Lesson 19 of 32 · Part One, the principles
The Elimination of Mental Static
Cleanliness
A prompt is a signal, and typing in a hurry puts static on it. A misspelt word is read as an unfamiliar word, two names run together become one person, and a line with no punctuation can be read two ways - so the model guesses, and a guess is where a wrong answer comes from. Spell the names, places and numbers correctly, put one instruction on each line, and read it once before you send it.
How to use The Elimination of Mental Static in Claude
Tags do what punctuation does, only more bluntly: a list of names inside one tag cannot be read as a sentence, and a figure inside its own tag cannot drift into the sentence beside it. There is a second use that fits this chapter exactly - paste your own rough shorthand, tag it as notes, and ask for it back as a clean prompt with one instruction per line. You then read that, fix what it misread, and send it. Long input is no problem, so paste the whole list rather than a half-typed summary of it.
- One tag per kind of fact - names, dates, figures, format.
- Paste your shorthand and ask for it back as a clean prompt.
- Say what to do when unsure: ask me, do not guess.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Tags do what punctuation does, only more bluntly: a list of names inside one tag cannot be read as a sentence, and a figure inside its own tag cannot drift into the sentence beside it. There is a second use that fits this chapter exactly - paste your own rough shorthand, tag it as notes, and ask for it back as a clean prompt with one instruction per line. You then read that, fix what it misread, and send it. Long input is no problem, so paste the whole list rather than a half-typed summary of it.
From the book Paste your rough notes and ask Claude to "clean this into a clear prompt with one instruction per line", then use that cleaned prompt. Claude handles long documents well, so paste full lists instead of half-typed summaries.
- One tag per kind of fact - names, dates, figures, format.
- Paste your shorthand and ask for it back as a clean prompt.
- Say what to do when unsure: ask me, do not guess.
- Ask it to list the names it read before it answers.
Instead of
rite lettr to dist offce for printr repare n chairs urgnt
Claude version
<my_shorthand> rite lettr to dist offce for printr repare n chairs urgnt </my_shorthand> <what_i_actually_mean> Station House Officer, [station name], writing to [officer's designation], District Police Office, Nagpur One laser printer, dead since [date] Four chairs, wheels broken Wanted: repair or replacement inside 15 days </what_i_actually_mean> <task> Two steps, in this order. 1. Turn my shorthand into a clean brief, one instruction per line, using only the facts beneath it. Name any word in the shorthand you could not decipher rather than deciding what it probably said. 2. Then write the letter from that clean brief: formal, under 150 words, with blanks left for the date, the reference number and the signature. </task>
Then loop it
- Good. Show the two items as a numbered list inside the letter.
- Now give me the same letter in Marathi, with the same blanks.
Why it worksEvery key word is spelt correctly and each detail sits on its own line, so nothing is misread - "repare n chairs" could mean almost anything.
Instead of
make clening roster rahul aman kiran joseph dev 2 per day sunday off
Claude version
<students> Rahul Aman Kiran Joseph Dev </students> <note>Those are five separate students. None of those lines is one person with two names.</note> <days>Monday to Saturday. No duty on Sunday.</days> <rules>Two students a day. Nobody on duty two days running. Equal totals.</rules> <format>Table: Day, Student 1, Student 2. Below it, days on duty per student.</format> <task>Build the roster for a college hostel common room in Dehradun. Begin by listing the five names back to me, so I can see you read them as five.</task>
Then loop it
- Dev has exams on Wednesday and Thursday. Redo the roster without him on those days.
- Good. Now write a short notice for the board to go with this roster.
Why it worksSemicolons between names and one fact per line remove the static - "kiran joseph" can no longer be mistaken for one student.
Instead of
family trip meghalya 4 days senior citizens cheap pls asap
Claude version
<travellers>Two grandparents in their 70s. Two parents. One child aged 9.</travellers> <route>Guwahati to Shillong to Sohra (Cherrapunji) and back to Guwahati. Four days, [dates] in November.</route> <limits>Maximum four hours in the car in any one day. Steep paths and long walks are out.</limits> <budget>About Rs 60,000 for hotels and car hire.</budget> <format>Day 1 to Day 4, with morning, afternoon and evening under each day.</format> <task>Write the itinerary. Give the full spelling of every place, with the local name in brackets where it differs, because I copy these names straight into the family's email. Mark any timing you are not sure of rather than filling the gap.</task>
Then loop it
- Day 2 has too much travel. Move the caves to Day 3.
- Now write a short WhatsApp message to the family summarising the plan, with hotel names left as [hotel].
Why it worksCorrect place names and each requirement on its own line give the AI a clean signal with no crackle to guess through.
Questions
What is The Elimination of Mental Static in prompt engineering?
A prompt is a signal, and typing in a hurry puts static on it. A misspelt word is read as an unfamiliar word, two names run together become one person, and a line with no punctuation can be read two ways - so the model guesses, and a guess is where a wrong answer comes from. Spell the names, places and numbers correctly, put one instruction on each line, and read it once before you send it.
When should I use this technique?
When you are typing fast on a phone - a letter, a roster, a list of names and numbers - and the answer comes back subtly wrong: a merged name, a misread figure, a place you never meant.
How do I apply it in Claude?
Tags do what punctuation does, only more bluntly: a list of names inside one tag cannot be read as a sentence, and a figure inside its own tag cannot drift into the sentence beside it. There is a second use that fits this chapter exactly - paste your own rough shorthand, tag it as notes, and ask for it back as a clean prompt with one instruction per line. You then read that, fix what it misread, and send it. Long input is no problem, so paste the whole list rather than a half-typed summary of it.
What are the limits of this technique in Claude?
Claude is good at reading through a typo, which is the trap - it will often produce a confident, fluent answer built on its own correction of your word, without mentioning that it corrected anything. Fluency is not accuracy. Tell it plainly what to do when a word is unclear - ask me, do not guess - because left to itself it will choose the reading that makes the nicest sentence.
How do I improve the answer after the first try?
Before you press send, read your prompt aloud once, slowly. Fix spellings, especially names, places and numbers, and put each instruction on its own line. When the answer comes back, look for signs of static: a wrong name, a misread number, two people merged into one. If you see one, do not just say "wrong". Correct the exact word: "Kiran and Joseph are two different students." Then keep the clean version of your prompt to reuse, so the static never returns.
Loop it
Before you press send, read your prompt aloud once, slowly. Fix spellings, especially names, places and numbers, and put each instruction on its own line. When the answer comes back, look for signs of static: a wrong name, a misread number, two people merged into one. If you see one, do not just say "wrong". Correct the exact word: "Kiran and Joseph are two different students." Then keep the clean version of your prompt to reuse, so the static never returns.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.