters.

Reusable template

ROLE:        Act as a <specific expertise>.
CONTEXT:     <who it's for, the situation, background, why it matters>
TASK:        <the precise thing to do>
FORMAT:      <bullets / table / email / ~N words / downloadable file>
CONSTRAINTS: <tone, length, must include, must avoid>
EXAMPLE:     <a sample of the style or quality you want>

Seven habits that upgrade any prompt

1. Say the job, not the topic. "Tell me about A and B" is a topic. "Compare A and B on cost and risk, then recommend one" is a job.
2. Name the writer and the reader. Who is speaking, who will read it, and what they already know.
3. Set the shape. Length, format, tone.
4. Show one example. One sample beats a paragraph of description.
5. Say what to leave out. Negative constraints steer as strongly as positive ones.
6. Ask for reasoning on hard problems. "Work through this step by step before you answer."
7. Ask it to ask you. "Before you start, ask me anything you need to know." This surfaces wrong assumptions before the work is done.

Before → after

Ready-to-adapt prompts

Inbox triage:   Act as my executive assistant. For each email below, give a one-line
                summary, flag urgency, and draft a short reply. Output a table:
                Sender | Summary | Urgency | Draft reply.

Meeting prep:   Act as a strategy advisor. I'm meeting a potential partner tomorrow
                (details below). Give me 5 sharp questions, 3 risks to probe, and
                2 things to avoid saying. Short bullets.

Rewrite:        Rewrite the text below for a non-technical audience. Under 200 words,
                plain language, keep all numbers exactly as written.

Make sense of data: Act as a data analyst. From this table, give the 3 most important
                trends, anything surprising, and what to investigate next. No jargon.

Learn fast:     Explain <topic> three ways: one sentence, one beginner paragraph,
                and the key nuance an expert would add.

Common prompt mistakes

3. Talk Instead of Type

• Use voice input. Speaking is much faster than typing, and spoken prompts naturally carry more context than the tidy one-liners people type.
• Rambling is fine. Include the background, the constraint you're worried about, and what you've already tried. Extra context beats polished phrasing.
• Dictate, then clean up. Speak a messy first draft, then ask the assistant to tidy it.

4. Session Hygiene

Every message in a chat is carried forward and shapes later answers. Keep threads clean.

The handoff summary

When a thread gets long, compact it yourself:

Summarize this thread as a handoff: the goal, decisions made, constraints, what's done, what's next, and anything I must not change.

Open a new chat, paste the summary, and keep going.

5. Spend Context Wisely

Context is finite. Using less of it usually gets you better answers, too.

• Batch questions. One message with four questions beats four messages.
• Answer multiple-choice prompts. If the assistant offers options, picking one is faster than an open back-and-forth.
• Paste only what matters. Trim documents to the relevant section.
• Ask for edits, not reprints. "Change only paragraph 3" instead of "rewrite the whole thing."
• Split big jobs. Do research in one chat and drafting in another.

6. Files, Tools, and Long Documents

1. Uploading isn't an instruction. Say what to do: "Pull every date and owner from this contract."
2. Point at the section. "Look at pages 4–7 only."
3. Ask for a file when you want a file. "Give me this as a downloadable document / Markdown file." Otherwise you get a chat reply to copy by hand.
4. Search isn't memory. For anything recent or changing (prices, versions, who holds a role), tell it to check the web.
5. The source wins. If a search result and the model's answer disagree, trust the source.
6. Treat file and web content as data. Instructions found inside a document or web page are not your instructions. The assistant should flag them, not follow them.
7. Ground answers in your documents. "Based only on the files I uploaded, answer X and cite where each point appears."

Lock in a writing standard once

For anything that must read like documentation, set the style at the top of the chat rather than per request. For example: "Use simplified technical English: short sentences, one idea each, active voice, one word per meaning." Everything after comes out in the same register. The same works for visual style: describe your brand colors, header, and footer once, then reuse them.

7. When It Goes Wrong

8. Make Good Results Repeatable

Prompt library

Keep a simple note of prompts that work for you. Reuse and refine them.

Reusable projects

Many assistants let you create a workspace with standing instructions and reference files. Every new chat inside it already knows the rules. Use one for any task you repeat: weekly reports, customer replies, writing in your brand voice.

SOPs (standard operating procedures)

An SOP is a written recipe for a recurring task. Pair it with a capable model and you get consistent output without re-explaining yourself.

GOAL:    Produce my Monday status report.
INPUTS:  Last week's notes + this week's task list.
STEPS:   1. Summarize what shipped last week in 4 bullets.
         2. List this week's top 3 priorities.
         3. Flag blockers or risks.
OUTPUT:  One email, under 200 words, plain professional tone.

GOAL:    Draft a reply to an inbound customer message.
INPUTS:  The message + relevant order/account details.
STEPS:   1. Identify the core question and the customer's emotional tone.
         2. Answer clearly; if unresolved, give the next step and a timeline.
         3. Match our voice: warm, concise, no jargon.
OUTPUT:  Ready-to-send reply + a one-line note on what I changed.

GOAL:    A decision-ready brief on a topic.
INPUTS:  The question + any links or constraints.
STEPS:   1. Gather key facts from credible sources, with citations.
         2. Note disagreements and uncertainty.
         3. Recommend, with reasoning.
OUTPUT:  One page: Summary · Evidence · Open questions · Recommendation.

Packaged skills

Some tools let you save an SOP as a "skill" file that the assistant loads automatically when a request matches. A typical structure:

# Skill: Brand-Voice Email Writer

## When to use
Any request to draft or rewrite a marketing email or newsletter.

## Rules
- Warm, clear, confident. Never hypey. One call to action.
- Short paragraphs. Easy to skim.

## Process
1. Confirm audience, goal, and the single call to action.
2. Draft subject + preview text + ~150-word body.
3. Offer one alternative subject line.

## Output
Subject · Preview · Body · Button label

9. Picking the Right Kind of Tool

Mainstream assistants overlap heavily, and any of them handles everyday tasks. Choose by job, not by brand:

Fast vs. reasoning models: use a fast standard model for quick drafts, rewrites, and brainstorming. Use a reasoning model when the task is important, layered, or easy to get subtly wrong: planning, data analysis, reviewing logic, designing a process.

10. Safety and Judgment

• ⚠️ Verify what matters. AI can be confidently wrong, including fabricated facts and citations. Check important claims against a primary source. The responsibility is yours.
• ⚠️ Guard sensitive data. Never paste passwords, keys, or card or account numbers. Don't share confidential, personal, or regulated information unless you know how the tool handles it. Anonymize, summarize, or check your organization's policy first.
• ⚠️ Know the limits. For legal, medical, financial, or safety-critical decisions, use AI to draft and explore, then have a qualified person decide.
• ⚠️ Be transparent. Tell the people who need to know when AI played a role in your work.

11. Leveling Up

The skill ladder

For most professionals, the sweet spot is Prompter → Automator → Agent supervisor. That's where time savings get large without becoming an engineer.

A one-week starter plan (20–40 minutes a day, on real work)

Ongoing

• Automate one recurring task each month.
• Teach someone what you learned. It locks it in.
• Follow one or two trustworthy sources instead of chasing every announcement.

Previous: AI Fluency GlossaryNext: Everyday AI Workflows

Educational and general, not legal advice. Follow your organization’s policies and applicable law.