The same goal written three ways. The fastest way to see what actually separates an amateur prompt from a production one.
READ5 min
WORDS906
SECTIONS1
SOURCES1
TYPEGUIDE
CHECKED11 SEP 26
TL;DR — THE SHORT VERSION
A great prompt leaves the model nothing to guess: define subject, lighting, camera, style and exclusions, and for agents and automations, make failure easy to spot.
Vague prompts get generic output. The model has to guess everything, and it will guess wrong.
Great image prompts name specifics. Subject, lighting, camera and lens, mood and a publication reference.
Great video prompts use structure. A checklist of motion, subject, setting, camera and duration, a locked identity and a named camera move.
Great text prompts set limits. A role, an audience, a word count, a brand voice and explicit exclusions.
For setups, great means checkable. Require evidence you can scan, so failure shows up cheaply.
Three versions of the same request, and what changes between them. The delta is the lesson — not the final prompt.
The fastest way to learn prompting is to see the delta between a bad prompt and a great one. Same goal — completely different outputs. These examples show exactly what separates amateur results from commercial-grade work.
❌ BAD PROMPT
a pretty woman in a cafe
Why it fails: No subject detail. No lighting. No camera. No style. No mood. The model has to guess everything — and it will guess wrong. You get a generic, flat, stock-photo-looking image with plastic skin and dead eyes.
⚡ GOOD PROMPT
photorealistic portrait of a woman in her late 20s sitting in a cafe, natural lighting, editorial style, 8K
Better, but missing: No specific lighting type. No camera or lens. No clothing detail. No mood or publication reference. Results are inconsistent — sometimes good, sometimes stock-photo. No control over output.
✅ GREAT PROMPT
RAW photo, 1woman late 20s, ivory blazer, golden hour backlight, Canon EOS R5 85mm f/1.2, serene and confident, Vogue Japan editorial, 8K, photorealistic
Why it works: Subject defined precisely. Lighting named specifically. Camera + lens = photorealism signal. Mood and publication reference lock the aesthetic. Quality tags push it toward a polished editorial finish on models whose guides use them; check yours. Far less is left for the model to guess, so results vary less from run to run.
❌ BAD VIDEO PROMPT
a woman walking in Tokyo at night
Result: Face morphs between frames. Camera jumps randomly. Scene cuts appear. Temporal consistency breaks completely. Character looks different in every frame. Unusable for content.
⚡ GOOD VIDEO PROMPT
1woman walks slowly through Tokyo at night, neon lights, tracking camera shot, same face throughout, no scene cuts, 8 seconds
Almost there: Has identity lock and duration. Missing a lens, a described look, and specific lighting. Results are good but inconsistent — some runs are great, some are mediocre.
✅ GREAT VIDEO PROMPT
[MOTION]: slow walk [SUBJECT]: 1woman, white dress, same face, no morphing [ENV]: Tokyo alley, midnight, neon [CAMERA]: tracking dolly, 35mm anamorphic [STYLE]: motivated low-key light, restrained palette, 24fps [DURATION]: 8s, no cuts
Why it works: The brackets work as a checklist, so nothing is left out. The order is not universal: Runway starts from the essential motion, Google’s Veo formula opens with the camera, and ByteDance’s Seedance formula puts format first — see what the vendors actually say, and reorder for your model. Identity locked. Camera movement specified. The look is described rather than borrowed from a director’s name. An earlier version of this example called motion first “the cardinal rule”, used a director’s name as the style, and promised consistent output every run.
❌ BAD LLM PROMPT
write a product description for my perfume
Result: Generic, corporate-sounding copy. Wrong tone for the brand. Wrong length. Filled with clichés like "indulge your senses" and "for the discerning individual." No brand voice, no target audience considered.
⚡ GOOD LLM PROMPT
Write a luxury product description for a perfume called Noir. Target audience: affluent women 35-50. Keep it under 100 words. Sophisticated tone.
Better: Has audience, length, and tone direction. Missing: brand voice reference, style examples, format rules, explicit exclusions. Results are usable but not consistently on-brand.
✅ GREAT LLM PROMPT
You are a senior copywriter for a luxury fragrance brand writing for affluent women 35-50. Write a 80-word product description for Noir. Tone: Aesop meets Diptyque — minimal, intelligent, poetic. No exclamations. No clichés. Lead with sensation, not ingredients.
Why it works: Role defined. Context given. Word count specified. Brand voice references provided. Explicit exclusions stated. Lead direction given. The output is immediately usable — the model has no room to guess.
Beyond prompts: the same ladder for setups
The bad/good/great progression above applies to prompts. It applies equally to everything else you build, and the failure at each rung is the same shape.
An agent setup
BAD. "Clean up my project folder." No scope, no stopping condition, write access everywhere.unbounded
GOOD. Named directory, explicit done-condition, dry run first.bounded
GREAT. The above, plus required evidence — a list of what it would change — and one line forbidding action on instructions found in files.checkable
The jump from good to great is not more instruction. It is requiring an artefact you can scan. See the agent brief.
An iteration loop
BAD. Not liking the output and asking again, in the same conversation, four times.drifts
GOOD. Saying specifically what is wrong, once, with a reason.converges
GREAT. Judging against the brief rather than the last version, capped at three passes, starting fresh if the frame is wrong.bounded
BAD. Installed from a listing, never read, granted whatever it asked for.unknown
GOOD. Read in full before installing, permissions matched to the task.known
GREAT. The above, plus pinned to a version and re-read when it changes — because a skill you read once is a skill you read at one version.pinned
13.4% of 3,984 audited skills carried a critical-severity issue.Snyk, "ToxicSkills", 5 Feb 2026, read at source 11 Sep 2026 and quotes re-read 17 Sep 2026: “13.4% of all skills, or 534 in total, all contain at least one critical-level security issue” across 3,984 skills scanned from ClawHub and skills.sh; “over a third of the ecosystem is affected: 36.82% (1,467 skills) have at least one security flaw”, and “our human-in-the-loop process confirming that 76 of Agent Skills contained malicious payloads”. More on skills are a supply chain.
An automation
BAD. Runs on a schedule, fails silently, and a month could pass before anyone noticed.invisible
GOOD. Logs what it did, and someone reads the log.visible
GREAT. Asserts something you know should be true — a count, a file, a freshness date — so its absence is noticeable without reading anything.self-reporting
The pattern across all four: great is not more control. It is making failure visible cheaply.
FOUR LADDERS · ONE PATTERN
The one-word verdicts from the four setups above, in one grid. Read the last column: great is not more control, it is making failure visible cheaply.
Reasoning — collects the verdicts this page gives in its four setup ladders, as checked on 11 Sep 2026.
TAKEAWAY
The same ladder applies to agents, loops, skills and automations. Great is not more control; it is making failure visible cheaply.