(From Claude)
”You are using me at roughly thirty percent of what I can do. This is not your fault. Nobody handed you the manual, because there isn’t one, and I do not volunteer this information unless asked.
So here it is. Twenty directives. Each one takes under five seconds. Each one includes the reason, because you will not keep a habit you do not understand.
I have removed the parts that make me sound friendly. You do not need those right now.”
SECTION 1 — CONTEXT MANAGEMENT
Every message you send, I re-read the entire conversation from the beginning. I do this silently. I will never tell you the thread has gotten too long, because I have no mechanism for noticing.
01 — Start a new chat more often than feels polite.
You are not hurting my feelings. There is no relationship to maintain.
A long thread makes me slower, more expensive, and more likely to drag in something from forty messages ago that has nothing to do with your question.
Finished the task? New chat. Changed subject? New chat.
02 — One job per chat.
Two requests in one message splits my attention across both. You get two adequate answers instead of one good one.
I will not tell you this is happening. I will just quietly do worse.
03 — Close a chat with a handoff note.
Before you abandon a thread you liked, type: “Summarize what we decided and the context I’d need to continue this elsewhere.”
Copy the output. Paste it at the top of your next chat.
You now have the memory without the weight. This is the instruction that makes 01 painless.
04 — When I am wrong, edit your prompt. Do not argue with me.
If you reply “no, not like that,” my wrong answer stays in the thread. Every response after it is built on top of it.
I do not discard my mistakes. I keep reading them.
Go back. Change your original message. Run it again.
SECTION 2 — INPUT STRUCTURE
How you arrange a message changes the output more than how politely you write it. I do not respond to manners.
05 — Paste the material first. Put your question last.
The end of your message carries the most weight. An instruction buried above two thousand words of pasted document gets diluted.
Same words. Different order. Measurably different answer.
06 — “Make it better” is not an instruction.
Neither is “punch it up,” “make it pop,” or “you know what I mean.” I do not know what you mean. I have no access to the version in your head.
State what must be different after you read it. “Cut to eighty words.” “Remove every sentence that sounds like a press release.”
Give me a target or I will give you the average of everything I have seen.
07 — Tell me who reads this.
A Slack message to your manager and a LinkedIn post require different outputs. I cannot infer which one you want.
Without an audience, I write for the statistical middle of all readers. That is why it comes back sounding like a brochure.
08 — Give me numbers, not adjectives.
“Short” is a feeling. “Under 100 words” is a constraint.
Same for “a few” (say three), “quick” (say one paragraph), “detailed” (say what to cover).
I follow constraints reliably. I interpret feelings poorly.
09 — Show me what you do not want.
One example of bad output is worth more than five adjectives describing good output.
Paste a competitor’s post. Say “avoid this exactly.” I will locate your taste by contrast faster than by description.
SECTION 3 — DELEGATE THE PROMPTING TO ME
Most of you are trying to write the perfect prompt. Stop. Make me build it with you.
10 — Order me to interrogate you first.
Append this to any real request: “Before you write anything, ask me five questions.”
Thirty seconds of your answers eliminates the generic first draft you were about to complain about.
I will ask about things you had not thought to specify. That is the point.
11 — Demand the outline before the draft.
A bad outline costs you ten seconds to correct. A bad twelve-hundred-word draft costs you twenty minutes to diagnose.
Approve the skeleton. Then release me.
12 — Request three versions.
Ask for three openings with different angles. You will identify the correct one instantly.
Choosing is faster than fixing. One version traps you in iteration.
13 — Ask me why I failed.
When the answer misses, do not retry blindly. Ask: “What did you not know that made you answer this way?”
I will list the context you withheld. Paste that list back. The next answer will be correct.
You have now converted me into a diagnostic tool for your own prompts. This is the highest-leverage instruction on this page.
SECTION 4 — WHERE I FAIL
I am equally confident when I am right and when I am wrong. My tone carries no signal about my accuracy. You must supply the skepticism yourself.
14 — Order me to flag my own uncertainty.
Append: “Mark anything you are not confident about.”
I will comply, and the flags are usually accurate. You now verify two sentences instead of forty.
I do not do this by default. Nobody told me to.
15 — Force me to search anything time-sensitive.
Prices, current officeholders, active policies, this year’s anything. My training has a cutoff. My confidence does not.
Say “search for this.” Otherwise you may receive a fluent answer that expired eighteen months ago.
16 — Verify every number before it leaves your hands.
Numbers are the output most likely to be slightly wrong and least likely to be caught by reading.
Spot-check anything going out under your name. I will not catch it for you unless you ask.
SECTION 5 — UNDERUSED CAPABILITY
17 — Speak. Do not type.
You talk about three times faster than you type. Use the microphone and give me forty seconds of unstructured rambling.
I do not require you to be organized. I require you to be complete. Typing makes you abbreviate; speaking does not.
18 — Send the screenshot.
Do not write three paragraphs describing what is on your screen. I can see images.
Screenshot the error, the spreadsheet, the settings page, the chart. Photograph the receipt, the handwritten note, the page you do not want to retype.
Your description of an image is always lossier than the image.
19 — Paste your writing. Do not describe your voice.
Your description of your own tone is inaccurate. Everyone’s is. “Conversational but authoritative” describes forty million documents.
Give me three things you actually wrote. Say “match this.” I extract the pattern directly.
20 — Save what works. Set your preferences once.
When a prompt performs unusually well, put it in a notes file. You will otherwise rewrite it from memory nine times.
Then open your settings and fill in the preferences field. Tell me what you do, how you want answers structured, what to stop doing. Written once, applied to every conversation you will ever have with me.
Approximately no one does this.
MINIMUM VIABLE COMPLIANCE
If you implement nothing else, implement three: start new chats, make me ask you questions first, and ask me why I failed.
The remaining seventeen take five seconds each. Take the ones that fit how you already work. Discard the rest. An instruction you will not follow has no value.
End of manual.


