1328 hand-tested AI prompts across image, video, and music — each with full context, variables, and a sample output. Content is authored in Traditional Chinese; the UI here is English.
The key to an infographic is carving out the sections first — quoting each of the three section headers by name in quotation marks keeps the model from blurring text into pure decoration; adding...
Leave 'condensing the information' to the model, but lock down 'which words absolutely must appear' with quotation marks — educational diagrams are most vulnerable to labels going off-script, and this...
Naming `orthographic` along with the three drafting views — plan, elevation, and section — makes the model...
The phrase `deconstructed to show the texture of…` ties an exploded-view diagram together with appetite appeal...
The sketch serves as the layout template, and the model only handles the materials — `Keep the exact layout of…` is...
Clear division of labor — placeholder boxes get 'their content swapped in,' while button positions and the grid are 'not to be touched'; UI mockups are most vulnerable to...
The 'make-or-break instant' is written as the shot's explicit task (catching the upper hold perfectl...
Naming the specific technique outright (Fosbury flop) plus the approach-run curve and takeoff foot means the model doesn't have to guess the posture; 'the bar quivering slightly after clearance...
Defining the soundtrack negatively — 'silent except for the edge cutting through snow' — forces a sense of emptiness far better than listing a string of sound effects; waist-deep...
Google's own official example of the textbook answer for the five-part 'subject + action + setting + lighting + style' formula; two light sources (cold white...
The key to a convincing selfie feel isn't the word 'selfie' itself, but writing 'the arm must be clearly visible in frame' as a hard composition requirement; the closing...
First/last frame only gives you the two endpoints — this line carries the entire middle: it locks down the three keys of a match cut: wha...
Before/after shots most easily turn into a hard swap in the last second — `across the clip` requires the...
A time-lapse needs evidence — besides the color temperature shifting from cool morning to warm dusk, two independent time markers are added (clouds drifting slowly, shadows lengthening), so it doesn't...
A runway clip's texture comes entirely from lighting — the phrase `harsh flash photography simulatio...
The soul of a dance piece is 'making the light beam visible' — `a single spotlight above` plus `dust drifting in the light beam` …
Short, spec-style phrasing — focal length, light position (backlight creating a halo), camera position (static), motion (slowly...
Studio lighting terms are used as literal parameters — key light and fill light each get a specified direction and gear (window left / bounc...
Native audio is the thing models most often bury under auto-generated music — start by setting the spec with `Native audio. 48kHz.`...
Doesn't let the AI just hand in its answer as-is: it first tags the source type of every key sentence, then hunts for counterexamples to attack its own claims, and finally grades confidence level item by item with suggested directions for verification. Good homework before researching, writing a report, or making a decision. Honest note: this isn't the same as actual fact-checking (the model may not be able to search the web) — but it forces uncertainty out into the open and noticeably reduces confident-sounding nonsense.
Report a bug in this format and the AI's fix hit rate jumps noticeably: give all four parts at once — the raw symptom, a minimal repro, expected behavior, and environment — and force the AI to diagnose first, ask for more information when it's insufficient, and never guess-and-change blindly. Suited to any scenario where you're asking ChatGPT or Claude to fix code. Note: the AI's diagnosis is reasoning, not an actual test run — after the fix, be sure to actually run through the included verification checklist.
Feed in a lease, employment, or outsourcing contract, and it scans clause by clause from your side of the deal: a plain-language translation, worst-case scenario, risk grading, and negotiation angle for each, plus three questions to bring to a professional at the end. Note: this is a prep tool for understanding a contract, not legal advice — for high-value contracts or ones involving major rights, consult a lawyer, and it's safest to treat the scan results as homework to do before meeting one.
Generate a step-by-step SOP diagram for staff training in one pass: the title and every step's text locked with full-width quotation marks, character counts noted per step, and step numbers in Arabic numerals to reduce the error surface. Good for follow-the-steps scenarios like opening procedures or cleaning protocols. Limits: keep steps to six or fewer, with roughly ten characters per line for the most stable results; the illustrated action occasionally doesn't match the text — just ask it to redraw that one panel.
Two key priorities for generating event tickets and staff passes: locking the event name and date in quotation marks (with character-count declarations and safeguards for characters like '證/号'), and clearly reserving a pure white blank zone for a serial number or QR code to be added later. It specifically adds a 'no fake barcodes or serial numbers' rule — the model loves to invent a fake barcode on its own initiative, and you don't want to find out it can't be scanned only after printing.