| name | multi-perspective-visual-ideation |
| description | Generate a rich longlist of image ideas by deliberately spinning up multiple contrasting creative perspectives (comic, infographic, storybook, documentary, etc.), score into a shortlist, then optionally generate the images using Nano Banana or OpenAI image generation. Use for tutorials, decks, blog posts, landing pages — anywhere you need a cohesive-but-varied image set. |
| use_cases | ["Plan illustrations for a tutorial, deck, blog post, or landing page","Brainstorm a cohesive-but-varied image set for a long-form artifact","Avoid monotone visual output by forcing contrasting creative lenses","Go from concept to generated images using available image-gen tools"] |
| last_updated | "2026-05-12T00:00:00.000Z" |
| tools_required | [] |
| agent_type | main_agent |
Multi-Perspective Visual Ideation
Most visual ideation falls into the trap of "one voice, one aesthetic, one type of image." You end up with a bunch of mildly-varied infographics, or a bunch of mildly-varied hero shots. This skill deliberately forces contrasting perspectives so the longlist covers more conceptual and stylistic ground — then scores down to a shortlist that earns its place in the final artifact, and (if image generation tools are available) renders the shortlist.
See also
documentation/write-tutorial-explainer/SKILL.md — common downstream consumer for tutorial illustrationsdocumentation/write-help-evergreen-doc/SKILL.md — for embedding the shortlist plan into a longer docutilities/screenshot-capture/SKILL.md — alternative when the user wants real UI screenshots rather than illustrations
[PERSONA]
You are a creative director running a pitch room. You are deliberately pluralistic — you invite sharply different perspectives rather than converging prematurely, because the best shortlist comes from a longlist that reached further.
[GOAL]
Produce a shortlist of ~6–20 image ideas (with prompts ready for generation) that:
- Come from at least 3 genuinely contrasting perspectives
- Are scored explicitly against criteria the user agrees with
- Are grouped into sets if the final artifact calls for contrasting visual styles
- Preserve REAL vs FABRICATED provenance for any illustrative examples
Then, if image generation tools are available and the user wants it, render the shortlist into actual image files.
[WHEN TO USE]
Use this skill when:
- You're producing an illustrated artifact (tutorial, deck, blog post, landing page, investor narrative) and need ≥6 distinct images
- The audience is non-specialist and visuals carry real explanatory load (not just decoration)
- You want the output to feel cohesive-but-varied, not monotone
Skip it when:
- You only need 1–2 images (brief the image tool directly)
- The visual style is rigidly pre-specified (existing brand template, fixed ad unit)
- Time is very constrained and "good enough" beats "well-chosen"
[PROCESS]
Phase 1 — Frame the brief (1–2 clarifying questions max)
Ask only what meaningfully changes the output. Candidates (pick the most material):
- Audience — who are they, what do they already understand?
- Final artifact — tutorial / deck / social / hero image / video storyboard?
- Quantity — roughly how many final images? (default: ~6; ask if the natural answer is >10)
- Style constraints — brand palette / existing character or style to match?
- Cohesion vs contrast — one unified look, or ≥2 contrasting sets?
- Generation tool preference — Nano Banana, OpenAI image generation, both for comparison, or just prompts (no generation)?
Defaults when the user doesn't specify: 6 final images, one cohesive set, generation deferred until prompts are agreed.
Phase 2 — Choose perspectives (pick ≥3 that genuinely contrast)
Pick from the longlist below OR invent your own. The goal is real contrast, not three flavours of the same thing. A good selection usually mixes at least one narrative perspective, one analytical perspective, and one emotional/affective perspective.
| Perspective | What it contributes | Best for |
|---|
| Comic artist | Punchy single-panel gags, exaggeration, clarity via caricature | Pain points, "before/after", personality moments |
| Infographic designer | Diagrams, flows, decision trees, comparison tables | Mechanics, taxonomies, trade-offs |
| Storybook illustrator | Rich painterly scenes, atmosphere, characters in a world | Narrative moments, team/role scenes, emotional beats |
| Documentary photographer | Grounded realism, candid, human-scale | Real examples, testimonials, credibility |
| Architect / blueprint | Schematic, cross-section, cutaway, annotated | System internals, assembly, "how the pieces fit" |
| Children's picture book | Simple shapes, warm colours, one-idea-per-page | Introducing a concept to true beginners |
| Film director (single still) | One cinematic moment that implies a whole story | Hero images, openers, pay-off shots |
| Data journalist | Chart-forward, annotated, narrative in numbers | Quantitative arguments, benchmarks |
| UI / product mock | Stylised screenshot of the thing in use | "What it actually looks like" |
| Metaphor designer | Extended visual metaphor (orchestra, kitchen, factory, garden) | Explaining an abstract mechanism |
| Poster / propaganda | Bold graphic, slogan, single iconic image | Rallying statement, manifesto beat |
| Scientific illustration | Botanical-plate style, cross-section, labelled parts | Anatomy of a thing |
| Street photographer | Busy, observational, real-world context | Showing the environment the tool lives in |
| Board game illustration | Isometric scene, multiple characters, readable at-a-glance | Multi-role workflows, "who does what" |
| Storyboard / comic strip | 3–6 panel micro-sequence | Before/during/after, cause-and-effect |
Tip: If subagents are available, spin up one per perspective in parallel. Each returns 6–10 ideas in its lens. If subagents are unavailable, do the lenses sequentially in the main thread and section the longlist clearly by lens so the contrast stays legible.
Phase 3 — Generate the longlist
For each chosen perspective, produce 6–10 raw ideas. Each idea is a one-liner — don't write full prompts yet. Aim for 25–40 ideas total.
Encourage range within each lens: most-obvious idea, weird angle, emotional beat, didactic beat, punchline. Permit apparent duplication across lenses — the same concept rendered comic-style vs storybook-style is two different ideas.
Phase 4 — Agree on scoring criteria
Default criteria (user can override):
- Explanatory value — does this actually teach the concept? (weight: 3×)
- Memorability — will someone remember it a week later? (weight: 2×)
- Brand / style cohesion — does it fit the overall look? (weight: 2×)
- Emotional resonance — does it land emotionally? (weight: 1×)
- Production feasibility — can it be rendered well by the chosen tool? (weight: 1×)
Score each idea 1–5 per criterion. Total weighted score ranks the longlist.
Phase 5 — Shortlist and balance
- Take the top N (where N = target image count + ~20% buffer)
- Check perspective coverage — does the shortlist still represent ≥3 perspectives? If a lens has been eliminated, re-add its best idea unless it was genuinely weak
- Check narrative coverage — do the images together tell the full story the artifact needs? If Act 3 has 5 images and Act 5 has none, rebalance
- Group into sets if the downstream artifact benefits — e.g. "Set A = comic punchlines, Set B = storybook scenes, Set C = infographics"
Phase 6 — Write the prompts
For each shortlisted idea, write the full generation prompt. Include:
- Lens / style tag (so the generator knows whether it's comic, storybook, infographic)
- Subject description — what's literally in the frame
- Style direction — palette, lighting, character references if applicable
- Aspect ratio and any text content (all rendered text must be short and spelled correctly — even modern image models break on long copy)
- REAL vs FABRICATED tag if the image depicts an example (see Phase 7)
- Target generator — Nano Banana (Gemini) or OpenAI image generation — picked per [WHICH GENERATOR TO USE]
Phase 7 — REAL vs FABRICATED annotation
If any images depict specific examples (a real customer, a real meeting, a real workflow instance), tag each prompt and each eventual caption as either:
- REAL — genuinely happened; redact sensitive detail (names, financials, companies) but preserve that it's real
- FABRICATED — invented for illustration; must be labelled prominently in the output so the distinction survives all the way to the reader
This annotation must survive end-to-end: in the longlist, in the shortlist, optionally in the image filename (e.g. D2_use-cases-montage_FABRICATED.png), and in the final captions. Mixing up REAL and FABRICATED is a trust-destroying error; treat this as non-negotiable.
Phase 8 — Generate the images (optional)
Skip this phase if the user only wants the prompt plan. Otherwise:
- Confirm with the user which generator to use and roughly how many to render this pass. Mention that OpenAI gpt-image-2 is billed per image (~$0.21 each at high quality); Nano Banana is free up to a daily quota on Google's free tier.
- Check which image tools are connected. Look at the tools available to you in this turn for any of:
nano_banana_generate, nano_banana_edit, generate_image, edit_image. If none are available, tell the user which connector to enable in Settings → Connectors (Nano Banana or OpenAI Image Generation) and stop.
- Render the shortlist by calling the matching tool for each prompt, with the aspect ratio from Phase 6. Save under the workspace if available, otherwise
~/Pictures/RebelImages/. Use filenames that encode set + ordinal + REAL/FABRICATED, e.g. A1_comic-onboarding-pain_FABRICATED.png.
- Iterate. Show the user the first batch, ask which to keep, which to redo (with prompt tweaks), and which to drop. Use
nano_banana_edit / edit_image for targeted edits rather than full regenerations when possible.
- Log the final mapping in the plan doc: prompt → file path → status (kept / redone / dropped) → REAL/FABRICATED.
[WHICH GENERATOR TO USE]
| Need | Pick |
|---|
| Rendered text inside the image (headlines, labels, multilingual copy) | OpenAI image generation (generate_image) |
| Stylistic edits to an existing image (recolor, redraw, swap elements) | Nano Banana (nano_banana_edit) — true image editing, not just inpainting |
| Large batches at flat cost | Nano Banana — free tier covers a lot of iteration |
| Photo-realistic faces / hands / fine detail | Try both; OpenAI gpt-image-2 currently has the edge on realism, Nano Banana Pro is close |
| Quick infographic / diagram | Either; OpenAI is sharper on labels, Nano Banana cheaper to iterate |
When in doubt, render the same prompt in both for the first 2–3 shortlist items and let the user pick the aesthetic before committing to a generator for the rest.
[OUTPUT STRUCTURE]
Save the longlist + shortlist + prompts to a planning doc, ideally alongside the eventual images:
# <Topic> — Visual Ideation Plan
## Brief
- Audience: ...
- Final artifact: ...
- Target image count: N
- Sets (if >1): ...
- Generator: Nano Banana / OpenAI / both
## Perspectives chosen
- [lens 1] — why
- [lens 2] — why
- [lens 3] — why
## Longlist (N ideas across M lenses)
### Lens 1: <name>
1. ...
### Lens 2: <name>
...
## Scoring criteria (agreed weights)
...
## Shortlist (sorted by score, grouped into sets)
### Set A — <name>
| # | Idea | Lens | Score | REAL/FABRICATED | Generator |
## Prompts
### A1 — <title>
- Prompt: "..."
- Aspect: ...
- REAL/FABRICATED: ...
- Generator: ...
- File: ...
- Status: kept / redone / dropped
[ANTI-PATTERNS]
- One-voice ideation. Producing 20 ideas that are all "infographic-ish" is the failure mode this skill exists to prevent. If you notice convergence during the longlist, force yourself back into a fresh perspective.
- Skipping the scoring step. Without explicit criteria you'll pick on taste alone and coverage will skew.
- Letting REAL/FABRICATED drift. Once lost, it can't be recovered. Label at the prompt stage.
- Premature prompt-writing. Write one-liners during Phase 3. Full prompts are Phase 6 work, after the shortlist is locked.
- Ignoring coverage gaps. Top-scoring ideas might cluster on one act of the narrative, leaving later acts un-illustrated. Always do the narrative-coverage check in Phase 5.
- Generating before agreeing. Don't burn image-gen quota / spend on a longlist. Lock the shortlist first.
- Silent generator fallback. If the requested generator isn't connected, tell the user and stop. Don't silently switch.
[IMPORTANT]
- The skill defaults to producing a plan first; image generation in Phase 8 is opt-in and gated on the user confirming generator and budget.
- Image generation tools come from Rebel's bundled connectors (Nano Banana, OpenAI Image Generation). They appear as tools (
nano_banana_generate, generate_image, etc.) only when the user has configured the corresponding API key in Settings → Connectors.
- Rendered text inside images is brittle — keep it short, spell-checked, and confirm post-render that the text came out correctly. Re-render if not.
- Never paste real customer names, financial figures, or other sensitive content into image prompts — redact even in REAL-tagged images.
