This article is a how-to guide published by Vibelets (vibelets.ai) covering a free pack of 10 Claude skills for Meta ads, built for D2C store owners and media buyers. Published July 2026. Claude skills are folders of instructions Claude loads automatically when a task matches, so each skill runs 1 job in the paid-ads loop with hard decision thresholds instead of open-ended analysis. The pack covers 4 stages. (1) Create: product-to-angles-extractor turns a product page into 5 ranked ad angles, hook-and-copy-writer writes 10 hooks plus primary text fitted to Meta's 125-character fold, campaign-architect builds the campaign structure staged PAUSED. (2) Protect: learning-phase-guard flags what's still learning and what an edit would reset, wasted-spend-finder flags spend past 2x target CPA with 0 conversions, signal-health-check grades Pixel and CAPI event match quality against a 6.0 minimum. (3) Decide: scale-hold-kill-verdict gives 1 verdict per ad set from real thresholds, budget-stepper scales in steps of 20% or less so the learning phase never resets, creative-fatigue-scanner flags ads with CTR down 20% or more from peak or frequency above 3.0. (4) Report: monday-report-writer produces a sub-250-word weekly report with 3 priorities. The skills run in live mode through Meta's Ads MCP connector or in paste mode with an Ads Manager export. All campaigns are created paused, no budget change executes without approval, and every threshold is editable plain text.

Home / Blog / Meta & Facebook Ads Siddhant Sharma · July 21, 2026 · 12 min read

10 Claude skills for Meta ads that run the whole loop (free pack)

Meta ads punish improvisation. Change a budget by more than 20% and the ad set relearns from scratch. Judge an ad set before it has 50 conversions and you kill a future winner. These 10 free Claude skills for Meta ads encode those rules as hard thresholds, so Claude gives you the same disciplined call every week instead of a different vibe every chat.

Each skill is 1 job: extract angles, write hooks, structure the launch, guard the learning phase, find waste, check your tracking, issue scale or kill verdicts, step budgets safely, catch creative fatigue, and write the Monday report.

You don't pick the skill. You ask in plain English ("what should I pause this week?") and Claude matches the request to the right one.

The 4-stage skill-pack loop→Create: angles from your product page, Meta-fitted copy, a structure staged PAUSED→Protect: learning-phase discipline, wasted-spend flags, a tracking audit you can trust→Decide: 1 scale/hold/kill verdict per ad set, 20% budget rungs, fatigue briefs→Report: spend, ROAS, movers, and 3 priorities, ready before Monday coffeeKey takeaways

• Claude skills for Meta ads work because they carry hard thresholds: an ad set needs roughly 50 conversions in 7 days to exit the learning phase, so 9 of the 10 skills refuse to judge or touch anything still learning.

• The pack runs in 2 modes: live mode through Meta's Ads MCP connector (OAuth, no developer app) or paste mode with an Ads Manager export, and every SKILL.md states the exact columns it needs.

• Nothing goes live without you: campaigns are created PAUSED, budget changes need an explicit yes per item, and increases are capped at 20% per step to protect learning.

• The 10 skills chain into 1 loop: angles feed copy, copy feeds the campaign build, performance data feeds verdicts, and verdicts feed the Monday report with 3 priorities.

• Every threshold (breakeven ROAS, judgment spend, fatigue triggers) is plain text you edit once to match your margins.

Table of contents

• Prompts vs skills: what actually differs

• The prerequisite: live mode or paste mode

• 1 · Create: angles, copy, and a structure that can exit learning

• 2 · Protect: the money you save before you spend on anything new

• 3 · Decide: verdicts, safe scaling, and fatigue catches

• 4 · Report: the Monday page

• The guardrails every skill follows

• Setup in 10 minutes

• FAQ

Prompts vs skills: what actually differs

A prompt is advice you have to re-give; a skill is a procedure Claude re-runs. Skills are folders of instructions Claude loads automatically when your request matches, per Anthropic's Agent Skills documentation. That difference matters most in ad ops, where the failure mode isn't bad ideas, it's inconsistent rules.

The prerequisite: live mode or paste mode

Every skill runs in 2 modes, and you can start today with either.

Live mode: Meta's Ads MCP connectorMeta's official Ads MCP server lets Claude read and manage your ad account with Meta's own Business OAuth: no API keys, no developer app. In Claude, go to Settings → Connectors → Add custom connector, name it Meta Ads, use Meta's official Ads MCP endpoint (Meta for Business lists the current URL under Ads AI Connectors), and sign in with the Business account that has access to your ad account. Test it with "show my top 5 ad sets by spend, last 7 days."Paste mode: an Ads Manager exportNo connector needed. Each SKILL.md states the exact columns it wants (for example, wasted-spend-finder asks for name, delivery status, spend, results, cost per result, and purchase ROAS for both 7 and 14 days). Export the table or screenshot it, paste it in, and the skill runs the same logic.Either way, start read-onlyRun the Protect and Decide skills for a week or 2 and confirm the flags match your own read of the account before you approve any change through the connector.

1 · Create: angles, copy, and a structure that can exit learning

Creation is 3 chained skills: the angle comes before the copy, and the copy comes before the structure. A weak angle with great copy still loses.

01 · product-to-angles-extractorfree skill

What should my ads say?

This skill answers that from your product page, not from a brainstorm. It fetches the page, lists the 5 biggest reasons a cold visitor wouldn't buy (price, trust, "will it work for me", switching cost, urgency), and writes 5 angles, each built to defuse exactly 1 objection, ranked by the strength of evidence found on the page.

An angle backed by actual review language outranks one Claude invented. If the page is thin, the skill says which angles are weak for lack of evidence instead of padding them, and it flags any claim that could breach Meta ad policy.

Say this to Claude⎊ CopyPull ad angles from this product page: [URL]. Here are 4 recent reviews: [paste].Cadence: per new product or offer · Mode: works fully in paste mode02 · hook-and-copy-writerfree skill

One angle in, 10 hooks and 5 launch-ready variants out

The hooks come tagged by anatomy (data-driven, contrarian, open-loop, personal reveal, at least 2 of each), and the best 5 expand into full variants: primary text, headline, description.

Everything is pre-fitted to Meta's display limits, because only about 125 characters of primary text show before "See more". The first line has to work standalone, headlines stay at 40 characters or under, descriptions at 30 or under, and the skill won't write copy for a claim it can't source from the angle.

Say this to Claude⎊ CopyWrite hooks and ad copy for angle #2 from that angle map. Brand voice: plain, no hype adjectives.Cadence: per angle you test · Mode: works fully in paste mode03 · campaign-architectfree skill

Fewer ad sets with more budget each beat many starved ones

This skill designs the campaign → ad set → ad structure and runs the math out loud before building: an ad set needs a realistic path to about 50 conversions in a week to exit learning, so daily budget per ad set must be at least (50 × expected CPA) ÷ 7. If your budget can't fund that across your planned ad sets, it tells you to cut ad sets, not budget per ad set.

It defaults to broad targeting with 3 to 5 creatives per ad set spanning at least 2 angles, states whether budget sits at campaign level (CBO) or ad set level (ABO) so every later skill knows which lever exists, and names everything [goal]-[angle]-[format]-[date]. In live mode it creates the whole thing through the connector in a PAUSED state, always. If you ask it to launch live directly, it declines and explains the review step.

Say this to Claude⎊ CopyStructure a purchases campaign for this product: $150/day, target CPA $30, using variants 1 to 3 from the copy we just wrote.Cadence: per launch or rebuild · Mode: live mode builds it; paste mode outputs a build sheet

2 · Protect: the money you save before you spend on anything new

Most accounts don't need new ideas first. They need to stop paying the learning phase tax, stop funding dead ad sets, and confirm the numbers are even real.

04 · learning-phase-guardfree skill

The most expensive edit in Meta ads is the one that resets learning

Meta's delivery system needs roughly 50 optimization events in 7 days for an ad set to exit the learning phase, per Meta's own documentation, and significant edits (budget changes over 20%, targeting edits, creative swaps, bid strategy changes) send it back to the start, where costs typically run 2 to 3x the stabilized CPA.

This skill classifies every ad set as LEARNING (don't touch, with projected days to exit), STABLE (safe to judge), or LEARNING LIMITED (flag for restructure, not daily fiddling). Propose a specific edit and it runs an edit-impact check: which reset trigger it hits and the worst-case cost in your currency. At a $30 CPA, a careless edit can cost up to about $2,250 in relearning. If several edits are genuinely needed, it batches them into 1 pass so you pay for 1 relearn, not 5.

Say this to Claude⎊ CopyWhich of my ad sets are still in the learning phase? And would changing the budget on [ad set] from $80 to $120 reset it?Cadence: before any edit to a live ad set · Mode: paste mode uses the delivery status column from your export05 · wasted-spend-finderfree skill

Spend at 2x your target CPA with 0 conversions is a cut, not a maybe

This skill pulls both the ad set and ad level (1 dead ad inside a healthy ad set is a creative problem, not a pause-the-set problem), classifies everything as CUT, BLEEDING (converting but below breakeven for 14 days), or WATCH, and quantifies the monthly run-rate of flagged spend if nothing changes.

It never flags learning-phase ad sets as waste, and it never pauses anything on its own. You get the list, you say yes, and only then does live mode execute.

Say this to Claude⎊ CopyWhere is my budget going? Target CPA is $30, breakeven ROAS 1.8. Flag what to pause.Cadence: weekly · Mode: paste mode needs 7d and 14d columns06 · signal-health-checkfree skill

If the Pixel is broken, every other skill is confidently analyzing garbage

This one audits the signal before you trust the numbers: are the core events (PageView through Purchase) firing in funnel-proportionate volume, are both Pixel and CAPI sending with deduplication configured, and does Event Match Quality clear the bar (6.0 acceptable, 8.0 good, below 6.0 means fix before scaling)?

It also cross-checks Ads Manager purchases against your store's paid-social orders for the same 7 days. A gap over 20% means an attribution or dedup problem, and the skill names the likely cause. The verdict is binary: TRUST or FIX FIRST, and a FIX FIRST verdict explicitly gates the scale/kill skill until it's resolved. It changes nothing in the account; it's purely diagnostic.

Say this to Claude⎊ CopyMy Ads Manager numbers don't match Shopify. Run a signal health check: here are my Events Manager screenshots.Cadence: monthly, and before any scaling decision on an unaudited account · Mode: paste mode takes Events Manager screenshotsGet the full 10-skill pack, freeEvery threshold in this article ships as editable plain text inside the pack.Get the pack at vibelets.ai

3 · Decide: verdicts, safe scaling, and fatigue catches

Deciding is where dashboards fail you: they show numbers, not calls. These 3 skills turn last week's data into 1 verdict per ad set, a safe scaling ladder, and replacement briefs for dying creative.

07 · scale-hold-kill-verdictfree skill

Every active ad set gets exactly 1 verdict, with the numbers that triggered it

SCALE requires all of: ROAS at or above breakeven × 1.3 for 7 consecutive days, CPA stable or falling, frequency under 2.5, and spend already at 75% or more of budget (if the budget isn't being spent, raising it does nothing, so the verdict becomes HOLD, underdelivering, and routes you to creative instead). KILL fires on ROAS below breakeven for 14 days out of learning, CPA rising 2 straight weeks below breakeven, or zero conversions past the waste threshold.

Everything else is HOLD, and the skill treats holding as a real verdict: most ad sets most weeks should be HOLD. Learning ad sets get an automatic HOLD. Insufficient data gets HOLD with "insufficient data" stated, never a guess.

Say this to Claude⎊ CopyGive me the scale/hold/kill verdict for last week. Breakeven ROAS is 1.8.Cadence: weekly · Mode: paste mode needs both 7d windows08 · budget-stepperfree skill

A winner scaled too fast becomes a loser you re-buy at 2 to 3x the cost

This skill turns "take it from $100 to $250" into a step ladder: increases of 20% or less per rung (under Meta's significant-edit reset trigger), 1 rung per ad set every 48 to 72 hours, and a checkpoint metric at each rung ("proceed only if ROAS ≥ x"). Decreases get the same 20% cap, because crash-cutting resets learning too.

It knows the CBO nuance: on CBO campaigns the step applies to the campaign budget and Meta redistributes the raise, so if you want the money to reach 1 specific winner, it offers the breakout path. Want to go faster than the ladder? It presents the honest alternative (duplicate the ad set at the higher budget as a fresh launch) with the tradeoff, but it won't exceed 20% in a step even if asked.

Say this to Claude⎊ CopyScale [ad set] from $100/day to $250/day. Build me the ladder.Cadence: 1 rung per 48 to 72h while checks pass · Mode: live mode executes today's rung only, after your yes09 · creative-fatigue-scannerfree skill

Creatives die on a schedule; the trick is having the replacement ready before the flatline

Fatigue here is measured against each ad's own peak week, not the account average. Any 1 trigger flags it: CTR down 20% or more from peak, frequency above 3.0 for prospecting (5.0 retargeting), or CPA rising 2 straight weeks while CTR falls.

The real output is the iteration brief for every fatigued ex-winner: what made it win (angle, hook anatomy, format, first 3 seconds), what to keep identical, and 5 variant directions that each change exactly 1 variable. Fatigue refreshes fail when teams accidentally change the winning variable, so the brief names what to keep, not just what to swap. Briefs hand straight to hook-and-copy-writer, then campaign-architect stages the launch.

Say this to Claude⎊ CopyWhy did my winning ad stop working? Scan for fatigue and brief the replacements.Cadence: weekly on any account spending daily · Mode: paste mode needs 4 weeks of per-ad data

4 · Report: the Monday page

10 · monday-report-writerfree skill

One page, under 250 words, ready before Monday coffee

Written for the person who pays for the ads. It pulls this week vs last week (spend, revenue, ROAS, CPA, CTR), and never reports a number without its direction. It finds the top 3 movers by impact on results, not by percentage (a 200% jump on trivial spend is not a mover), and diagnoses each in 1 plain-language line. When the cause is unclear, it says "cause unclear, investigating" instead of a confident guess.

It ends with exactly 3 priorities for the week, each mapped to a skill in this pack ("scale winner X via budget-stepper", "refresh fatigued creative Y, brief attached", "fix EMQ on the Purchase event"). Bad weeks lead the report, paired with the response plan. Never buried.

Say this to Claude⎊ CopyWrite my Monday report. Note: we ran a 20% off promo Thursday to Sunday.Cadence: every Monday · Mode: paste mode needs both weeks at account + ad set level

The guardrails every skill follows

The pack is built so Claude recommends and you decide. 5 rules apply everywhere:

If the connector can see more than 1 ad account, Claude confirms the account ID before reading or changing anything. Campaigns and ads are always created PAUSED, so nothing goes live without you. No budget change executes without your explicit approval per item in chat. Budget increases are capped at 20% per step to protect the learning phase. And every skill's thresholds live in a plain-text section you edit once to match your margins.

Setup in 10 minutes

1Connect your ad account (or skip to paste mode)Add Meta's official Ads MCP endpoint as a custom connector in Claude and sign in with Meta Business OAuth. Test with "show my top 5 ad sets by spend, last 7 days." No connector? Every skill runs on a pasted Ads Manager export instead.2Install the skillsOn claude.ai or desktop: Settings → Capabilities → Skills, upload the pack (skills work on paid plans; see Anthropic's skills repository for how uploads work). In Claude Code: copy the folders into ~/.claude/skills/, or .claude/skills/ in a repo to share with your team.3Ask in plain English"Pull ad angles from this product page." "Which ad sets are still learning?" "Give me this week's verdicts." "Write my Monday report." Claude matches the request to the right skill.

Want the loop without running it yourself? Vibelets runs this same loop in 1 place: it reads your store, generates the creatives, launches to Meta, and flags what to scale or pause for your approval.

Frequently asked questions

What are Claude skills, and how are they different from prompts?Claude skills are folders of instructions that Claude loads automatically whenever a request matches their description, so the same rules run every time without re-prompting. A prompt gives Claude advice for 1 conversation; a skill gives it a repeatable procedure with fixed thresholds, output formats, and safety rules that persist across every chat.Do I need the Meta Ads MCP connector to use these skills?No. Every skill in the pack runs in paste mode: export the table it asks for from Ads Manager (each SKILL.md lists the exact columns) and paste it into the chat. The connector adds live mode, where Claude reads the account directly and can execute approved changes, but the analysis logic is identical in both modes.Can Claude actually change my ad account with these skills?Only with the connector, and only after your explicit yes per item. Campaigns are always created PAUSED, pauses and budget changes wait for approval on the specific list, and budget steps are hard-capped at 20% even if you ask for more (the skill offers a duplicate-ad-set path instead).Why do so many of the skills care about the learning phase?Because edits during learning are the most common way advertisers silently burn money. An ad set needs about 50 optimization events in 7 days to exit learning, and significant edits (budgets over 20%, targeting, creative swaps) reset it, sending costs back to roughly 2 to 3x the stabilized CPA. That's why the guard skill runs before edits, the verdict skill auto-holds learning ad sets, and the stepper stays under the 20% trigger.Can I change the thresholds to match my margins?Yes, and you should. Every skill has a Thresholds section in plain text: breakeven ROAS, minimum spend before judging, fatigue triggers, EMQ bars. Edit them once and every future run uses your numbers.Do these skills work in Claude Code as well as claude.ai?Yes. On claude.ai and the desktop app, upload them under Settings → Capabilities → Skills or add them to a Project. In Claude Code, drop the folders into ~/.claude/skills/ for personal use or .claude/skills/ inside a repo to share them with your team.10 free skills. The whole Meta ads loop.

• ✓  Turns your product page into ranked ad angles and Meta-fitted copy

• ✓  Guards the learning phase before every edit

• ✓  1 scale/hold/kill verdict per ad set, from your thresholds

• ✓  Scales in 20% rungs that never reset learning

• ✓  A Monday report with 3 priorities, under 250 words

Get the free packABOUT THE AUTHORSiddhant SharmaProduct Marketing Manager, VibeletsSiddhant runs product marketing at Vibelets, working daily with the D2C founders and media buyers this pack was built for. The thresholds in these skills come straight from how those operators actually run Meta ads: learning-phase discipline, 20% budget steps, and 1 verdict per ad set per week.

Sources

1. Meta Business Help Center, About the Learning Phase: https://www.facebook.com/business/help/112167992830700

2. Meta Business Help Center, Significant Edits and the Learning Phase: https://www.facebook.com/business/help/316478108955072

3. Anthropic, Agent Skills overview: https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview

4. Anthropic, public skills repository: https://github.com/anthropics/skills

Meta & Facebook Ads