An AI copywriting skill that works like an ad agency

Good ads rarely come from one person writing one good sentence. A planner reframes the brief, a creative director reviews it, several creative teams write a lot of lines, the CD cuts them down more than once, and the client picks. This Claude Code skill runs that workflow with one AI agent per role and hands you a headline, two alternates and a record of how they got there.

/plugin marketplace add 5hyuns/ad-copy-agency-skill

/plugin install copy-masters@ad-copy-agency-skill

3 teams, round 1

CD round 1 keeps

CD round 2 pool

CD shortlist

Client's final

Line counts from one English run on a fictional US smart dog bowl. The round 2 pool adds rewritten lines, defended lines and approved customer quotes to the 13 survivors. It writes English copy for the US market and Korean copy for the Korean market.

WORKFLOW

Ten steps, a different agent for each role

Roles never see each other's reasoning, only the files the previous step left behind. Teams never pick their own lines, and the CD judges lines with the team names hidden. In Turnbull & Wheeler's study of London agencies, only two of 24 steps produced ideas; the rest were about agreeing on the problem and checking the work. This skill spends most of its steps the same way.

Client brief 01-client-brief.md
The inputs, written down as given, with no interpretation.
Raw material 02-raw.md
Product facts, what real people wrote, and competitor copy, collected verbatim with sources.
Planner 03-creative-brief.md
Reframes the task as "what this ad must say" by changing the audience, changing the kind of problem, or finding the empty spot competitors leave. Points to the one scene or quote the reframe rests on.
Creative surgery 04-surgery.md
The CD reviews the brief before any team sees it. If it goes back, the planner revises once.
Three creative teams, in sequence 05-team-A/B/C.md
Each team opens 3 to 5 routes and writes 33 lines. Later teams are told which routes are already taken and look elsewhere.
CD review, round 1 06-cd-review-1.md
Criteria first, then a keep or kill on every line, with a short reason for each kill. 3 to 5 routes survive with feedback.
Rework and defense 07-team-A/B/C.md
Teams rewrite their surviving routes and may defend one killed line. Approved customer quotes enter the pool verbatim.
CD review, round 2 08-cd-review-2.md
Rules on each defense and shortlists about five lines for the client.
Client review 09-client-review.md
Fact and advertising-law check first, then one final line and two alternates. Unverified facts are flagged for the client to confirm.
Handoff 10-final.md · 10-process-check.md
The path each chosen line took, and a table checking that every step ran as the agency workflow expects.

FROM CD ROUND 1

Every line gets a verdict and a reason

Real verdicts from the English run. The CD does not pick by taste or by counting pros and cons; it checks each line against criteria written down before reading. Length is never a reason.

He looks up. There's more.keepA scene only this product can produce: the fast eater looks up at a bowl he thinks is finished and finds more.

The gulp is the cue.keepTurns the problem into the switch, from the dog's side.

Kibble on a drip.kill"Drip" reads as medical.

The end of dinner keeps not arriving.killReads as a meal that never ends.

You can stop doling out dinner by the handful.killOwners portioning by hand is not in the material.

INSTALL

Two commands inside Claude Code

Install it as a plugin from this repository. Start a new session (or run /reload-plugins) and the skill is /copy-masters:copy-masters. Update with /plugin marketplace update ad-copy-agency-skill.

/plugin marketplace add 5hyuns/ad-copy-agency-skill

/plugin install copy-masters@ad-copy-agency-skill

Or clone it into a skills folder

Put it in your user skills folder to use it everywhere, or in a project's .claude/skills/ to use it there only. Keep the folder name copy-masters; the skill is then /copy-masters and you update with git pull.

All projects (macOS · Linux · Git Bash)

git clone https://github.com/5hyuns/ad-copy-agency-skill.git ~/.claude/skills/copy-masters

All projects (Windows PowerShell)

git clone https://github.com/5hyuns/ad-copy-agency-skill.git "$HOME\.claude\skills\copy-masters"

One project

git clone https://github.com/5hyuns/ad-copy-agency-skill.git .claude/skills/copy-masters

Requires Claude Code and Python 3 (the helper scripts use only the standard library). Raw material collection uses a web search tool when one is available and falls back to your own material. Start a new session and copy-masters appears in the skill list.

USAGE

Tell it what, for whom, and where it runs

Call it with the slash command or just ask for ad copy or headlines. If a required input is missing it asks first; it never guesses the mode, channel or target. Set Copy language: en for English copy (the default is Korean); the market then defaults to us, so the client checks against US channel specs and FTC rules.

/copy-masters:copy-masters
Subject: SlowBowl, a smart dog bowl that releases kibble in small portions when it notices a dog gulping
Business goal: more first orders from Instagram ads
Target: US dog owners, 25 to 45, whose dog eats too fast
Mode: performance
Channel: social feed
Brand: Tumbleweed Pets
Must-haves: not a veterinary medical device; no health claims
Your material: ./material.md
Copy language: en

Input	Required	Value
Subject	yes	What is being advertised
Business goal	yes	One short line
Target	yes	Who the ad speaks to
Mode	yes	`performance` (clicks, conversions) or `brand` (memory, attitude)
Channel	yes	`social feed`, `search ads`, `landing page` or `video script`
Brand	yes	Names the output folder and learning log
Product truth	no	What the company itself says is the core of the product
Must-haves	no	What must or must not appear, tone rules
Industry	no	Adds industry-specific regulation checks
Your material	no	Decks, reviews, interviews. Approved customer quotes reach the shortlist verbatim
Copy language	no	`en` for English copy; Korean if not set
Market	no	`us` or `kr`: whose channel specs and ad law the client checks against. Follows the copy language if not set

Output

One file per step under copy-runs/<brand>/<date>-<topic>/ in your working directory. At the end you get the final line and two alternates, the planner's reframe, line counts per step, any lines the fact check caught, and the facts the client still needs to confirm.

Time and cost

The Korean demo run used 13 sub-agents and took about 20 minutes, about 1.04M sub-agent tokens in total. It suits work where you want to follow the process or where fact checking matters more than speed.

DEMO

One full run on a fictional smart dog bowl

CLIENT'S FINAL “Two of you listening at dinner. Only one of you can hold the kibble back.” Alternates: "He can't inhale what isn't in the bowl yet." and "He doesn't have to change. The bowl does."

The planner changed the kind of problem: owners already know to slow their dog down, and every bowl they know does it with an obstacle in a full bowl. The evidence was one review from an owner still listening for chewing at dinner. The CD sent the first brief back once. Three teams wrote 99 lines from 15 routes. A full line-by-line record exists for a Korean run.

Open the Korean run record

LIMITS

No evidence it beats a single prompt

Compared head to head with a single prompt given the same raw material, the skill's final lines won 0.37 of the time on average. Treat its output as about as good as, or worse than, one well-written prompt.
The reason to use it is the process. You can follow each step, and the brief check and client review catch factual errors a single prompt repeated across runs, such as unsupported claims or edited customer quotes.
The CD and the client are the same model. Research suggests model judges match people poorly on originality and surprise, and re-running the same pool can change the final line.
English copy has been tested once, on one fictional US product. The run worked end to end, but its three picks won 0.39 of blind comparisons against two single-prompt runs. The client's longest pick lost every pair while an alternate was the second strongest line. Given the same shortlist nine more times, the client picked that 43-character alternate every time (0.67 against the single-prompt lines), so 0.39 leans on one pick. Lines grow longer at team rework, not at the CD's shortlist. The US checklist did catch a customer quote under the FTC's typical-results rule (16 CFR 255.2).
It produces headlines. Visuals, subheads and body copy are out of scope.
Its only material is web sources and what you give it. Reframes that come from product demos, focus groups or a remark in a meeting room are out of reach.