agentleFS
bartlomiejborzuckibartlomiejborzucki/meta-ads-agent/skills/meta-ads-creative/SKILL.md

meta-ads-creative

Skill0 starsChanged 44 days agoMITView on GitHub

Generate genuinely different creative angles, hooks, primary text, headlines, descriptions, and CTAs for Meta (Facebook/Instagram) ads, constrained by the brand's voice, banned phrases, and claims policy. Use for "write me three angles", "new ad copy", "variations on the winning ad", "creative concepts for this offer", or when a campaign build needs copy.

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 6 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    1. Meta's official Ads MCP - the execution layer: Look for tools named ads_* in this session.

    line 22
  2. Very specific:

    If voice.md exists, it governs.

    line 59
  3. Distinctive:

    If the offer has no proof, write copy that does not need any.

    line 74
  4. Distinctive:

    Not distinct - one angle, four rewordings:

    line 115
  5. Distinctive:

    If you produce the second kind, say so: *"these are four executions of one angle, not four angles."* The plan validator warns when every …

    line 122
  6. Distinctive:

    Always state what proof each angle relies on.

    line 164
  7. Distinctive:

    When the user wants to promote a post that already exists, do not rewrite it.

    line 180
  8. Distinctive:

    Do not copy a competitor's text or creative.

    line 191

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Creative: angles and copy

Three genuinely different reasons someone would care beat thirty rewordings of one. Read meta-ads-core for the safety model; this skill writes, it does not create objects.

<!-- shared:preflight start - generated from packaging/shared/preflight.md by scripts/sync_skill_blocks.py - edit there -->

Preflight: two checks, kept separate

Independent questions; never let one answer stand in for the other.

1. Meta's official Ads MCP - the execution layer. Look for tools named ads_* in this session. With none, nothing here can reach an account: say so and help the user connect it - connect-meta-mcp.md under references/ in the meta-ads-core skill, or https://github.com/bartlomiejborzucki/meta-ads-agent/blob/master/skills/meta-ads-core/references/connect-meta-mcp.md. The local CLI does not replace it.

2. The local meta-ads-agent CLI - optional, usually absent. Assume it is missing until this succeeds:

meta-ads-agent --version

"command not found" is the normal answer, not a fault. Show no meta-ads-agent ... command before that probe has succeeded - a command that fails at the user's prompt costs more than the step it saves. Say "that step needs the optional CLI, which is not installed here" and carry on.

The MCP alone covers audits, reporting, Ad Library research, previews, tracking, creative, optimisation and the changes that follow; .meta-ads/ files can be written directly from the templates in meta-ads-core's assets/. Only these need the CLI: plan validation (so campaign builds), local image and video upload, video / existing-post / multi-variant / carousel creatives, deletion, and the report arithmetic (which has a by-hand route).

<!-- shared:preflight end -->

Read the constraints first

SourceWhat it gives you
.meta-ads/voice.mdtone, vocabulary, examples the user likes and dislikes
.meta-ads/brand.yamlbanned_phrases (hard bans), claims_policy (binding), profile
.meta-ads/offers/<slug>.yamlaudience, problem, outcome, proof, restrictions, CTA
ads_library_searchwhat the category is already saying - via meta-ads-research

If voice.md exists, it governs. If it does not, ask for two or three ads or emails the user likes; examples calibrate tone far faster than adjectives.

Never invent evidence

You may cite only proof that exists in the offer brief or that the user supplied. No invented:

  • results or percentages ("cut reporting time by 80%")
  • customer counts ("trusted by 10,000 teams")
  • testimonials or named customers
  • scientific or medical claims
  • product capabilities the product does not have
  • awards, certifications, or rankings

If the offer has no proof, write copy that does not need any. Specificity can come from the problem rather than from a number - "the Friday afternoon that disappears into the weekly report" is concrete and requires no evidence.

A fabricated ad claim is a legal and reputational problem for the user, not a creative flourish. If you find yourself wanting a statistic, ask for one.

claims_policy in brand.yaml is binding. restrictions in the offer brief is binding. banned_phrases must not appear verbatim.

Structure

Separate the layers. They are edited independently and they fail independently.

LayerWhat it is
Anglethe reason someone would care. The concept.
Hookthe first line that earns the second.
Primary textthe body. Shown truncated in most placements.
Headlineshort, below the image. Often the only thing read.
Descriptionoptional, shown in some placements only.
CTAMeta's enum. Discover valid values with ads_get_field_context.

Front-load the primary text: Feed truncates it, and the "see more" cut lands early. Say the thing that matters before the fold.

Angles have to actually differ

An angle is a different reason to care, not a different sentence about the same reason.

Distinct angles for one offer:

AngleThe reason
Time recoveredthe hours the work currently costs
Error riskwhat happens when a manual number is wrong
Being the bottleneckthe person everyone waits on
Cost of the status quowhat the manual process costs annually

Not distinct - one angle, four rewordings:

  • "Save time on reporting"
  • "Cut hours from your reporting"
  • "Stop wasting time on reports"
  • "Reporting shouldn't take hours"

If you produce the second kind, say so: "these are four executions of one angle, not four angles." The plan validator warns when every variant shares an angle, and calling rewordings a creative test produces a test that cannot teach anything.

For a genuine test, vary one layer at a time. Different angles with the same image tests messaging. The same copy on different images tests visuals. Changing both tells you which combination won and nothing about why.

How many

ProfileAngles
conservative2
standard3
experimental4-5

Fewer, better-differentiated angles beat more variants. Each angle needs enough budget to produce a readable result - see the volume floors in brand.yaml.thresholds. Splitting a small budget across five angles produces five inconclusive results.

Output format

Give the user something they can review, then hand the chosen variants to meta-ads-campaign as plan input.

ANGLE 1 — Time recovered
  Hook           Friday afternoon, gone again.
  Primary text   ...
  Headline       Stop rebuilding the weekly report
  Description    ...
  CTA            SIGN_UP
  Proof used     none - this angle asserts nothing measurable
  Test intent    does the time framing beat the risk framing

ANGLE 2 — Error risk
  ...
  Proof used     "340 attendees at the last session" (offer brief)

Always state what proof each angle relies on. It makes an unsupported claim visible before it ships, and it shows the user exactly which claims they are vouching for.

Placement matters to copy

  • Feed - primary text truncates early. Headline carries weight.
  • Stories and Reels - vertical, sound often off, text overlay competes with UI. Short.
  • Right column - image and headline only, tiny.

meta-ads-preview will show you the actual truncation. Do not guess at it - count characters if you must, but the preview is the evidence.

Existing posts

When the user wants to promote a post that already exists, do not rewrite it. Promoting the real post preserves its likes, comments, and shares; recreating the content as a new dark post starts from zero. If they ask you to "make an ad from this post", find out which they mean.

Competitor research

meta-ads-research and ads_library_search show what a category is saying. Extract themes: angles in use, offers, formats, positioning, what nobody is saying.

Do not copy a competitor's text or creative. Beyond the legal problem, ads that sound like the category perform like the category - the gap is usually the opportunity.

More: references/angles-and-copy.md.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.