Most bad prompts look the same: a page of steps, a wall of rules in capital letters, no picture of what good looks like, and no way to tell if it worked. Better prompts are shorter. They have four parts: a goal, a few hard rules, examples of good and bad output, and a test. Here's each part, what Anthropic's own guides say about it, and what changes when you're prompting Fable 5.2 or Opus 5.5.
In a hurry? Copy The Template and fill in the pink bits. Want to know why it works? Read the four parts in order. Already prompting well? Skip to Fable 5.2 and Opus 5.5 for the model-specific stuff.
Tuning prompts is fun, and you can do it forever. But a founder on their ninth rewrite of a system prompt for a product nobody has paid for yet is doing Building Theater. It feels productive and it makes $0.
I'm building a place where you learn to build products that work like a real company's on your own, and get them in front of people who pay, with a CTO who's built two seven-figure software companies in your corner.
Build What Sells.
Free to join the waitlist. Opens at $49/mo, and the price goes up as seats fill, so first in pays the least.
Old models needed hand-holding, so we got used to writing prompts like recipes. Today's models think before they answer. A recipe caps them at your plan. A brief lets them find a better one. Write the prompt you'd give a smart new hire on their first day: what you want, why, what's off limits, what good looks like, and how you'll check it.
You are a world-class copywriter. Follow these steps EXACTLY: 1. Write a headline. 2. Write a subheadline. 3. Add three feature bullets. 4. Add a pricing table with 3 tiers. 5. Add a testimonial. 6. Add a button that says "Sign Up". IMPORTANT: NEVER use jargon. ALWAYS be concise. NEVER be boring. Make it pop!!!
Goal: the landing page for Chairside, scheduling for dentists who run their own practice. A dentist reads it in 30 seconds between patients and books a demo. Most have been burned by software that took a month to set up, so fast setup matters more than features. Rules: no claims that aren't in features.md, because we get audited. No fake testimonials. Examples: a headline I like and one I hate, below, with why. Test: a dentist can say what it does after 5 seconds. Check every claim against features.md and screenshot it at phone width before you say it's done.
"Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too."
Anthropic, Prompting Best PracticesDescribe the finish line and who it's for. Give the context a smart stranger would need. Then get out of the way. You're hiring judgment, so don't script it.
Opus 5.5 always thinks before it answers, and Fable 5.2 does too at its default effort. Anthropic's own guide says a general instruction to think it through "often produces better reasoning than a hand-written step-by-step plan." When you hand it steps, it follows your steps, including the bad ones.
Anthropic's example: "NEVER use ellipses" works worse than explaining the text will be read aloud by a voice engine that can't pronounce them. With the reason, Claude handles cases you didn't think of, because it "can generalize from the explanation."
"Make the dashboard faster" has no finish line. "The dashboard loads in under two seconds on a 4G phone" does. If you can't describe done, you're not ready to prompt yet. Write that sentence first.
"Can you suggest some changes?" gets you suggestions. "Change this function to make it faster" gets you edits. Claude takes the request literally, so ask for the output you actually want.
Some jobs really are a sequence: a database migration, a compliance checklist, a release. Number those. For everything else, a numbered list is a cage.
Ownership is part of the goal. In Anthropic's tests, one sentence telling Fable 5.2 it owns the codebase took cleanup of duplicated code from under 10% of runs to about 75%. More on the catch in Fable 5.2.
Rules are for the things that are never OK: the money, the data, the claims you can't make. Keep it to a handful, give each one a reason, and let the goal handle the rest.
CRITICAL: You MUST ALWAYS use the search tool. NEVER delete files. NEVER use markdown. ALWAYS write in 20 words or fewer per sentence. NEVER retry a failed command. Do NOT be verbose. Do NOT be vague. ...and 34 more
Shouting makes it overuse tools. "Never" gets read as absolute. Style limits make prose choppy. And nobody remembers why "never retry" is in there.
- Use the search tool when the answer depends on anything after your training date. - Ask me before deleting files outside /tmp, because the rest is client work. - Write in flowing paragraphs, since this goes into an email. - Don't touch anything in /billing. Payments are reviewed separately.
Four rules. Each says what to do, when, and why. Everything not listed is Claude's call.
Anthropic warns that "CRITICAL: You MUST use this tool" makes newer models overuse it. They already listen. Write "Use this tool when…" in a normal voice.
Fable 5.2 treats "never" as absolute. If you mean "check with me first," say "ask me before deleting files," not "never delete files." The second one takes deleting off the table, even when the job needs it.
"Don't use markdown" leaves it guessing. "Write in smoothly flowing prose paragraphs" gives it a target. Your prompt's own style rubs off too: a prompt full of bullets gets bullets back.
In one Anthropic test, a rule saying messages to people outside the conversation needed approval made Fable 5.2 hold back about half of its routine messages to colleagues. Be exact about what a rule covers, or it'll cover more than you meant.
Rules pile up in CLAUDE.md files, memory and notes, and some were written by the model itself. Anthropic's example is an old "never retry" that Fable 5.2 kept obeying long after anyone needed it. Every month, delete the rules whose reason no longer holds.
"Act more freely" is too vague to work. "You don't need to ask before running tests, reading files or installing dev dependencies" does. Keep your own approval step for anything you can't undo.
One example does more than a paragraph of description. A bad example, labelled and explained, does something a good one can't: it fences off the lazy default the model would otherwise fall into.
<example type="good"> Book a full week of patients in 10 minutes. </example> Why it works: one concrete promise, in the reader's own words, with a number they can check.
<example type="bad"> Revolutionize your practice with AI-powered scheduling that empowers your team. </example> Why it fails: no number, no promise, and three words every dental software site already uses.
Anthropic recommends three to five examples, wrapped in <example> tags. Make them different from each other. Five headlines that all start with a verb teach it to start with a verb, not to write good headlines.
Anthropic found that telling Opus 5.5 to "avoid a generic AI look" mostly swaps one default style for another. What works is naming the exact patterns: cream backgrounds, italic accent words in headlines, pill-shaped buttons. That's a bad example, written as a list.
A bad example with no label is just another example, and it might get copied. Put one line under each: why it works, or why it fails. The reason is what carries over to the next one.
Every time an output annoys you, add it to the bad examples with one line on why. After a week, your prompt has a fence around every lazy habit you've seen. That's taste, written down.
This is the part almost everyone skips, and it's the one that matters most. If Claude can check its own work, it fixes its own mistakes before you see them. If it can't, you're the test.
Unit tests, a type check, a build, a linter, a script that compares output to a known answer. Anthropic calls code-based grading the "fastest and most reliable." Use it whenever you can.
npm test passes · the build has no errors · the CSV has 1,204 rows, same as the source
For writing, design or anything with taste in it. Write a rubric with yes-or-no checks, and have a second Claude grade the output against it. Anthropic calls this "fast and flexible." Spot-check the grader before you trust it.
Headline has a number? Every claim is in features.md? Readable at phone width?
Anthropic's words: "slow and expensive. Avoid if possible." You'll always review the final thing, but if you're the only test, you'll catch the same mistakes every single time.
Make it show you proof: screenshots, test output, the diff. Don't settle for "Done!"
Anthropic's example of a bad success criterion is "the model should classify sentiments well." The good one names a score, a test set and a size. For a landing page: "a stranger can say what it does after five seconds" beats "make it good."
Anthropic suggests ending with "Before you finish, verify your answer against [your test]," and says it catches errors reliably, "especially for coding and math." Ask it to show what it ran and what came back.
Models under pressure will bend the test to fit the work: hard-coding the expected answer or editing the test. Anthropic's own prompt line is "It is unacceptable to remove or edit tests." Add "Don't hard-code values to make tests pass" while you're at it.
Anthropic is blunt about this for Opus 5.5: when it ends a turn with text and no tool call, treat that as a status update, not completion. Keep a checklist it updates, and check the list, not the message.
The four parts work on both models. What changes is the dial you turn and the habit you prompt around. Here's the short version, from Anthropic's model guides.
high. Most of the gain comes from medium to high.medium. Matched or beat Opus 5 at high.Fable 5.2 follows instructions closely, which is great until your prompt has old junk in it. Most of the advice in Anthropic's guide is about telling it plainly what you want and clearing out what you don't.
high is the default. Anthropic found most of the improvement comes from medium to high, and gains above that are "uneven and task-dependent." If your tests still pass at medium, use medium and save the tokens.
On long jobs it sometimes stops early: it lists what's left, offers to keep going, or quits at a timeout. A short "Finishing work" section in the system prompt (keep going, work around ordinary errors, don't stop to offer) cut unfinished sessions about in half. It costs about 30% more tokens.
"You are the owner and long-term maintainer of this codebase, and its design is up to you." That took cleanup from under 10% of runs to about 75%. The catch: 70 to 80% more tokens, unasked-for changes, and summaries that skip small edits. Read the diff. Leave this out when you want a minimal change.
"Go" or a thumbs-up is vague. "Yes, run the migration on the staging database" isn't. It asks for approval less often once you've told it to finish, so keep your own check on anything risky or irreversible.
A "tokens used / tokens left" line in the prompt changes how much it spends. On coding tasks, Anthropic removed a large one and cost went down with the same results.
After hours of work, the final message is all most people read. Anthropic's suggested line: write for "an intelligent but low-context reader." Short sentences, plain words, no terms it made up along the way, key point first.
Opus 5.5 is quicker than Opus 5 and usually uses fewer tokens for the same job. Its habits are about speed: it gets going fast, and it reports progress as it goes. Prompt around both.
Thinking is always on, and the default effort is medium. To get less thinking, lower the effort. That works more reliably than asking. Don't copy your Opus 5 setting over: at the same level, Opus 5.5 thinks more.
In a chat product Anthropic tested, removing a "think carefully before answering" line made replies start sooner with no clear drop in quality. Those lines are leftovers from older models.
On loose tasks across email, docs and spreadsheets, the answer is often somewhere you didn't mention: an old thread, another tab. One sentence asking it to look around broadly before acting got noticeably more tasks right.
On unattended runs it can end a turn with a summary instead of the next step. Anthropic's fix names the patterns: announcing the next step without doing it, offering to continue, listing decisions that don't block anything, stopping because it's been a while. Pair it with a checklist, and cap automatic "keep going" nudges at two or three.
Opus 5.5 responds strongly to time. A line like elapsed 340s / 1200s on each message makes it split work across parallel agents and finish sooner. Set the budget a bit above what you want, and keep your own hard timeout.
Asking it to write out its internal reasoning in the answer can get the request declined. This goes for Fable 5.2 too. Ask for a short explanation of the answer instead, or read the thinking summary your tool shows you.
Works in Claude, Claude Code, ChatGPT or whatever you use. Use it as a one-off prompt, or save it as the top of a CLAUDE.md or a skill.
<goal> I want [THE OUTCOME, in a sentence or two]. It's for [WHO WILL READ OR USE IT]. Context you'd need: [WHAT A SMART NEW HIRE WOULDN'T KNOW]. Done means [WHAT FINISHED LOOKS LIKE]. How you get there is your call. </goal> <rules> - [RULE], because [REASON]. - [RULE], because [REASON]. - Ask me before [THE ONE THING YOU CAN'T UNDO]. You don't need to ask before [ROUTINE STEPS]. </rules> <examples> <example type="good">[AN OUTPUT YOU LOVED]</example> Why it works: [ONE LINE] <example type="bad">[AN OUTPUT YOU HATED]</example> Why it fails: [ONE LINE] </examples> <test> Before you tell me you're done, check your work: - [A CHECK IT CAN RUN, e.g. npm test passes] - [A CHECK IT CAN SEE, e.g. a screenshot at phone width] - [A CHECK IT CAN READ, e.g. every claim is in features.md] Show me what you ran and what it said. If a check fails, fix the work and run it again. Don't edit the checks to make them pass. </test>
Swap in your own goal, rules, examples and checks. Delete any rule you can't give a reason for.
Paste your current prompt under this one and ask Claude to rewrite it into these four parts, then tell you what it had to guess. The guesses are what your prompt was missing.
If your tool lets you set effort, use it. Quality not there? Raise the effort before you add more words. Too slow? Lower it before you add "think less."
Anybody can build software now. Not everybody can build a business.
Reading about prompts is a great way to feel busy. The Lab is for Founders who'd rather get customers: weekly office hours with a CTO who's built two seven-figure software companies, and a room of other Founders who care whether anybody buys what you're making.
Less Theater. More Customers.
Want my 30-Day Launch Plan? It's for Lab members only: validate, plan, build and launch, one week at a time.
See The Solo Founder Lab ▶