Ad agencies charge $5,000–$25,000 to produce a campaign. You get a creative brief, a few rounds of revisions, and a PDF deck. What you actually needed was copy, angles, and assets — and now you can get all of that in an afternoon using Claude Code. Not a simplified chat interface. The actual tool, running locally, wired to your files and your brand.
This guide shows you exactly how to use Claude Code to build a repeatable ad creative system: from raw product information to polished ad copy, angle matrices, and headline variations — the kind of output that used to require a senior copywriter and a strategist on retainer.
Why Claude Code and Not Just Claude.ai
Claude.ai is a chat window. Claude Code is a local agent that reads your files, runs commands, writes to disk, and operates iteratively without you babysitting every step. That difference matters for ad creative work.
When you use Claude.ai, you paste things in, get things out, and manually track versions. When you use Claude Code, you point it at a folder — your brand docs, your product one-pager, your past winning ads — and it uses all of that as live context. It can generate 40 headline variations, sort them by angle, and write them to a file you can immediately hand to your media buyer.
It also loops. You can give it a review instruction and it will critique its own output, flag weak hooks, and rewrite them before you see the final result. That self-editing loop is where the agency-grade quality comes from.
Set Up Your Creative Context Folder
Before you write a single prompt, build the folder Claude Code will read from. Think of this as the brief packet you'd hand a new creative director. Put it somewhere clean, like ~/ad-creative/[client-name]/.
What goes in the folder:
- brand-voice.md — your tone descriptors, words you use, words you avoid, and 3–5 examples of copy that already sounds right
- product.md — what the product does, the mechanism, the outcome, the price, and the offer structure
- audience.md — who the buyer is, what they're afraid of, what they want, what has not worked for them before
- winners.md — paste in 3–5 ads that have converted, even if they're from competitors or analogous markets
- blocklist.md — phrases and claims you will never use (compliance language, overused clichés, anything legal has flagged)
You do not need these to be polished documents. Rough notes work. The point is giving Claude Code raw material to reason from. A single paragraph of real audience insight will outperform a generic prompt every time.
The Angle Matrix — Your First Real Output
The most expensive thing agencies do is angle development — figuring out which emotional, rational, or situational frame will make someone stop scrolling. Claude Code can build you a full angle matrix in one pass.
Open Claude Code in your creative context folder and run this prompt:
Read all files in this folder. Based on the product, audience, and voice context, generate an angle matrix with 8 distinct angles for paid social ads. For each angle:
- Name the angle (e.g. "Fear of Status Quo", "Speed to Result", "Credibility Transfer")
- Write a one-sentence rationale for why it works for this specific audience
- Write 3 hook variations (first line of the ad) for that angle
- Flag which angles are safest for cold traffic vs. retargeting
Write the output to angle-matrix.md.You now have 24 hooks across 8 angles, organized and file-saved. A freelance strategist would charge $800–$2,000 for this deliverable and take three days. You just got it in under ten minutes.
From here, pick the two or three angles that match your test budget and move into body copy.
Writing Full Ad Copy — with Self-Review Built In
Claude Code's self-editing loop is the feature most people skip and it is the most valuable one. You can instruct it to write, then critique, then rewrite — all in a single command.
Using the angle-matrix.md you just wrote, pick the top 3 angles marked best for cold traffic.
For each angle, write a complete Facebook/Instagram ad in this structure:
- Hook (1–2 sentences, matches the hook variations already written)
- Body (3–5 sentences: problem, mechanism, proof point, outcome)
- CTA (1 sentence, direct, no soft language)
After writing all three ads:
1. Review each ad against the brand-voice.md and blocklist.md
2. Flag any lines that violate voice or blocklist rules
3. Rewrite those lines
4. Output the final clean versions to cold-ads-v1.md
Do not show me the drafts. Only show me the reviewed final versions.That last instruction — 'do not show me the drafts' — forces Claude Code to complete the loop before surfacing anything. You get reviewed output, not a first pass you have to clean yourself.
Run the same pattern for retargeting copy, email subject lines, or landing page headlines. Same folder, same context, different output instruction.
Iteration at Scale — Testing 40 Headlines in One Session
Media buyers need options. The more headline variations you can test, the faster you find the winner. Claude Code can generate a full headline bank in one session — organized by angle, character count, and format type.
Using the angle matrix and audience file, generate 40 ad headlines. Organize them in a table with these columns:
- Headline text
- Angle (from the angle matrix)
- Format (question / statement / list tease / social proof / number-led)
- Character count
- Cold or retargeting traffic
Apply the brand voice and blocklist before finalizing.
Write the output to headline-bank.md as a markdown table.Hand headline-bank.md to your media buyer or upload it directly into your ad platform. You now have a structured testing plan, not a random pile of copy.
When you find a winner, add it to winners.md with a note on what worked. Every winning ad you log makes the next creative session sharper because Claude Code will read it and pattern-match against real performance.
Scaling Across Clients or Products
If you are running ads for multiple clients or multiple products, the folder system is how you stay organized. Each client gets their own subfolder with their own context files. Claude Code reads only what you point it at, so there is no cross-contamination of voice or positioning.
You can also templatize the whole workflow. Write your prompts into a prompts.md file inside each client folder. When a new campaign cycle starts, open Claude Code, run the prompts in sequence, and get a full creative package in under an hour.
The practical workflow:
- Update product.md and audience.md with any new offer details or research
- Run the angle matrix prompt — get fresh angles that reflect the current offer
- Run the cold ad copy prompt with self-review — get 3 clean full ads
- Run the headline bank prompt — get 40 organized variations
- Export to your media buyer or directly into your ad manager
- After results come in, log winners to winners.md to compound the context
That is a repeatable creative system. Not a one-off prompt. Every client and every product gets the same structured approach, and the quality improves the longer you run it because the winners file grows.
What This Does Not Replace
Claude Code generates copy. It does not buy media, manage creative fatigue, or tell you why a campaign underperformed. The strategic layer — what to test first, how to read results, when to kill an angle — still requires judgment.
It also does not replace good source material. If your product.md is vague and your audience.md is generic, the output will reflect that. Garbage in, polished garbage out. Spend time on the context files. That is where the real work is, and it is a fraction of the time you used to spend writing briefs for an agency.
Claude Code is not magic. It is leverage. The sharper your context files, the better the output. Build those files once, refine them as you learn, and you have a creative infrastructure that outpaces any agency retainer.
The agencies are not gone. But their pricing power over repeatable, brief-driven creative work is. You can now produce that work yourself — faster, cheaper, and with a feedback loop that compounds every campaign cycle.