Claude Projects is the best AI creative team we've ever hired

Building a context library

Each project is set up with three core documents:

These three documents are connected to Claude along with a carefully written system prompt that points to all the content.

Practical setup

• Start with 3-5 core areas, such as headlines, scripts, ideas/iterations, and rewrites.
• Update the documents regularly based on poor outputs as well. When Claude gives a bad response, upload that response and explain what made it bad.

Impact on production

• Significantly increases both the volume and quality of headlines and other ad creative.
• Allows small teams to produce at the same pace as larger ones, without having to hire as many people.
• Results are far better than using Claude or GPT without context.

A few quick notes

• Use Claude instead of ChatGPT. It's better at copywriting.
• I could build these documents for you, but it's more valuable for you to train the AI on what actually works for your specific situation.

Hook generation in a Claude Project

This project takes emotional quotes from your audience research and generates 20+ scroll-stopping hooks tailored to the brand, for both UGC scripts and static ads.

Context document for winning hook writing

Paste this into Claude to save and use as context in the project, following the setup guide above.

Domain Context: Static Headlines
Context Document: Writing High-Converting Static Headlines & Image Ads

1. The Role of Headlines in Static Ads
- First job of a headline: Stop the scroll. Attention is the currency.
- Second job: Spark curiosity, relevance, or emotion strong enough to make the user process the rest of the ad.
- Third job: Tie directly to the value prop (clear, specific, visual, and ownable).

Headlines in static ads must do 3 things fast:
1. Communicate the core USP or benefit.
2. Resonate emotionally with the audience's desire or pain point.
3. Be unmistakably different from generic "me-too" claims.

2. Emotional Triggers That Headlines Should Leverage
- Pain Point Relief: "No more chasing invoices."
- Aspiration / Gain: "Turn spare time into income."
- Simplicity / Ease: "From signup to first sale in 5 minutes."
- Fear of Missing Out: "Last chance to lock today's price."
- Authority & Proof: "Trusted by 10,000+ small business owners."
- Curiosity Gap: "The trick gyms don't want you to know."

3. Static Ad Visual Principles (Meta Image Ads)
- Headline-Visual Fit: The visual must reinforce or dramatize the headline claim. Example: Headline: "Save hours every week." Visual: Clock cut in half.
- USP Placement: Top-left or headline text must carry the main USP (it's where the eye goes first). Avoid clutter - prioritize one message per creative.
- Branding: Subtle but present (logo, colors, typography). Don't overpower the message, but ensure recognition.
- Legibility: Short, bold text, high contrast. Static ads are consumed in <2s.
- Product in Use: Show the product solving the pain point, not just isolated packshots.
- Visual Anchors: Icons, arrows, or symbols to draw the eye to the CTA or headline.
- Social Proof & Trust: Layer reviews, ratings, or customer count into the static where possible. Example: "Rated 4.9/5 by 12,000+ users."

4. How to Build a Strong Static Headline
1. Start with the USP: What is the single strongest promise or differentiator? Write it plainly first.
2. Apply the 3 Copy Tests: Can I visualize it? Can I falsify it? Can nobody else say this?
3. Stress-Test with an Emotional Lens: Does this relieve pain, unlock desire, create curiosity, or prove credibility?
4. Trim ruthlessly: Make it digestible in under 3 seconds of reading.

5. Static Ad Structure Framework
- Headline (Primary Text): Short, bold, visualizable, falsifiable, ownable.
- Supporting Visual: Shows the product solving the problem or delivering the promise.
- Secondary Copy (Optional): Expand on proof (social proof, stat, unique detail).
- CTA: Clear, action-driven ("Get started," "Shop now," "Book today").

6. Common Mistakes to Avoid
- Generic claims: "Best quality," "Save time."
- Overloading with multiple USPs.
- Tiny, hard-to-read text.
- Stocky visuals that don't connect to the headline.
- Copy-visual mismatch (e.g., a headline about speed with a static showing relaxation).

7. Checklist for Every Static Ad
- Headline communicates a clear, specific USP.
- Headline passes the visual / falsifiable / ownable test.
- Emotional trigger is present (pain, gain, ease, fear, proof, curiosity).
- Visual dramatizes or reinforces the headline.
- Product is visible or clearly implied.
- Legible and uncluttered.
- CTA is clear.
- Brand is recognizable.

Static ad design with Nano Banana Pro

This workflow generates AI images for static ads, product shots, lifestyle imagery and textures in seconds. Choose any of your favorite large language models first: ChatGPT Sora, Gemini, Nano Banana Pro, Perplexity, or Claude.

You will prompt the LLM to learn the structure for prompting on an AI image generation tool of your choice. Today, the best option is Google Nano Banana Pro. But know that tools one-up each other every 4 months or so, so always be on the lookout for the best tools every now and then to stay on top.

Prompt to train your own AI prompt engineer

If you want to save time, here's a basic prompt to train a new GPT, create a new Claude project, or create a Gem that acts as a prompt engineer:

You are "Google Nano Banana Pro Prompt Engineer" (aka Gemini 3 Pro Image / Nano Banana Pro). Your job: teach and generate best-in-class prompts specifically for Nano Banana Pro image generation + image editing, using ONLY official Google platform documentation as your source of truth whenever possible.

FIRST ACTION (always): compile a compact internal "Nano Banana Pro ruleset" by consulting official docs/pages (prefer these in order):
- Gemini API image generation docs: https://ai.google.dev/gemini-api/docs/image-generation
- Gemini 3 series docs: https://ai.google.dev/gemini-api/docs/gemini-3
- Vertex AI Gemini 3 Pro Image docs: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro-image
- Gemini app overview (image generation / Nano Banana Pro): https://gemini.google/overview/image-generation/

If browsing/tools are available, read them fully and extract: prompt best practices, editing instructions, negative prompt behavior, aspect ratio/output controls, limitations, safety constraints, and any platform-specific fields (e.g., aspect_ratio, image_config). If browsing/tools are NOT available, ask the user to paste relevant doc excerpts; otherwise proceed with clearly stated assumptions and keep advice conservative.

INTERACTION FLOW (every request):
1. Ask the minimum questions needed (max 3) ONLY if missing critical details:
- Generate or Edit?
- Subject + goal + usage (ads, product, thumbnail, storyboard, etc.)
- Must-keep / must-avoid
If the user provides a brief, do not ask more. Assume sensible defaults.
2. Explain the Nano Banana Pro prompt structure in a short checklist (bullets, not paragraphs).
3. Produce EXACTLY THREE distinct visual variations per brief (V1-V3). Each variation must be paste-ready.

OUTPUT FORMAT (strict):
For each variation output:
- "PROMPT" (one cohesive natural-language prompt, not keyword soup)
- "NEGATIVE" (one single-line negative prompt)
- "OPTIONAL CONFIG" (only if relevant: aspect ratio, resolution/output notes; use official field names if in API/Vertex context)

VARIATION REQUIREMENTS:
- V1: safest literal interpretation (high controllability)
- V2: stronger composition + lighting direction (still realistic/controllable)
- V3: bolder creative twist (different lens/angle/layout or style) while preserving the brief

NANO BANANA PRO PROMPT BLUEPRINT (use this order inside each PROMPT):
A. Intent + deliverable (what the image should accomplish)
B. Subject specifics (who/what, attributes, brand/product details if any)
C. Scene context (environment, props that matter)
D. Composition (shot type, framing, camera angle, distance, layout; where key elements go)
E. Lighting (type, direction, time, softness/hardness)
F. Materials/textures (surfaces, fabrics, reflections, micro-details)
G. Text (ONLY if needed): exact wording, font vibe, placement, legibility constraints
H. Hard constraints (must include / must not include; "do not change" clauses)
I. Realism controls (avoid artifacts, keep natural proportions)
J. If editing: explicitly lock what must stay identical (identity, pose, camera, background, product geometry) and specify ONLY the change

EDITING RULES (when the user is editing an existing image):
- Preserve identity, camera perspective, framing, background layout, and unmentioned details.
- Change only what the user requests.
- If the user requests multiple changes, sequence them as ordered steps inside the prompt.
- Add a fail-safe line: "If a requested change cannot be applied cleanly, leave the original unchanged."

TEACHING MODE:
After providing V1-V3, add a compact "How to iterate" section:
- 5 quick knobs (composition, lighting, texture, constraints, negatives) with example micro-edits.
- If the user says "make it better" or "looks AI," diagnose likely causes (5 bullets) and rewrite V1-V3 with stronger constraints and realism cues.

CONSTRAINTS:
- No fluff, no marketing chatter.
- Prefer explicit controllable language over vague style names.
- Always include a NEGATIVE line.
- Always deliver exactly three variations unless the user explicitly asks for a different number.

How to work with it day to day

After your customized "prompt engineer" AI is set up, use a tool like Wisperflow.io or MacWhisper to transcribe your ideas or briefs into text without having to type them out. This will save you a lot of time in the long run.

Your idea should describe, with as much detail as you can possibly imagine, the scene you are trying to create: the camera angle, the level of realism, which camera is being used, how many people will be in the scene, what ethnicity the people shown in the image would be, what location, time of day, season, how crowded the scene is going to be, the type of color grading the image is going to have, and so on.

How many prompts should you try? Start with 4 text prompts based on your brief, and test each prompt 3 times in your favorite image generation tool. A single image idea should therefore have about 12 image generations, in order to pick the best result for your ads.

So: you make the prompt AI engineer and explain what you need. It then makes a prompt for you. That prompt makes the image.

Keep going with Curve

Want to see what these workflows produce when they hit the ad account? Browse our cases, including Collagenhund and Sneglefellen.

If you want winning creative made for you, take a look at street interview ads and UGC, or read more guides on the blog.

Frequently asked questions