Remember when "think step by step" was the magic phrase?
Everyone had a little toolkit. Capitals for the important bits. "Be thorough." "Double-check your work."
Some of that is now making your answers worse.
Not my opinion, either. The companies that make these models say it themselves, in their own prompting guides.
Anthropic, on its newest model: "The model decides for itself how much to think."
OpenAI says step-by-step instructions "can sometimes hinder" its reasoning models.
Google says Gemini "may over-analyze" prompts written for older models.
Why? The new models take you very literally.
Write CRITICAL in capitals and they treat it like a fire alarm.
(Think of a very keen new hire who does exactly what you said, including the bit you didn't mean.)
Now a quick word from today's sponsor.
Can Robotics Make Regenerative Farming Scalable?
Greenfield Robotics is on a mission to give farmers alternatives to herbicide-dependent weed control. BOTONY is the beginning of a broader robotic farming system designed for a wider range of applications in the field.
The Reg A+ offering is now live for investors who want to be part of what comes next.
This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. Please read the Offering Circular and related disclosures before investing. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment.
Right. Back to it.
Five habits to drop
"Think carefully" or "think step by step." The models already decide how hard to think. When Anthropic removed that line in its own tests, answers started sooner "with no clear decline" in quality.
Capitals, CRITICAL and MUST. Anthropic's own checklist puts it well: "an anxious prompt produces a cautious, hedging model."
"Be thorough, don't be lazy." Anthropic's own audit tool flags it for deletion. In one Anthropic test, "be maximally thorough" turned into dozens of searches nobody needed.
"Double-check your work." Anthropic says lines like this just waste tokens (the chunks of text the AI reads and writes).
Describing a screenshot instead of attaching it. Anthropic says Claude Opus 5.5 reads charts better on its lowest setting than Opus 5 did on its highest. Attach the chart.
And a bonus sixth. "Try to" and "if possible" now get read literally. If you mean "do it", say "do it".
(OpenAI's GPT-6 Astra has a quirk of its own: it sometimes stops early. So tell it what finished looks like.)
Better answers: do this instead
Give the reason, not the volume.
Anthropic's example: "NEVER use ellipses" becomes "this will be read aloud by a text-to-speech engine, so never use ellipses."
Now the model knows why, and it applies the rule to cases you never thought to mention.
Say what done looks like. "A one-page summary a busy manager reads in two minutes" beats "be concise and thorough" by a mile.
Show one example of what good looks like. Anthropic calls examples "one of the most reliable ways to steer."
And keep the context about you, like your job and who you're writing for. Anthropic's checklist has a line I love: "Context is never cruft." (Cruft means junk.)
My opinion: most people's custom instructions are a museum of 2024 tricks. Nobody opens them, and they still shape every single chat.
Clean up your instructions in 5 minutes
Anthropic's engineers built a tool for exactly this, called prompt-audit. It combs through your instructions and flags the outdated lines.
It blew up on X last week. One post about it got saved nearly 10,000 times.
The catch: it's built for developers.
So here's the chat version. Copy your custom instructions from ChatGPT, or your Project instructions from Claude, and paste them in with this:
Below are the standing instructions I give you. They were written for older AI models.
Audit them line by line. Flag anything that:
- shouts (capitals, CRITICAL, MUST)
- asks you to think harder, be thorough or double-check
- tells you the steps instead of what the result should be
- contradicts another line
- says "try to" or "if possible" when I actually mean it
For each flag, tell me why it's outdated and give me the rewrite. If a line is fine, leave it alone, especially anything that gives you context about me. Then give me the full cleaned-up version.
[paste your instructions here]That "leave it alone" line matters. Anthropic's own rule for the tool: an audit that finds nothing should change nothing.
Make it stick
Paste the cleaned version back into your custom instructions. Every new chat picks it up from then on.
Then run the audit again whenever a big new model lands.
(At the current pace, that's roughly every three months. Somebody on Hacker News complained about exactly that this week.)
Cleaner instructions get you better answers in every chat. In a business, better answers mean drafts you can actually send, and that's how Run On Claude is built.
Every month, Claude takes one job off your plate. Right now that's your social posts and your customer emails, with a new job each month picked by member vote.
Each job comes with the instructions already written for Claude. You set it up once with a 20-minute guide, Claude writes the drafts, you read them and hit send.
If it hasn't saved you 5 hours in your first 30 days, you get your first month back.
Business owners only. $147 a month until we hit 50 members.
Talk tomorrow,
Zephyr
P.S. Run the audit and hit reply with the worst line it found in yours. I want to see it.