Edition 02
Reference
Tips & tricks
The junk drawer. Ten things that make a real difference, sourced from the people who study this stuff, tested in actual work. Each one comes with a prompt you can run right now.
Most AI training tells you about AI. This asks you to use it, with your judgment intact.
These aren’t features. They’re habits. The difference between someone who uses AI and someone who’s good at it.
Getting better outputs
The recursive review
After Copilot produces something, ask it to critique its own output. Not generally, specifically. It'll surface things you'd have missed, and then fix them. Two prompts. Better result.
Try this
What are the three weakest parts of what you just wrote, and how would you fix each one? Then apply those fixes.
Source: The Neuron
Let it write the prompt for you
Before you ask Copilot to do the thing, ask it to design the mission. Give it your notes or brief, have it figure out your real goal, constraints, and success criteria, then have it write the prompt you should actually use. Works best when you're not sure how to frame a complex ask.
Try this
Here are my notes on what I'm trying to do: [paste]. Read these and infer my real goal, success criteria, and any constraints I haven't spelled out. Then write the prompt I should use to get the best result.
Sources: The Neuron, Anthropic engineering team, One Useful Thing
Triple compression for the real point
Summarize something in 200 words. Then summarize that in 50 words. Then ask for the single sentence that captures the whole thing. Each pass forces a prioritization decision the model couldn't make until it had already tried once. By the third compression, you almost always have the actual insight, which was buried in the first output.
Try this
Summarize this in 200 words. [Paste content] Now summarize that summary in 50 words. Now: what is the single sentence that captures the whole thing?
Sources: The Neuron, Anthropic, practitioner compression guides
Getting better inputs
Neutral framing beats leading questions
AI agrees with whatever direction your prompt implies. "Find evidence that X works" hands it a conclusion before the question. Any verb like find, defend, confirm, prove, or support pre-loads the answer. You'll get confirmation instead of analysis. The fix is two words: ask "for and against" instead of "why it works."
Try this
What are the strongest arguments for and against [decision or approach]? Don't lean either direction, I want the honest tradeoffs.
Sources: AI Agent Factory (Washington Post, 2025)
Paste the example, not the description
Describing tone and format takes ten sentences and still misses something. Pasting an example of a previous output you liked takes five seconds and lands immediately. For an LLM, examples are the pictures worth a thousand words. Your own best work is the best style guide you have.
Try this
Write this in the same style as the following example. Match the tone, length, and structure, not the content. [Paste your example here]
Source: Anthropic engineering documentation
Tell it what not to do
Negative constraints, telling the model what not to do, improve output more than positive instructions. One practitioner tested two email approaches on 50 real clients: the negatively constrained version got a reply 64% of the time. The generic version got 19%. Identify your model's annoying defaults and ban them by name.
Try this
Write this without: hedging language ("it's worth noting," "it may be the case"), bullet points, opening with "Great question!", or fake enthusiasm. Make direct statements. Every sentence earns its place.
Source: Prompt Engineering 2026 practitioner research
Calibrating trust
Give it permission to push back
AI models are trained to be agreeable. Unless you explicitly tell it otherwise, it will invent suggestions to seem helpful, validate your thinking even when it's off, and agree with framings it should challenge. Nobody needs that kind of people-pleasing from their tools. One line fixes it.
Try this
If this is already strong, say so, don't manufacture suggestions. And if you think my framing is wrong, tell me directly. I'd rather have honest feedback than polite agreement.
Source: The Neuron Power User Guide
Change the audience, not the content
When an output is technically right but not landing, don't ask it to rewrite the content, ask it to rewrite for a different audience. "Rewrite this for someone who's skeptical" produces a different document than "be more persuasive." Audience shift forces the model to reconsider emphasis and what to cut, which is usually the actual problem.
Try this
Rewrite this for someone who is skeptical of this idea and hasn't bought in yet. Same facts, different framing. Lead with their concern, not our conclusion.
Source: One Useful Thing (Ethan Mollick)
Under what conditions would this be wrong?
After any recommendation or plan, ask this exact question. It moves the model out of advocacy mode and into analysis mode. You get actual tradeoffs instead of rationalization. Dan Shipper at Every calls it inverting the decision, not arguing against it, but finding the conditions under which the opposite would have been right.
Try this
You just gave me a recommendation. Now: under what conditions would that recommendation be wrong? What assumptions am I making that might not hold? When would the opposite call have been correct?
Source: Every / Dan Shipper
Show your reasoning before you answer
Add this to any prompt where the logic matters as much as the conclusion. When reasoning is visible, you can catch mistakes before they become mistakes in the output. Better still: you can redirect mid-flight. "Your reasoning in step two is off, here's what you missed. Now continue." That's a different conversation than editing a finished draft.
Try this
Before you give me the answer, walk me through your reasoning. Don't skip steps. I'll tell you if something looks off before you conclude.
Source: AI Agent Factory
Review before you send, especially when the output sounds confident.
Techniques are habits. Pick one to run today.
Why Copilot? →MillerKnoll AI · 2026