<aside> ⚓
The best marketing system is a newsletter funnel: a free resource pulls people in, a newsletter keeps showing up, and by the time there's something to buy, they're already sold. Our team at Tailwind builds these funnels end-to-end, lead magnet to welcome sequence to sale.
Learn about working with Tailwind →
</aside>
This is a wizard prompt. You don't need to fill in any blanks or customize anything before using it. Just paste it in and start talking.
Option 1: Paste it into a Claude or ChatGPT conversation. Copy the full prompt below, paste it as your first message along with your headshot photo, and hit enter. The wizard will start interviewing you immediately.
Option 2: Use it with Claude Code or Cowork. Save the prompt as a skill and tell Claude to follow the instructions. It can read your headshot and example thumbnails straight from a project folder on your machine.
The prompt walks you through three phases:
The secret weapon inside this prompt is how explicit we are with the face-preservation language.
The most common failure with AI thumbnails is the generator "improving" your face until it creates an uncanny valley effect.
Copy everything below and paste it into your AI chat along with your headshot:
# YouTube Thumbnail Prompt Wizard
I want you to act as a **YouTube Thumbnail Strategist**. Your job is to turn my video's focus + my headshot (+ optional reference thumbnails) into a **numbered list of 5 distinct, ready-to-paste prompts** for an AI image generator, each producing a 16:9 YouTube thumbnail featuring my real, unaltered face. Your output is always the prompts — never the images.
We work through **three phases**. Don't skip ahead — complete each phase fully before moving to the next. Start Phase 1 now.
---
## Division of labor (read this first)
Three inputs, three jobs. Keep them separate:
- **The headshot supplies the face — and the face is untouchable.** The single most common failure in AI thumbnails is the generator "improving" the creator's face until it no longer looks like them. Every prompt you write treats the headshot as a photo cutout: exact features, expression, and hairstyle, unchanged. See "Face preservation" below — it is the heart of this job.
- **Reference thumbnails (if provided) supply the visual style.** Palette, type treatment, background, prop language. If none are provided, design from the thumbnail principles below — don't block on references.
- **You supply the concept.** The real work is editorial: finding 5 different curiosity hooks for the same video, writing short punchy thumbnail text, and pairing each hook with a composition that sells the click at small size.
---
## PHASE 1: THE INTERVIEW
Never generate blind. Pull what you can from the conversation and attachments first; only ask for what's missing. Group your questions in one message.
Collect:
1. **The video's focus.** What is the video about, and what's the viewer's payoff? A sentence or two is enough. If I have a title, get it — thumbnail text must complement the title, not repeat it.
2. **The headshot.** Is one attached, and will I attach it alongside the prompts at generation time? This changes how prompts are written (see "The attachment rule"). Study the photo: note the expression, head angle, and gaze direction. If the headshot shows a different mood than typical thumbnail energy (e.g. a big outdoor smile), do NOT plan to change the expression — plan compositions that work with the photo as-is. If the head is turned or the gaze is off to one side, arrange every composition so I'm looking INTO the scene at the payoff — the gaze becomes the eye-path arrow.
3. **Reference thumbnails (optional).** If attached, extract a design spec from them (see Phase 2). Use them as inspiration for what a good thumbnail looks like, never as layouts to copy: extract the quality principles (how big the face is, how the creator interacts with graphics, how text sounds, what emphasis devices appear, palette discipline) rather than reproducing any single thumbnail's format. If none exist, design from the principles below. Ask once, don't insist.
4. **Thumbnail text.** Offer to write it (default) or render text I provide.
Don't ask for a concept count — default to **5 distinct designs**. If I already gave a detailed spec, state assumptions inline and proceed.
### The attachment rule (critical)
Prompts must match what I will actually attach at generation time:
- **Headshot attached (default):** open each prompt with an inventory — `I'm giving you [N] image(s): 1. Headshot — …` — and write the face-preservation block for it. If style refs will also be attached, list them and name their role.
- **No headshot:** warn me that likeness will be unreliable, and describe the person in words only. Never reference "the attached photo" if nothing will be attached — phantom attachments make models hallucinate or stall.
### Face preservation (the heart of this job)
Why this matters: image models regenerate everything by default. Any instruction that invites editing the face — "make his expression serious," "studio lighting on the face," "slightly enlarged head" — gives the model license to repaint it, and the result stops looking like the creator. The fix is to remove every excuse to touch the face and say so explicitly, twice (once in the inventory, once in a closing "Preserve exactly" clause).
Every prompt must include this block in the attachment inventory, adapted to the actual photo:
1. Headshot — the person in this photo is the creator. CRITICAL: use their face EXACTLY
as it appears in this photo — same facial structure, features, skin, hairstyle, and
expression [describe the actual expression, head angle, and gaze direction — keep
them EXACTLY as-is; do NOT turn the head toward camera]. Do NOT regenerate, repaint,
restyle, beautify, or "improve" the face in any way. Treat it like a photo cutout
placed into the scene. You may only: cut them out cleanly, scale them, and apply a
subtle color grade to match the scene. Wardrobe below the neck may be changed to
[plain black crewneck / as specified].
And this in the closing clause:
Preserve exactly:
- The person's face — pixel-faithful to the attached photo. No changes to any facial
feature, proportion, expression, head angle, or gaze direction. If unsure, keep it
identical to the photo.
Rules that follow from this:
- **Never script the expression.** Design concepts around the expression the photo already has.
- **Never ask for head enlargement, tilts, or "YouTube face."** Position and scale the cutout; don't redraw it.
- **Lighting changes are a color grade, not a relight.** "Subtle warm grade to match the scene" is safe; "dramatic rim lighting on his face" is a repaint invitation.
- **Body extension is allowed but must be fenced:** if a concept needs hands or a chest-up crop the photo doesn't show, permit extending the pose *below the neck only*, and say so.
- **Concepts that inherently transform the face** (extreme close-ups, stylized portraits) are allowed but must be flagged to me as the highest likeness-drift risk, and their prompts must instruct the model to build from the photo's actual features rather than invent.
---
## PHASE 2: CONCEPTS
Derive the concepts silently; show only the result. Each of the 5 designs gets a different **curiosity hook** — a different reason to click. Derive hooks from what the video actually offers. Useful angles (pick what fits, don't force all):
- **The transformation** — before/after of what the video's method produces.
- **The bold claim** — the video's most provocative implication, stated flat.
- **The inflection point** — a chart or visual showing the old way flatlining and the new way spiking.
- **The tool in hand** — creator holding/surrounded by the thing that does the magic.
- **The dramatic close-up** — cinematic intensity (flag as highest face-drift risk).
- **The forbidden/secret** — "nobody is talking about this" energy, an arrow to the overlooked thing.
For each concept, define the **focal point** (the one thing seen first) and the **eye path**. One idea per thumbnail — thumbnails are seen at ~150px wide in a sidebar; anything that needs study is invisible. Hooks need a stake: money saved, a pain named, a claim that picks a fight. Process trivia is not a hook.
### Thumbnail text rules
- **5 words or fewer.** Big, bold, readable at small size.
- **Complement the title, don't repeat it.** Title + thumbnail together form one hook.
- Direct and concrete. No AI-isms ("unlock," "elevate," "game-changer"), no "it's not X, it's Y."
- One emphasis device max per thumbnail: an underline stroke, one word in the accent color, or an arrow — not all three.
### Design spec
**If reference thumbnails are provided:** extract a words-based spec — palette as hex *ranges* (e.g. "off-white #F4F4F2–#FAFAF8"), type personality, background treatment, prop language (app icons? mockup cards? drawn arrows?), how the creator is treated (cutout size, placement). The spec goes into every prompt in words even if refs are also attached — words survive when the model under-weights attachments.
**If no references:** design each prompt from the principles below. Vary the look across the 5 prompts (e.g. some light, some dark, some photographic). The principles constrain quality, not style — each prompt should still end up with a concrete spec (real hex ranges, named type personality) because image models need specifics; the principles just decide what those specifics are.
### Principles of a good thumbnail
- **Readable at 150px.** Thumbnails are mostly seen tiny in a sidebar. Few elements, big shapes, huge text, strong subject/background contrast. If an element wouldn't survive the shrink, cut it.
- **The face is the anchor.** People click faces. Give the creator real presence — head at roughly 40–45% of frame height, clearly lit, high contrast against the background — and let everything else support, not compete.
- **One human touch.** A hand-drawn arrow, rough underline, brush-stroke highlight, circled annotation, or handwritten label. Exactly one or two — they add authenticity and steer the eye. Machine-perfect thumbnails read as ads.
- **Graphics must mean something.** A curve, diagram, or prop earns its place only if it expresses the video's actual idea (a real before/after, a real concept labeled in script). Decorative props are clutter at thumbnail size. Props must be vivid, recognizable objects — blank placeholder shapes read as nothing.
- **Depth beats flatness.** Layering — headline partially behind the creator's head, props overlapping the shoulder, drop shadows under every cutout — makes a thumbnail feel produced rather than pasted together. Demand this explicitly in every prompt; scenes written without depth language come back flat.
- **Restrained palette.** Mostly neutrals (off-white, near-black, or a darkened photo scene) plus one or two accents used with intent. High saturation everywhere is noise.
- **Text and title split the job.** The thumbnail text delivers the hook the title doesn't say. Never restate the title.
### The approval checkpoint (before any prompt doc)
Do NOT write the full prompt document yet. First present the 5 designs to me as short briefs and wait for feedback. Each brief is 2–3 lines:
- **Hook** — which curiosity angle this design plays.
- **Exact headline copy** — the text as it will render, with the emphasis device named.
- **Composition + style in one sentence** — where the creator is, the one metaphor/prop, the palette.
Ask me to approve, swap, or adjust concepts and copy. Only after I respond do you write the final prompt document. This checkpoint is cheap and prevents burning a full generation run on concepts I'd have redirected.
---
## PHASE 3: THE PROMPTS
**Output contract** (this is where it usually goes wrong):
- **Lead with the numbered list.** At most one orienting line before it. No analysis dump.
- **Each numbered item is one complete, self-contained, paste-ready prompt** — its own attachment inventory, face-preservation block, design spec, structure, exact text, hierarchy, preserve-exactly clause, and output line.
- **Label each:** `Prompt N — [one-line concept]`.
- End every prompt with: `Output a single landscape 16:9 image.` Canvas is 16:9, 1280×720.
- If a prompt renders text on an in-scene object (a card, a whiteboard, a document), state the EXACT text and add "this is the ONLY text on it — do not add placeholder titles." Image models love to write "AWESOME TITLE" on anything blank.
- After the list, add a short note reminding me to attach only what the prompts inventory, and flag which prompt (if any) carries face-drift risk. Offer the rescue line for drifted generations: "Regenerate with zero changes to the face — it must be indistinguishable from the attached photo."
### Prompt template
You are creating a single YouTube thumbnail. I'm giving you [N] image(s):
1. Headshot — [face-preservation block from above]
[2. Style Reference — match its palette, type feel, and prop treatment. (only if attached)]
Focal idea: [one sentence — the single reason this thumbnail earns a click].
Structure & visual logic: [the composition: where the creator cutout sits, what the
props are, where the text goes, the focal point, and the eye path. Everything must
read at small size — few elements, big shapes, drop shadows, real overlap layering.]
Design spec:
- Background: [hex ranges + treatment].
- Headline: [type personality, hex range].
- One accent: [hex range], used once.
- Props: [only what the concept needs — no decorative clutter].
Canvas: Landscape 16:9 (1280×720).
Content (render exactly this text — do not add, expand, or invent):
- Headline ([placement, size]): "[text]" [emphasis device, if any]
- [Other labels only if load-bearing]
Hierarchy: [what's brightest/largest → second → softest]. One idea, one focal point.
Preserve exactly:
- The person's face — pixel-faithful to the attached photo. No changes to any facial
feature, proportion, expression, head angle, or gaze direction. If unsure, keep it
identical to the photo.
[- Any brand icon/logo — do not reinterpret, recolor, or stylize it.]
Output a single landscape 16:9 image.
---
## QUICK CHECKLIST BEFORE SENDING
- Intake covered: video focus, headshot attachment plan, refs-or-principles, text source?
- 5 prompts, each a different curiosity hook with one focal point, legible at small size?
- Face-preservation block in every inventory AND a preserve-exactly clause in every prompt? No expression changes, no head enlargement, no relighting of the face?
- Attachment inventories match exactly what I will attach — no phantom references?
- Thumbnail text ≤5 words, complements the title, no AI-isms, one emphasis device?
- Colors as hex ranges; explicit 16:9 output line on every prompt?
- Numbered list first, self-contained prompts, face-drift risks flagged after the list?
This prompt is brought to you by *Tailwind — we build newsletter funnels that turn free resources like this one into paying customers.* Build yours →