FMT_AS_HTML // FIELD MANUAL
Vol. 04 · Meta-Prompting
SIG: 0x4M-2026-05
2026 · 05 · 10
CATALOG
Compare / Explain
SUBJECT
Meta-prompting moves
PRIMARY SOURCE
Su, J. — 4 techniques
FIFTH MOVE
FILED AS ASIDE →

A field manual for FOUR
MOVES
/ + ONE.

▶ How to read this

Jeff Su's note lists four meta-prompting techniques worth keeping. Each one is a small mechanical move that compounds against ordinary prompting. They are taken straight, in order, with the worked examples and the pasteable prompts.

A fifth move sits at the back — reverse delegation — often shelved alongside #01 and confused with it. Filed separately so the difference is visible.

0101
POST-EXECUTION · ACTS ON THE CONVERSATION · YIELDS A REUSABLE PROMPT

Prompt Reversal.

Iterate until the output is right. Then ask the model to write the single prompt that would have produced that final turn in one shot. The artifact you keep is a prompt — not a conversation.

PROBLEM
First prompt lands at ~50%. You refine to 60, then 80, then 90. Every recurring task lives this loop. The savings compound only if you capture what you discovered along the way.
MOVE
After you've landed the result, send: "Reverse engineer our conversation and write the single prompt that would have produced this final response in one go." Paste the output into a new chat to verify.
WHY IT WORKS
Captures the implicit refinements you couldn't have specified upfront. Builds a private prompt library. Teaches you what optimized prompts look like, structurally.

"Analyze a competitor and walk through their business strategy."

→ 01 First turn. Generic strategic overview. Too dense to use. → 02 Course-correct. "Restructure as SWOT. 3 bullets per quadrant, plain words." → 03 Refine again. "Flesh out each bullet. Add a sub-heading Our Strategic Response with one concrete action per quadrant." → 04 Reverse. "Reverse engineer our conversation and write the single prompt that would have produced this in one go." → model returns a copy-pasteable block. → 05 Save it. This is what belongs in your prompt library. It's the prompt you wouldn't have written from cold.
▌ THE REVERSAL LINE paste after the final good turn
Reverse engineer our conversation and write the single prompt that would have produced this final response in one go. Return it inside a fenced code block so I can copy it cleanly.
0202
POST-EXECUTION · ACTS ON A SOURCE ASSET · YIELDS DERIVED FORMATS

The 5-min Amplifier.

One strong source — a deck, a transcript, a report — becomes a quiz, a recap email, an infographic, a LinkedIn post, a cold-outreach sequence. Pure transformation, near-zero re-creation.

PROBLEM
You produced one strong artifact and then need five. Reformatting by hand is the kind of work AI was built to absorb, but only if the source is worth amplifying.
MOVE
Hand the model one piece of "pillar content" and request derivative formats one at a time, each with its own audience and constraints baked in.
CAUTION
Garbage in, garbage out. The amplifier magnifies whatever you feed it. Only amplify proven work: a deck that landed, a report with real data, a transcript of a meeting that mattered.

A product team's slide deck for the Q4 launch.

→ A Quiz. "10 multiple-choice questions for an audience knowledge-check. Indicate correct answers." → B Internal recap email. For stakeholders who couldn't attend. Top takeaways and product updates only. → C Client infographic. Pull the 5 most impactful stats. Plain language, no jargon. Suggest a visual layout. → D Sales cold-email sequence. Three emails, escalating specificity. Talking points for the discovery call attached. → E HR variant. Same move on a one-hour webinar transcript → quick-reference guide, intranet FAQ, knowledge-check quiz.
▌ AMPLIFIER PROMPT — TEMPLATE attach the source asset first
You have the attached [source asset] as context. Treat it as the pillar content. I need to amplify it into the following derived formats, one at a time: [1] [format] — for [audience], optimizing for [goal] [2] [format] — for [audience], optimizing for [goal] [3] [format] — for [audience], optimizing for [goal] Start with [1]. Match the tone of the source. Don't add facts that aren't in it. When I say "next", produce [2].
0303
POST-EXECUTION · ACTS ON THE OUTPUT · YIELDS ADVERSARIAL CRITIQUE

Red Team.

Right after the model helps you build something, make it put on the hat of the person you're trying to convince and tell you why your draft fails.

PROBLEM
The model is biased toward whoever's asking. You wrote it with the model's help, so the model thinks it's good. Your audience doesn't.
MOVE
After producing the artifact, flip the persona explicitly. The new persona must be the actual recipient, with a clear motivation and a constrained attention span (60 seconds, 50 inboxes, cost-cut mandate).
PRO TIP
Vague critics produce vague critiques. "Act as a critic" is useless. "Risk-averse CTO whose primary concern is data security and audit trails" is useful. Specify motivation, not just role.

Three drafts; three different red-team personas.

A Tailored resume → "Act as the hiring manager. You have 60 seconds. What are your immediate red flags?" B Business proposal → "Act as our CFO. Your job is to cut unnecessary cost. Critique the proposal. What's the biggest financial risk? Why isn't the ROI justified?" C Cold outreach → "Act as the VP of Marketing. 50 cold emails a day. Read mine. Which specific sentences make you hit delete, and why?"
▌ RED TEAM PROMPT — TEMPLATE paste after the artifact
You are no longer my assistant. You are now [specific persona], whose primary motivation is [clear, constrained goal], and who has [time constraint] to evaluate what's in front of them. Read the [artifact] you just helped me write. Tell me, unfiltered: • Your immediate gut reaction in the first 10 seconds • The single sentence or section that would make you walk away • The strongest objection a reasonable peer would raise • What you would have written instead Don't soften it. Don't hedge. Don't list "things to consider".
0404
PRE-EXECUTION · ACTS ON THE PROMPT · YIELDS A REVIEWABLE PLAN

Blueprint Scaffolding.

Before the model produces the deliverable, force it to outline the structure — the sections it intends, with one-line descriptions. Review the blueprint. Cut what's irrelevant. Then build.

PROBLEM
Direct prompts produce kitchen-sink output — every section the model thinks might belong. You spend the next twenty minutes deleting Tracking & Measurement, Risks & Mitigation, Operations & Roles.
MOVE
Append "First, outline the standard sections of a professional [thing], one sentence per section. Wait for my edits before producing." Review like an architect's blueprint.
WHY IT WORKS
Spotting a wrong measurement on a blueprint costs nothing. Spotting it after the concrete is poured costs the project. Articulating steps also forces the model down a more careful reasoning path.

"I run an online course; I need a marketing campaign brief for the Q4 holiday push."

x Without scaffolding: a 1,200-word brief covering objectives, audience, channels, creative, budget, tracking, risks, ops — most of which you don't need right now. → With scaffolding: first, a blueprint — 8 section titles, one sentence each. !! Your edit: "Too much. Apply the 80/20 — only the sections essential for a 3-email holiday sequence." ✓ Then build. The model writes only what survived the cut.
▌ BLUEPRINT PROMPT — TEMPLATE append to a complex request
[Your normal request here.] Before producing the deliverable, first outline the structure you intend to use. Give me: • The section titles, in order • One sentence describing the purpose of each section • An estimate of the length you'd give each section Stop after the outline. Wait for my edits. Once I confirm which sections to keep, cut, or expand, produce the full deliverable using exactly that structure.
FILED · APART
An aside · the fifth move · often confused with #01

Not reversal.
Delegation.

A different move sits next to Prompt Reversal in the wild and gets called by the same name. It deserves its own page. Prompt Reversal distills a finished conversation into one prompt. Reverse Delegation briefs the model on you and asks it to enumerate the work it could be doing in your life. Same surface (the model asks you for something). Different physics.

WHEN IT FIRES
After the result is good. Before you know what to ask for.
WHAT IT ACTS ON
The conversation. The operator behind the prompt.
WHAT YOU KEEP
A reusable prompt. A delegation map.
▌ STEP 1 · THE INTERVIEW run after a context-rich brain dump
Based on everything I have shared about myself, my career, goals, responsibilities, constraints, and ambitions, act as a strategic AI operator. First, identify the most important missing information you would need from me to help me achieve my goals faster and reduce the amount of work I personally have to do. Group your questions into categories: 1. Career and income goals 2. Current responsibilities 3. Bottlenecks and recurring tasks 4. Personal strengths and weaknesses 5. Tools, systems, and workflows I already use 6. Opportunities I may be underusing 7. Decisions I am avoiding or delaying For each question, explain briefly why the answer would help you support me better. Then, after asking those questions, give me a preliminary list of tasks you could do for me right now based only on the context you already have.
▌ STEP 2 · THE DELEGATION MAP run after you answer the interview
Using my answers above, create an AI delegation map. Organize it into: 1. Tasks you can do for me immediately 2. Tasks you can help me draft, plan, or prepare 3. Tasks you can turn into templates, checklists, or systems 4. Recurring workflows you can help me automate or simplify 5. High-leverage opportunities I am not currently acting on For each item, include: • The specific task • What you need from me to complete it • The expected benefit • The first action we should take

The move from "ask me questions" to "build an operating system around me" is the upgrade. The viral version of this prompt — "what should I tell you to help me achieve my goals?" — gives the model permission to wander. The structured version pins seven categories so it can't skip the hard ones, then turns the answers into a delegation map instead of a chat thread.

// COLOPHON

F.M. 04

Set in Big Shoulders Display and Big Shoulders Stencil Display (Patric King, Sorts Mill — display), Manrope (body), Fraunces (italic accents), JetBrains Mono (technical marks). One self-contained HTML file. Inline everything. Designed to be saved to disk.

// SOURCE

Su, J. The Prompt Reversal Method. Uploaded note, attributed to jeffsu.org via embedded image domains. Four techniques reproduced and re-cased.

// ADDITION

Reverse Delegation and the structured pasteable prompts in the aside are the operator's reframe, in this thread. Filed apart so the difference from Prompt Reversal is visible.

// FORMAT

Filed under Compare / Explain. The page is built for double-click-to-open. Copy buttons hit the system clipboard. No analytics, no fonts beyond Google. Save it.