For two years I did what most people do with ChatGPT: type a rough request, get a generic answer, retype it with more detail, get a longer generic answer.
The real work was never the writing. It was hand-tuning the wording until the output stopped sounding like a brochure.
So I built one prompt to do that tuning for me. I call it The Prompt Architect v4.
You give it five slots — task, audience, source material, constraints, success criteria — and it returns a reusable prompt in five blocks: Role, Context, Steps, Output format, Rules.
I tested it on a 350-word client meeting note, with a target of a one-page project brief. By hand that brief takes me 45 minutes to draft and format.
The master prompt produced the reusable prompt in 25 seconds, the generated prompt wrote the brief in 40 seconds, and I spent 12 minutes reviewing and editing.
The other 11 modules in this chain are optional repairs. You run them only when the first output misses. Here is the full chain, in the order the failures actually show up.
| Module | Use it when | What it returns |
|---|---|---|
| 1. Master prompt | Always start here | A 250–450-word reusable prompt in five blocks |
| 2. Task clarifier | The generated prompt is still vague about what done means | A one-paragraph task statement under 80 words |
| 3. Role and audience | The prompt says 'act as an expert' | Three ranked roles and the Role and Audience lines |
| 4. Source packager | Your notes are messy or full of numbers | Facts to use, facts to verify, open questions, background |
| 5. Step writer | Steps like 'write the brief' with no order | 3–6 numbered steps with inputs and outputs |
| 6. Output formatter | You get a wall of text or wrong headings | One output format block under 80 words |
| 7. Rule writer | You spot a number you cannot find in your source | 4–6 testable rules, including flag-uncertain-claims |
| 8. Prompt critic | You cannot tell what is weak | A 0–10 score table and one next module |
| 9. Test case | Before a client deliverable | One test input and expected output with [CHECK] flags |
| 10. Privacy scrubber | Before pasting client material | Scrubbed text and a removal list |
| 11. Compressor | The prompt has grown past 450 words | The same prompt under 300 words, five blocks intact |
| 12. Iterator | The output is close but still wrong | One rewritten block and one proof test |
1. The master prompt that writes prompts
The master prompt: The Prompt Architect v4 is one long instruction I paste into a fresh ChatGPT chat. It takes five things from me: the task, the audience, the source material, the constraints, and what counts as done.
In my run, the task was to turn a 350-word client meeting note into a one-page project brief. The constraints were one page and the client's tone. Done meant the client signs off without a follow-up call.
What comes back: A reusable prompt of 250 to 450 words in five blocks: Role, Context, Steps, Output format, Rules. It lands in about 25 seconds in one code block, ready to paste. That prompt then wrote the brief in 40 seconds.
The check: It cannot verify names, dates or numbers from my note, so I read the brief against the note line by line. Review took 12 minutes, against 45 by hand. Modules 2 to 12 are optional repairs, run only when the output misses.
💡 Tip: Save the generated prompt in a note called 'Prompts I actually use'. You will run it dozens of times before you need to change a word of it.
Five prompt terms, translated
2. The task clarifier prompt
Answer what done means before you touch the role or the rules. A reusable prompt can be well built and still aim at a fuzzy target.
Prompt to copy
You are a task editor. My rough request is: [ROUGH REQUEST]. Rewrite it as a precise task with four parts: deliverable, audience, source material, and what done looks like. Do not add facts. If a part is missing, write [NEEDS INPUT]. Return one paragraph under 80 words, then a three-item checklist of what I still need to supply.
By hand, pinning down what done means took me about 10 minutes of rewriting and second-guessing. ChatGPT returns the tightened task and a three-item supply list in roughly 20 seconds. My check of both takes another 2 minutes.
Two times: about 12 minutes by hand against about 20 seconds plus a 2-minute check.
My run turned "make a brief" into a one-page brief for the client approver, done when they approve without a follow-up call.
It marks the missing audience [NEEDS INPUT] rather than inventing one.
⚠️ COMMON MISTAKE
People skip the slots, then blame the prompt
The master prompt only works if you actually fill [TASK], [AUDIENCE], [SOURCE MATERIAL], [CONSTRAINTS] and [SUCCESS CRITERIA]. Leave two blank and you get a generic reusable prompt that produces a generic brief. Fill all five, even roughly, before you hit enter.
3. The role and audience prompt
You know the generic answer as soon as it arrives. It opens with "in today's business environment" and reads like any blog post, because the prompt told the AI to act as an expert and never said who was reading.
Even a sharp task definition leaves the role flat and the audience unnamed. The role and audience module names a specific job title and the one thing that reader checks first.
Prompt to copy
You are a role designer. For this task: [TASK]. Name the single best role for the AI to play, the audience it writes for, and the one thing that audience will check first. Give me three role options, ranked, with one sentence on why each fits. Then write the Role and Audience lines I should paste into my prompt. Do not use generic roles like expert; name the job.
In my run it returned three ranked roles in about 20 seconds. The winner was a project operations lead writing for a client sponsor, whose first check is scope and dates.
The generic expert line was gone, and the brief opened on schedule and deliverables instead of pleasantries.
The catch: a confident, specific role can make the AI sound more assured than its source supports. The role line does not replace the meeting note or the rule that flags unverifiable claims.
Read every date and figure against the note, because the role tells the AI who to sound like, not what is true.
What you get is three ranked role options and the exact Role and Audience lines in about 20 seconds, ready to paste into the reusable prompt.
That one swap is what turns a brief from a generic summary into a document a client sponsor reads without asking who it was written for.
What the chain feels like over a month
Times are for one 350-word meeting note going to a one-page brief.
4. The source packager prompt
Prompt to copy
You are a context editor. Here is my source material: [PASTE NOTES, DATA OR QUOTES]. Organize it into four labeled sections: facts I can use, facts I must verify, open questions, and background only. Keep every name, date and number exactly as written. Do not summarize away a number. Return the four sections as a bulleted list, then a 50-word context block I can paste into my prompt.
Run it when the file behind your role lines is a rambling transcript, a long PDF, or notes with numbers you cannot afford to lose.
On my 350-word client note it cost one paste and about 30 seconds. Claude handled a long PDF best, because it reads the whole document before sorting.
What came back was four buckets: rows I could trust, rows to verify, questions still open, and background.
Compare those rows with the original before you paste, because a misread date is easy to miss.
Names and dates stayed exact, open questions left the usable facts, and the 50-word block became my Context block.
The Prompt Architect v4 — At a Glance
📝 Blocks
5: Role, Context, Steps, Output format, Rules
📏 Word count
250–450 words
⏱️ Generation time
About 25 seconds
🔁 Optional modules
11 repairs, run only on a miss
🧪 Example task
350-word note to one-page brief
The whole chain
One master prompt, eleven repairs, one brief — the modules exist to fix specific failures, not to be run every time.
5. The step writer prompt
Sorted source still fails if the steps say "write the brief" with no order and no check. Steps are also where a prompt quietly merges writing with checking, so nothing verifies the output against the note.
Prompt to copy
You are a process writer. For this task: [TASK] and this source material: [SOURCE]. Write the numbered steps the AI should follow. Each step starts with a verb and names the input it uses and the output it produces. Give 3-6 steps. Do not merge writing and checking into one step. End with the step where the AI must stop and ask for missing input.
More steps do not mean a better prompt, so cut any that does not change the deliverable.
On my note it separated writing from checking, and the list ended with a stop-and-ask point.
About 25 seconds in ChatGPT for 3-6 steps, each naming its input and output.
Three places the chain goes wrong
6. The output formatter prompt
What came back was the one-page project brief with fixed headings in a fixed order and a one-page ceiling, no table, and a flagged line for any fact the note did not settle.
The client tone stayed in Constraints, where it already lived. Before this block, the steps reached the brief but handed back four pages of run-on paragraphs.
Prompt to copy
You are an output designer. For this deliverable: [DELIVERABLE]. Specify the exact output format: headings, order, word count, table columns if any, and what to do with uncertain facts. Give me one format only, not options. Then write the Output format block I should paste into my prompt, under 80 words.
You run it once. ChatGPT returns the block in about 20 seconds, and from then on every run of your prompt lands in the same shape, so you stop reformatting by hand.
The block settles the headings, their order, the word budget and how unsure facts get marked.
A tight format can also hide thin reasoning: a tidy brief reads finished when the thinking is missing. Keep the flag-for-uncertain-facts line inside the block, so a gap shows up as a gap instead of a clean sentence that quietly guesses.
7. The rule writer prompt
Three of the 350 words came back as numbers I could not find anywhere in the note. That is the rule writer's job, and it is the one place where AI should stop.
If the client name, the fee or the deadline is not in your source, write that rule yourself. The model cannot tell you what it does not know.
Prompt to copy
You are a guardrail writer. For this prompt: [PASTE PROMPT]. Write 4-6 rules that prevent the most likely failure: invented facts, ignored source material, bloated length, or wrong tone. Each rule must be testable. Include one rule that requires the AI to flag any claim it cannot verify from the source material, and one rule that tells it what to do when the source is silent. Return the Rules block only, under 100 words.
I pasted it into ChatGPT and the block came back in 25 seconds. Two rules were new to my brief. One says flag any claim you cannot check against the note.
The other says what to do when the note is silent. Neither can promise the model obeys, so I still read every figure against the note.
8. The prompt critic prompt
You can read a prompt top to bottom and still not see which block is doing the damage. That is the tedious part. So I stopped guessing and used a score.
I pasted the whole prompt into ChatGPT and asked for marks across five areas. The job, the context, the order of the steps, the format and the rules each got 0 to 2. The marks came back in 30 seconds.
Prompt to copy
You are a prompt critic. Score this prompt: [PASTE PROMPT] on five checks, 0-2 each: task clarity, context sufficiency, step order, output format, and rules. For each check, quote the line that earns or loses the point. Total the score out of 10. If the total is below 7, name the one module I should run next from this list: clarifier, role, source, steps, format, rules, test, privacy, compressor, iterator. Return the score table, then the single next action.
My prompt scored 6 out of 10. The critic quoted the line that lost the point on rules and told me to go to the rule writer first.
That rule block changed the final brief. The score is a checklist, not a truth test, and it cannot predict an invented number.
With five blocks in place, the critic is the cheap next move. It costs half a minute and a table.
I stopped arguing with myself about wording and followed the one action it named. The fix was faster than the worry. Then only the low mark needed work, not the whole prompt.
Before the chain vs. after the chain
Before: retyping requests
- Retype the same request with more detail every time
- Each chat starts from zero context
- You cannot tell which block is weak
- Fluent answers that quietly invent a number
- By hand, the brief takes 45 minutes
After: a saved reusable prompt
- Five slots filled once, reused for every client
- Every prompt carries Role, Context, Steps, Output format, Rules
- The critic scores it out of 10 and names the fix
- The rule block flags what the source does not support
- Master prompt in 25 seconds, brief in 40, edits in 12 minutes
9. The test case prompt
Prompt to copy
You are a test designer. For this prompt: [PASTE PROMPT]. Write one realistic test input under 120 words, then the output the prompt should produce in under 200 words. Mark any place where the prompt would need a fact you did not give it with [CHECK]. Do not make the test easy; include one ambiguous instruction and one missing number.
That is the drill I run before client work. A score of 7 out of 10 still says nothing about behavior.
So I feed the prompt a made-up brief with one vague line and one number left out, and see what breaks before a client does.
The model answers in about 30 seconds; my read takes two minutes.
On my 350-word note run, the drill exposed an ambiguous instruction and a missing number. Both went into the prompt as [CHECK] before any client material touched it. A test I write alone can share my blind spots; a real, non-sensitive example catches more.
10. The privacy scrubber prompt
Two placeholders in twenty seconds: that is what the scrubber did to my 350-word client note. It replaced the client name with [CLIENT] and a fee figure with [AMOUNT]. The paragraph order and the sense of the note survived intact.
Run it before the note, the spreadsheet or the client file goes into any free or personal chat plan. It swaps passwords, personal names, account numbers, health details, money figures and private web addresses for bracketed tags. Then it lists what it removed.
Cost: about 20 seconds and no money, on the free ChatGPT plan.
Prompt to copy
You are a privacy editor. Here is text I plan to paste into a chatbot: [PASTE TEXT]. Replace every password, client name, account number, medical detail, financial figure, and private URL with a bracketed placeholder like [CLIENT], [ACCOUNT], [AMOUNT]. Keep the structure and meaning. Return the scrubbed text, then a list of what you removed. If you are unsure whether something is sensitive, replace it.
11. The compressor prompt
A prompt past 450 words starts losing the middle. When I stacked a longer role line and more rules onto mine, the model stopped honoring instructions that sat in the center. The words were still on screen, but the answer ignored them.
Do not paste that bloated prompt and hope. Run the compressor first, because a shorter prompt with all five blocks beats a longer one missing its guardrails. Then read what survived.
Trimming removes protection along with padding. My fact-check line and my stop-and-ask step were the two longest items, and both had to stay even after the cut. Check every rule against the source before you paste it into a fresh chat.
Cut to under 300 words with Role, Context, Steps, Output format and Rules still present in about 25 seconds. You keep the safety line and lose the noise.
Prompt to copy
You are a prompt editor. Cut this prompt: [PASTE PROMPT] to under 300 words. Keep the five blocks: Role, Context, Steps, Output format, Rules. Remove repeated instructions, hedging, and examples that do not change the output. Do not remove a rule that prevents invented facts or a step that checks the source. Return the shortened prompt in one code block and a one-line note on what you cut.
🪜 The order to run the chain on a real client note
Fill the five slots
Write out your task, audience, source, constraints and success criteria before opening ChatGPT.
Run module 1
The master prompt returns a reusable prompt in five blocks in about 25 seconds.
Score it with module 8
The critic gives a 0–10 total and names the module to run next if the score is below 7.
Run the repair modules
Run modules 2–7 and 9–12 only for the gaps the critic or the review surfaces.
Scrub private material with module 10
Replace client names, accounts and figures with placeholders before anything goes into the chat.
Test with module 9
One test input and expected output with [CHECK] flags before a real client note goes through.
Review line by line
Check every name, date and number against the original note. This is the 12-minute step.
12. The iterator prompt
The iterator rewrote one block and returned one test input, about 30 seconds of work. On my run the output was close but broke the one-page limit, so it touched only the Output format block and left Role, Context, Steps and Rules alone.
Prompt to copy
You are a prompt improver. Here is my prompt: [PASTE PROMPT] and here is the output it produced: [PASTE OUTPUT]. Name the single biggest gap between what I asked for and what I got. Then rewrite only the block that causes the gap: Role, Context, Steps, Output format, or Rules. Return the gap in one sentence, the rewritten block, and one test input that would prove the fix.
The trigger is an output that misses the audience, the length or one fact you needed. One gap at a time, because rewriting every block at once hides which change fixed it. Best tool: ChatGPT.
Frequently Asked Questions
Do I need to pay for ChatGPT to use this chain?
No. The master prompt and every repair module work on the free tier, which has usage limits that change. Paid plans raise the limits and may include features like web browsing and larger file uploads. Start on free and upgrade only if you hit a limit mid-task.
Do I have to run all 12 modules every time?
No, and you should not. Module 1 is the only one you run every time. Modules 2–12 are repairs — you run each one when a specific symptom shows up in the output. Running the whole chain on a simple task wastes about 15 minutes.
Will this stop ChatGPT from making things up?
No. The rule writer adds a flag-uncertain-claims instruction and a source-silent rule, which helps, but every fact, figure, quote and link still has to be checked against your source. The 12-minute review is the part that actually catches errors.
What if my source material is a long PDF rather than a short note?
Run module 4 first and paste the PDF or its text into Claude or ChatGPT. Claude handles long documents well. The module returns a 50-word context block you can then carry into the master prompt, which keeps the final prompt short enough to work.
Can I run this chain outside ChatGPT?
Yes. The prompts use plain language and work in Claude and Gemini without changes. Some people prefer Claude for source-heavy work and ChatGPT for the reusable prompt itself. Pick one and stay with it so you learn how it responds.
Which module to paste first, and which one saves your client file
Run module 1 first. Score the prompt it gives you with module 8.
If the score is below 7, send yourself to module 2 when the prompt is vague about what done looks like, or module 11 when the prompt has grown past 450 words and the AI has started ignoring the middle.
Before any client material goes into a chat, run module 10. Never paste passwords, client files, or medical or financial records, scrubbed or not.