提示在結果部分將目標和實績一起列出(達成 vs. 未達成),強制進行誠實評估,並使建議感覺基於證據而不是意見——利益相關者信任指明不奏效之處的回顧。
ChatGPTGeminiClaude
Instagram/LinkedIn carousel from a blog post
Turn a long post into a 7–9 slide carousel: one idea per slide, swipe-worthy pacing.
提示詞 2 個欄位待填
Turn this article into a carousel of [7-9] slides for [PLATFORM — Instagram or LinkedIn].
Structure:
- Slide 1: a hook that names the payoff or the mistake (max 12 words)
- Slides 2 to N-1: ONE idea per slide, max 25 words each, no filler transitions
- Last slide: recap in 3 bullets + one CTA:
- After the slides, write the post caption (2 sentences + question)
Keep the numbers and concrete examples from the article; cut the storytelling.
Article:
範例輸出
Slide 1: Your onboarding email is why they never came back.
Slide 2: 63% of trial users open email #1. Only 11% open email #4. The sequence is the leak.
Slide 3: Email #1 job: one action, done in 90 seconds. Not a feature tour.
Slide 4: Subject lines with the user's goal beat product names 2:1 in our test.
Slide 5: Send #2 after the action happens, not 24 hours later. Behavior beats schedule.
Slide 6: One CTA per email. Two CTAs = half the clicks on both.
Slide 7: Recap — 90-second first win · behavior-triggered sends · one CTA. Steal our full sequence: link in comments.
Caption: We rewrote a 4-email onboarding sequence and doubled activation. The whole teardown is in the carousel. Which email in your sequence gets ghosted the hardest?
提示The "max 25 words per slide" constraint is what makes output usable — without it models write paragraph slides that die in design. Tighten to 18 words for Instagram, loosen to 30 for LinkedIn.
ChatGPTGeminiClaude
Cluster a keyword list into page groups
Sort a raw keyword dump into pages — one intent per page, no cannibalization.
提示詞 1 個欄位待填
Cluster this keyword list into page groups. One page per distinct search intent — if two keywords can be satisfied by the same page, they belong in the same cluster.
For each cluster give:
- Cluster name + the page type that wins (guide / comparison / tool page / listicle)
- Primary keyword (highest value) and supporting keywords
- One-line content angle
- Flag: any keyword that is its own intent and deserves a separate page
Also list keywords to DISCARD (zero commercial or informational value for us) with a one-word reason.
Our site: . Keywords:
[PASTE KEYWORD LIST, ONE PER LINE — VOLUME OPTIONAL]
範例輸出
Cluster 1 — "invoice template" (page type: free tool page)
Primary: free invoice template (9,900)
Supporting: invoice template word, invoice template pdf, simple invoice template
Angle: instant download, no email gate — every competitor gates.
Cluster 2 — "how to invoice" (page type: guide)
Primary: how to write an invoice (2,400)
Supporting: what to include in an invoice, invoice requirements freelance
Angle: first-invoice walkthrough for freelancers with a filled example.
Separate-page flag: "invoice vs receipt" — comparison intent, its own page.
Discard: "invoice meaning in hindi" (locale), "invoice 2019 crack" (junk).
提示Give the model your site's one-line description — clustering changes completely depending on whether you sell software, run a blog, or offer templates, and models default to generic blog clustering without it.
Run in Claude Opus 5.5 (claude-opus-5-5) on 1 Oct 2026; output shortened. Input: made-up rough notes from a 25-minute call with the operations director of a home-care company.
Stage: Discovery
Pain points:
- Scheduling for 140 caregivers across 2 counties runs on spreadsheets and group texts
- Missed visits ("a few a month"); one became a client complaint in August
Budget signal: Dana can sign herself for anything under $10k/yr; above that the CFO signs off. They passed on another scheduling tool last year as "too expensive" (no figure given).
Decision makers: Dana Whitfield (Ops Director) — signs under $10k/yr; Raj (CFO) — signs off on anything over $10k/yr
Objections / risks:
- Price sensitivity — rejected another tool last year as too expensive
- EVV support for their Ohio Medicaid clients is unconfirmed (we owe an answer)
- Timeline: they want to be live before January
Next action + date: Send pricing for 140 users and the EVV answer — no date mentioned. Dana to set up a call with Raj the week of 13 Oct.
Where it fell short: nothing was invented, but "Stage: Discovery" is the model's own call — the notes never name a stage and the prompt forces one of four — and her request for mobile clock-in with GPS was dropped because no field asks for requirements.
提示「未提及」規則阻止 AI 編造你從未聽過的預算或利益相關者——這就是使摘要安全地直接貼入管道的原因。
Run in Claude Opus 5.5 (claude-opus-5-5) on 1 Oct 2026; output shortened. Input: a 600-word made-up release-meeting transcript with three speakers (Lena, Daniel, Mei), two decisions and four owned tasks.
**TL;DR** — Fieldbook 2.3 ships as planned on Wednesday 14 October, but without offline sync, which moves to 2.4 because conflict handling still loses edits. Beta feedback moves from email to the in-app form starting with 2.3.
**Decisions made**
- Ship 2.3 on 14 October without offline sync; offline sync moves to 2.4.
- From 2.3 on, beta feedback goes through the in-app feedback form instead of email.
**Action items**
| Task | Owner | Due date |
|---|---|---|
| Email Kestrel Logistics and Brightwater Clinics to explain the new offline-sync timing | Lena | Mon 5 Oct |
| Merge the fix for the Android 14 login crash | Daniel | Thu 8 Oct |
| Redo the six iOS/Android store screenshots and the English + German release notes | Mei | Fri 9 Oct |
| Submit 2.3 for App Store review | unassigned | Mon 12 Oct |
| Enable the feedback-form feature flag in the 2.3 release candidate | Daniel | Tue 13 Oct |
**Open questions**
- Whether to hold the release if the release candidate still crashes on more than 0.5% of sessions — to be decided on 12 October from the numbers.
Where it fell short: all four real tasks came back with the right owner and date, but it also turned a passing remark ("submitting on Monday the 12th still works") into a fifth row. Nobody took that task on — delete the row or give it an owner before you send the notes.
Paste the assignment, get a 4-level rubric with observable criteria — no vague "shows understanding".
提示詞 4 個欄位待填
Create a grading rubric for this assignment.
Class: . Assignment: .
Requirements:
- 4 performance levels: Exceeds / Meets / Approaching / Beginning, with point values out of
- 4–6 criteria rows drawn from what the assignment actually asks (not generic "effort")
- Every cell describes OBSERVABLE work ("cites two primary sources" not "shows good research skills") so two graders would score the same paper the same way
- Language students in can read themselves
- End with a 2-sentence student-facing summary of what an A looks like
Format as a table.
範例輸出
Criterion: Thesis & claim (8 pts)
- Exceeds (8): Arguable thesis in first paragraph; every body paragraph advances it.
- Meets (6): Clear thesis; most paragraphs connect to it.
- Approaching (4): Topic stated but claim is a fact, not an argument.
- Beginning (2): No identifiable thesis.
Criterion: Evidence (8 pts)
- Exceeds (8): Two+ primary sources cited and explained; quotes under 2 lines each.
- Meets (6): Two sources cited; explanation present but thin in one spot.
[...]
What an A looks like: You take a position someone could disagree with, and every paragraph gives a piece of evidence for it — then explains the evidence instead of letting quotes speak for themselves.
提示The "two graders score it the same" constraint is what forces observable language — without it every model drifts into "demonstrates strong understanding" cells that cause grade disputes.