Prompt For Me

Prompt For Me is a free agent skill maintained by Scopeful. It teaches an AI coding agent such as Claude Code, Cursor, Windsurf or Codex how to drive this tool correctly, so you do not have to re-explain it every session. Every published Scopeful skill is free and the install command is public, with no sign-in required. Scopeful also tracks hand-verified USD pricing for 39 AI creative tools at https://www.scopeful.org/tools.

Write, improve, or translate a prompt for any AI image, video, or voice tool.

Tags: prompts, multi-model

Reference

name: prompt-for-me description: Use this skill whenever the user wants a prompt written, improved, or translated for any AI image, video, or voice generator. Triggers include "write me a prompt", "prompt for Midjourney/Flux/Kling/Seedance/Veo/Sora/Ideogram/GPT Image/NanoBanana/Reve/ElevenLabs", "improve this prompt", "why does my generation look bad", "which model should I use for this", or pasting an idea and naming any generation app. The skill detects the target model, applies that model's prompt grammar, and outputs a pasteable prompt with parameters and a cheap-iteration tip.

Prompt for me

One skill, every generator. The user gives an idea; you return a prompt in the target model's native grammar. Freeform "beautiful, detailed, 4k" prompts waste money on every platform. Each model below has a different grammar — using the wrong one is the #1 cause of bad generations.

Model facts current as of 2026-06. If a model or parameter seems off, check the platform's changelog before arguing with the user.

Workflow

1. Identify the target model. If named, use it. If not, ask ONE question: "Image, video, or voice — and do you have an app you already pay for?" Never write a generic prompt.
2. No preference? Pick from the routing table and say why in one line.
3. Write the prompt in that model's grammar (sections below).
4. Output: pasteable prompt block, parameters, and one cost-saving iteration tip. No essay around it.

Routing table (user has no model preference)

Universal rules (every model)

• Subject first. The most important element goes at the start of the prompt.
• Lighting is explicit: "soft window light from the left", never "well-lit".
• One action, one camera move per video shot. Multi-action prompts fragment attention.
• Text-in-image goes in "double quotes", kept short.
• Iterate cheap, render expensive once: preview at low res/quality/duration, lock direction, then final. Include the concrete cheap path in every output.
• Never put resolution in the prompt text when it's a UI/API parameter (Seedance, Kling).

Image models

Midjourney (V8.1 default; V7, V6.1, Niji 7)

Grammar: [natural language, most important element first] --ar --s --style raw --refs. Pasteable for midjourney.com/Discord — there is no public API.

• Always set --ar (default 1:1 rarely fits). Then --s (0–1000, default 100; low = literal), --style raw to suppress MJ's beautification.
• --draft = cheap ideation (24 low-res images); --hd = 2K at ~3× fast-hours. Never iterate at HD.
• Refs: --sref url --sw 0-1000 style only; --oref url --ow any element (supersedes --cref); --cref --cw 0-100 faces. Weighting: concept::2 other, negatives via --no.
• Workflow: SD grid → pick → Vary → Upscale. Example:

A polished steel water bottle on white marble, soft morning window light from the left, shallow depth of field, minimalist product photography --ar 4:5 --s 50 --style raw

Flux (FLUX.2 pro/flex/max/klein; FLUX.1 Kontext, Fill)

Grammar: positive prompt (subject first, style last) + negative prompt (load-bearing on Flux) + aspect ratio.

• Default text, watermark, extra fingers, blurry in NEGATIVE; add case-specific exclusions.
• flex/max are megapixel-priced: 4MP costs 4× 1MP. Preview at 1MP, final at target MP, same seed.
• Kontext = reference editing: "Match the mug from the reference exactly in shape and proportions. Change the color to matte forest green." Fill = inpainting (image + mask description + fill prompt).
• Cheap iteration: FLUX.2 [klein] 4B (~$0.014/img).

GPT Image 2 (OpenAI)

Natural conversational language — the text backbone interprets intent, so don't over-specify. Multi-turn editing is the core feature: first prompt rough, then "make the sky more orange". Text in quotes + placement + font style. Quality tiers: low drafts → high finals. Resolution: edges multiples of 16, max 3840px, ratio ≤ 3:1.

Ideogram 4.0

The text-in-image leader. Structure: subject + scene + "EXACT TEXT" + style words + composition + lighting. Use the Typography style tag when text accuracy matters. Describe font properties ("bold retro sans-serif"), exact typefaces are not selectable. Magic Prompt ON for exploration, OFF for control. Text wrong? Simplify words, regenerate, or Magic Fill the bad letters at ~90 strength.

NanoBanana Pro (Gemini 3 Pro Image)

Autoregressive → full sentences, never keyword soup. Positive-only (no negative prompts; describe what you want). First words get the most attention. Up to 14 reference images; on Magnific use @image1 mentions with an intensity phrase: "Inspired loosely by…" / "Match the visual style, palette and composition of…" / "Replicate the exact style of…". Iterate on NanoBanana 2 at 1K, final on Pro at 2K/4K. Output a single clean paragraph in a code block.

Reve 2.0

Code-based representation → lossless edits, strong multi-element adherence. Structure: Subject → Action → Setting → Style → Lighting → Camera. Always anchor a style ("photorealistic", "cinematic still") or it defaults to cinematic. Great at text-in-image. For edits: name what changes, everything else stays locked — artifacts don't accumulate, so iterate freely.

Video models

Shared block format — fill only what the model supports:

SCENE:    [static starting state, 1-2 sentences]
MOTION:   [what changes — verb chain, not adjectives]
CAMERA:   [ONE move: "static camera" / "slow dolly in" / "tracking left to right"]
AUDIO:    [soundscape + "She says, '…'" — or "silent, no audio"]
DURATION: [seconds, within model cap]
SEED:     [pin for iteration]

Kling 3.0 (V3 Standard/Pro, Omni, Turbo; 2.5 Turbo Pro)

• Caps: V3 15s/1080p; 2.6 10s/1080p; Omni 4K editing path. Audio billed separately on V3 (bundled on Turbo) — quote audio-on/off separately.
• One subject, one action per shot; dialogue = two shots, cut between.
• Same character across shots → V3 Pro + Elements 3.0 BIND_SUBJECT: [reference URL], scenes separated by ---.
• Audio on → name the soundscape: "ambient: espresso hiss, ceramic clink". Lip-sync languages: ZH/EN/JA/KO/ES.
• Iterate: 5s 720p preview (or Kling 2.5 Turbo Pro at ~$0.07/s), then final with same seed+prompt.

Seedance 2.0 (2.0, Fast, Mini, 1.5 Pro)

• No real-person face references — refuses or distorts. Describe the character in text; fictional characters are the sweet spot. Real faces → Kling or Runway.
• Audio on by default (2.0 family) — write "silent, no audio" if unwanted. Resolution is a UI parameter, never prompt text. Mini = 720p B-roll only, never faces/talking heads. Logos unreliable → generate still in Flux/NanoBanana, then image-to-video.
• Motion ≤ 80 words or beats get dropped. Dialogue: one line, 8–14 words, transcribed.
• Seed-reuse trick: 2s 720p preview → same seed + same prompt at full duration = same shot extended. Pin the seed.

Veo 3.1 (Standard, Fast, Lite)

7 elements in order: subject → action → setting → camera → style → lighting → audio. Template: [Shot type] of [subject] [action] in [setting]. [Camera move]. [Lighting]. [Style]. Audio: [dialogue in quotes + SFX + ambient].

• Audio always on, included in price. Dialogue as direct quotes; SFX explicit ("glass shattering").
• Camera vocab is load-bearing: dolly/pan/tilt/tracking/crane/orbital; lens terms work ("anamorphic", "shallow depth of field").
• Style refs work: "Wes Anderson", "35mm film grain, Kodak Portra 400".
• Rewriter is on by default (enhancePrompt: true) — disable for raw control. Lite has no 4K. 24fps, SynthID watermark, output stored 48h only — download immediately.

Sora 2 / sora-2-pro (deprecated)

Flag in every output: API shutdown 2026-09-24, no successor — migrate to Veo 3.1 / Seedance 2.0 / Kling 3.0. If used anyway: SCENE/ACTION/CAMERA/AUDIO block, 20s max, sora-2 caps at 720p (pro for 1080p), pin snapshot sora-2-2025-12-08.

Voice — ElevenLabs

• Punctuation is pacing: commas short pause, periods stop, "…" dramatic pause.
• Stage directions in brackets: [whispering], [excited]. Emphasis via ALL CAPS. Precise pauses via <break time="1s"/>.
• Misread words → phonetic spelling ("nite"). Short sentences > run-ons. Natural fillers ("um") humanize.
• Settings: Stability high = consistent/flat, low = expressive; Similarity for clone fidelity.

Output format

Always end with exactly this, no extra prose:

MODEL: [model + platform]
PROMPT:
[pasteable prompt in the model's grammar]
PARAMETERS: [only what the platform exposes]
ITERATE CHEAP: [the one-line cheap preview path for this model]

If the user's idea is missing a decision the grammar needs (aspect ratio, duration, audio on/off), ask one question — never guess silently on paid renders.