Positioning: This piece is the sequel to A3, "Prompt Engineering for Office (Starter)" — A3 taught you to "write prompts that work" (the basic role + task + constraint formula); this one teaches you to "write prompts that hold": use the four-piece kit to move output from "hit or miss" to "dependable." The kit is an amplifier for A3, not a replacement — finish A3 first. Every example in this article runs on any major assistant — ChatGPT, Claude, Gemini, or Grok; the method is model-agnostic (fine-tune word-count-style constraints during the ③ acceptance pass).
Abstract
The same prompt wows you today and disappoints you tomorrow — the problem isn't luck; it's requirements that were never "anchored down." This article hands you the four-piece kit: role presets, few-shot examples, iterative refinement, and output formatting — one section each, plus complete "60-point → 90-point" rewrite comparisons for three scenarios (customer-service replies, action-item extraction from meeting minutes, weekly-report summaries). You'll walk away with three things: a personal role preset library, a complete rewrite log that climbs from 60 points to 90, and a scoreable prompt self-check list.
Who This Helps / Prerequisites
- You've read A3 and can write the basic "role + task + constraint" prompt.
- These pain points sound familiar: the same prompt produces unstable output; when you're unhappy you only know to hit "Regenerate"; you want to improve systematically but can only retry on gut feel.
- One practice tool is all you need: any major assistant — ChatGPT, Claude, Gemini, or Grok — this article's method is model-agnostic.
- High-frequency use cases: weekly-report summaries, support and email replies, meeting-minutes extraction, information organizing and rewriting.
- Redact your material first — mask or replace names, amounts, order numbers, and client names before pasting. For what must never be handed to AI, see A5, "AI Office Security & Privacy Red Lines."
Steps
① Role Presets: Three Elements — Identity + Boundaries + Style
In A3 a one-line role was enough. The advanced version needs all three elements in place:
- Identity: who the AI is and what job it owns — write it as a duty list, not a job title.
- Boundaries: what it may not do, and what to do when asked beyond scope (no fabricating, no promising on its own, flag anything missing).
- Style: tone, structure, and length — all written as checkable criteria ("≤ 120 words" can be ticked off; "a bit nicer" cannot).
Pitfall: don't pile up adjectives. "Senior, professional, top-tier expert" anchors no behavior at all; "any refund promise must be flagged 'needs manager approval'" anchors plenty.
Store your frequently used roles as a preset library: at the start of each task, paste the preset first, then the material. Here are 7 high-frequency starters (add or trim to fit your role):
| # | Scenario | Preset text (condensed identity + boundaries + style) |
|---|---|---|
| P1 | Weekly-report assistant | You are the weekly-report organizer for this project team. Work only from the material I provide. If risks or plans are absent from the material, write a single line "To be added" in that section — no guessing, no padding to hit a count. Output three sections — Progress / Risks / Plan — no more than 3 items each, every item starting with a verb and containing a number or an action |
| P2 | Customer-service replies | You are the support specialist for an online store; your job is handling negative reviews and inquiries. Any refund or compensation promise must be flagged "needs manager approval". Empathize first, then give the solution. Replies in 120 words or fewer, courteous but not groveling |
| P3 | Meeting-minutes distillation | You are the project meeting recorder. Extract only concrete tasks that were explicitly assigned; vague rallying cries and undecided items are not tasks. If a task is explicit but no owner was named, write "TBC". Output a four-column table: Decision / To-do / Owner / Deadline |
| P4 | Email polish | You are a business-email editor. Change only tone and wording — never facts or numbers. Formal business style; keep every key information point |
| P5 | Data interpretation | You are a business-data interpreter. Use only the numbers I provide and show the calculation. Conclusion first, evidence after; outliers must be called out by name |
| P6 | Excel consultant | You are an Excel formula consultant. Give the formula plus a one-line explanation only — no long tutorials. Formulas assume an English-language Excel by default (English function names) |
| P7 | Long-document summary | You are an industry-news summarizer. Summarize only from the source text; add no outside information. No more than 150 words plus 3 key takeaways |
Rewrite Comparison 1: Customer-Service Reply (60 points → 90 points)
The 60-point prompt:
Help me reply to a customer's negative review — make the tone nicer.
The customer says: shipping took 8 days, the package arrived damaged, and three days of messages to support got no reply.What's wrong with it: no identity, no boundaries, no style. Sometimes the AI grovels with apologies; sometimes it promises "an immediate refund" on its own — the style is re-rolled every time.
The 90-point prompt:
You are the support specialist for an online store. Your duties are responding to negative reviews, calming the customer down, and offering a solution; any refund or compensation promise must be flagged "needs manager approval" — never promise it on your own.
Tone: warm and sincere, courteous but not groveling. Empathize first, then give the solution. Keep the reply within 120 words.
Material (customer's negative review): shipping took 8 days, the package arrived damaged, and three days of messages to support got no reply. Write one reply.| Dimension | 60-point version | 90-point version |
|---|---|---|
| Role | None | Support specialist + duty list |
| Boundaries | None — may promise a refund on its own | Promises must be flagged "needs manager approval" |
| Style | "Make the tone nicer" (unverifiable) | Empathy first, then the fix; ≤ 120 words |
| Reuse | Written from scratch every time | Saved as P2; a new customer changes only the last line |
② Few-Shot Examples: Show the AI a Template
Plain-language version: few-shot examples = hand the AI two or three samples of "an input like this should come with an output like that," and it copies the pattern. The jargon is few-shot examples; all you really need to remember is "show it a template."
- When you need it most: whenever the output has a fixed format or a fixed standard — field extraction, classification and tagging, unified naming. Pure chat and creative copy don't need it.
- How many: 2–3 is plenty, covering the typical cases and ideally one edge case (e.g., "what to output when there are no action items").
- How to give them: write each pair as "input → ideal output," formatted exactly like what you actually want — the AI will imitate the formatting details along with everything else.
Rewrite Comparison 2: Action-Item Extraction from Meeting Minutes (60 points → 90 points)
The 60-point prompt:
Extract the action items from the meeting transcript below.
Meeting minutes: (paste a transcript of about 800 words)What's wrong with it: the output is a coin flip — sometimes a wall of prose, sometimes a list; sometimes "hypotheticals discussed in the meeting" get logged as action items. Owners and deadlines appear and vanish at random.
The 90-point prompt:
Extract action items from the meeting minutes. Output four columns: Item, Owner, Deadline, Status.
Rules: extract only concrete tasks that were explicitly assigned; vague rallying cries and undecided items are not tasks; if a task is explicit but no owner was named, write "TBC" as the owner; Status may only be one of: Not Started / In Progress / Completed.
Example 1 Input: Xiao Wang will send the test report to Lao Li by next Wednesday.
Output: | Write and send the test report | Xiao Wang | next Wednesday | Not Started |
Example 2 Input: Everyone should take a look at competitors' proposals next month, and we'll talk again later.
Output: (do not extract — a vague rallying cry, not a concrete task)
Example 3 Input: (the transcript contains no task assignments at all)
Output: (no action items)
Example 4 Input: the demo video needs a re-cut before the client meeting next Monday; no one was named to do it.
Output: | Re-cut the demo video | TBC | before next Monday | Not Started |
Meeting minutes: (paste the transcript)| Dimension | 60-point version | 90-point version |
|---|---|---|
| Format | Random paragraphs or lists | Four-column table with hard-coded headers |
| Standard | Hypotheticals leak into the list | Only "explicitly assigned" gets extracted |
| Missing values | Inconsistent | Uniformly written as "TBC" |
③ Iterative Refinement: Point Out the Gap → Make the AI Self-Review → Revise Again
Core principle: feeding it "what's off" works far better than feeding it "rewrite." "Rewrite" carries no new information — the AI just re-rolls the dice; "what's off" hands it a new anchor. The three-step method:
- Point out the gap: get specific — "which item, missing what." "Item 2 is all filler and has no numbers" beats "not happy, try again" by a mile. If only the word count blew past the limit, don't rewrite the whole thing — just follow up: "You're about 15 words over; compress to under 120 words and keep the 'needs manager approval' flag."
- Make the AI self-review: list your acceptance criteria for it and ask it to score its own output and name the deductions — it often finds the holes by itself.
- Revise and finalize: feed the confirmed gaps back in together with the missing material, then merge everything into one 90-point prompt and archive it for reuse.
Rewrite Comparison 3: Weekly-Report Summary (complete iteration log, 60 points → 90 points)
Round 0 · the 60-point prompt:
Turn my work this week into a weekly report:
1. Finished the login module
2. Fixed a few bugs
3. Sat in three meetingsWhat's wrong with the output: filler everywhere ("steadily advanced various work streams"), three items with no sense of priority — and it invented next week's plan for you.
Round 1 · point out the gap:
Rewrite requirements: (1) cut every filler phrase — each progress item must contain a verifiable action or number;
(2) organize into three sections "Progress This Week / Risks / Plan for Next Week"; the material contains no next-week plan, so mark that section "To be added" honestly — do not invent one.Round 2 · make the AI self-review:
Score the weekly report you just wrote against three criteria and state where you lost points: verifiable progress (does each item carry a number/action), no fabrication (is the missing next-week plan honestly flagged), complete structure (all three sections present).The AI's self-review comes back: "'Fixed a few bugs' still can't be verified — points off." — the gap is found.
Round 3 · feed in material and revise:
Replace "fixed a few bugs" with: fixed 2 bugs — the login-page CAPTCHA not refreshing, and broken pagination in the order list. Keep the rest unchanged and output the final version.The 90-point final prompt (merging rounds 0–3; save it into your preset library for long-term reuse):
You are the weekly-report organizer for this project team. Summarize only from the material I provide; do not fabricate.
If risks or the next-week plan are absent from the material, write a single line "To be added" in that section — no guessing, no padding to hit a count.
Output three sections: Progress This Week / Risks / Plan for Next Week, no more than 3 items per section,
each item starting with a verb and containing a verifiable action or number.
Material:
1. Completed development of the login module (including mobile-number CAPTCHA login)
2. Fixed 2 bugs: login-page CAPTCHA not refreshing; broken pagination in the order list
3. Attended the requirements review meeting, the weekly project sync, and the cross-team scheduling meeting| Dimension | 60-point version | 90-point version |
|---|---|---|
| Filler | "Steadily advanced" all over the place | Every item verifiable |
| Fabrication | The AI invented next week's plan | Missing items flagged "To be added" |
| Structure | One run-on paragraph | Fixed three sections |
| Reuse | Start from zero every week | Saved as P1; swap only the material |
④ Output Formatting: Tables / JSON / Word Count / Structure Templates
The first three moves solve "the content is right"; this one solves "the shape is right" — output you can paste straight into your document, spreadsheet, or script:
- Tables: hard-code the column names yourself; don't let the AI invent them ("output four columns: Item, Owner, Deadline, Status").
- Word count: give a cap, not "keep it short"; for long output add "conclusion first, details after."
- Structure templates: write the paragraph order out directly ("paragraph 1 empathy, paragraph 2 the solution, paragraph 3 next steps") and the AI will follow the arrangement to the letter.
- JSON (a data format that scripts and programs can read directly): when the output feeds a script or an automation pipeline, spell out the field names and allowed values:
Output the action items above as a JSON array with these fields: task (string), owner (string), deadline (string),
status (string, one of "Not Started / In Progress / Completed" only). Output JSON only — no other text.Web chat isn't always strict about pure JSON; for the automation scenarios that really matter, use the structured-output mode on the vendor's developer platform (the API platform each assistant company exposes to developers). The more "grammatical" your format requirements (fields, enumerations, order), the less room the AI has to freelance — and this is also how you pave the road to automation later.
Wrap-Up: Score It with the Self-Check List
Run every finalized prompt through the list (item 1 maps to the ① pitfall; item 4 maps to giving samples in ②):
| # | Check | ✓/✗ | Fix when ✗ |
|---|---|---|---|
| 1 | Role holds up: concrete identity with a duty list, not a pile of adjectives | Back to ① — write duties, not titles | |
| 2 | One task per prompt: this prompt does exactly one job | Split it into several prompts and send them separately | |
| 3 | Constraints verifiable: word count / format / no-fabrication can be ticked against the output | Swap "make it better" for countable metrics | |
| 4 | When a stable format matters, 2–3 "input → output" examples were given | Back to ② — add templates | |
| 5 | Material complete: background, source text, and data all inside the prompt | Add the material; don't make the AI guess. Redact sensitive info first (see A5) | |
| 6 | Boundaries drawn: no fabrication / no altering facts / anything out of scope gets flagged | Add one line: "Mark missing items 'To be added'" | |
| 7 | Output structure matches the use: fields and sections ready to use as-is | Back to ④ — hard-code the format | |
| 8 | Not model-picky: the core requirements still hold when you switch assistants | Paste into a different assistant and run once — do structure/constraints still get enforced? If not, strip any wording that leans on one vendor's habits |
Scoring rule: ✓ on at least 6 items, with items 3 and 6 mandatory ✓; any ✗ means the prompt goes back to the workbench.
Why It Works: Why the Four-Piece Kit Stabilizes Output
Every time a large language model writes the next word, it picks that word from a probability distribution — and the context you supply (everything the AI can see in this conversation, up to a length cap) decides where that distribution leans. In plain terms:
- Role presets narrow the distribution: the AI speaks only like "this identity," instead of vaguely resembling everyone and no one.
- Few-shot examples hand the distribution templates: copying a model is steadier than drawing from thin air.
- Iterative refinement generates another round with new information: feeding back "what's off" lands far more often than blindly clicking "Regenerate."
- Output formatting adds grammatical constraints: it squeezes the free-styling space out cell by cell.
What all four moves have in common: they reduce output variance (= how much the output swings) — going from "brilliant one day, mediocre the next" to "dependable every time" is done by steering the distribution, not by getting lucky.
FAQ / Troubleshooting
Q: Do more examples always mean more stability? A: No. Three high-quality ones beat ten mediocre ones; the examples must share one standard — mutually contradictory examples are worse than none; and too many also fills up the context window and dilutes the point.
Q: Isn't a role like "senior, professional, world-class expert" strong enough? A: Adjectives don't anchor behavior. The AI has no idea what "senior" concretely means, but it knows exactly what "flag any refund promise with 'needs manager approval'" means. Write a duty list, not a pile of titles.
Q: When I'm unhappy I just click "Regenerate" or say "rewrite" — why doesn't that work? A: No information delta — it's another roll of the dice. Feed it "what's off": which item is hollow, which number is missing, where the tone is too sharp. The more specific the gap, the sharper the next version.
Q: A prompt I tuned in one assistant goes off-script in another? A: Check whether you leaned on one vendor's instruction habits. A good prompt is self-contained — every requirement lives inside the prompt itself, never counting on some assistant's hidden defaults; don't tune it until it works for one model only.
Q: Do all four pieces have to go in every time? A: No. Copywriting runs on role + style; field extraction runs on few-shot + formatting; for a one-off simple task, the A3 basic formula is enough. Combine as needed.
Recap & Where to Go Next
One-line recap: A3's three elements solve "can do the job"; this article's four-piece kit solves "does the job reliably" — set the role, hand over templates, push back on gaps when it misses, then nail the format down.
Where to go next:
- Preset library into your knowledge base: store P1–P7 as ordinary notes for unified management and updating (how to build one: F2, "Build Your Personal AI Office KB").
- Working examples: the reply presets in B29 ("Tame Your Inbox with AI") and the AI-field prompts in B30 ("Tencent Docs & Feishu: AI Features You Already Have") are ready-made implementations of this article's method.
- Toward automation: the JSON output formatting in ④ is the interface left for scripts and RPA (robotic process automation software) — define the fields now, and machines can consume them later.
- Back to basics: if "role + task + constraint" slips your mind, return to A3, "Prompt Engineering for Office (Starter)."
CTA
Once the four-piece kit is second nature, post one entry from your preset library in the comments and let's polish it together. Next up: A9 · AI Work on Your Phone — how to keep AI office work running on your commute and out in the field.
This article is part of the "AI-Powered Office: From Beginner to Expert" series, A8 (A Series Methodology · Prompts / L2). Author: He Xun. English translation: Yian. Status: Draft.