mb-ads

Ads Skill

Create ads, generate creative variations, review for compliance, and check ad account performance.

CLI facts first: Find the business repo, run mb status --json --peek, then use status readiness, drift, integrations, measurement, and ranked actions before scoring creative depth or asking setup/provider questions.

Shared Workflow Contract

Engine source workflow: workflows/mb-ads/workflow.md. This skill is the Claude Code shell over that source. Preserve the shared source's fact-first routing, launch-plan boundaries, provider gates, and read/write rules.

Shared source required mb commands:

• mb status --json --peek
• mb start --json
• mb doctor repair --plan
• mb connect doctor --json
• mb connect plan
• mb site check "$SITE_REPO" --business-repo "$BUSINESS_REPO" --json
• mb validate --cross-refs --json
• mb checkpoint --plan --json

Shared source required JSON fact paths:

• money_path
• money_path.objects.proof.quality
• money_path.objects.cta_path
• money_path.objects.channel_strategy
• money_path.objects.active_push
• validation.file_contracts
• content_strategy
• content_strategy.overall_state
• content_strategy.simple_entry_point
• content_strategy.layers
• ranked_actions
• update
• readiness
• drift.items
• integrations
• measurement
• measurement.available
• measurement.state
• measurement.facts.expected_events
• measurement.blocked_count
• measurement.manual_count
• relationship_health.gaps
• checkpoint.pending
• checkpoint.pending.blockers
• runtime.codex_cli
• runtime.claude_code
• state
• blocked
• manual
• evidence
• facts.expected_events
• facts.provider_state
• source
• child_descriptor

Shared source gates: updates_repairs_migrations, file_writes, checkpoint, provider_mutation, publishing_or_spend, customer_contact, private_data, destructive_operations, structured_collection, public_issue_or_proposal, account_change, upload_assets, budget_change, campaign_publish, conversion_upload, gtm_publish.

Shared public/private boundaries: no_secrets, no_raw_provider_exports, no_raw_transcripts, no_customer_member_data, no_private_runtime_settings, no_private_dms_or_gated_communities, no_raw_finance_legal_records, no_oauth_tokens, no_conversion_uploads, no_account_identifiers_in_public_examples.

Shared modes: static, copy-only, image-only, hook-library, video-scripts, long-form-video, review, launch-plan, check, or account-context. These are the account-context modes and paid creative modes for this shared source. Use this routing language exactly when needed: route to mb-think; route to mb-site. Durable paid artifacts use pushes/<YYYY-MM-DD-slug>/ads-batch-001.md, push playbook runs, research, or decisions after approval. Google Ads Search uses the google-ads-search-launch playbook source. Upload assets, provider mutation, spend, publishing, account changes, GTM publishes, conversion uploads, and customer contact remain gated. Codex support is read-only planning until runtime smoke proves more.

Output destinations and operator vocabulary

This skill writes new coordinated work to pushes/<YYYY-MM-DD-slug>/ (the engine primitive). Examples below say "push" for the wrapping record; this is the engine's word. If core/vocabulary.md defines a display word — for example terms.push.singular: drop — speak the operator's word ("drop") in conversation while still writing current files (pushes/..., type: push, linked_pushes). Never rename folders, frontmatter, link fields, JSON keys, or commands based on vocabulary.

If the repo still has legacy campaigns/ records, preview the migration before creating new push work:

mb doctor                       # confirms legacy campaigns/ drift
mb migrate campaigns --plan     # read-only preview of moves

The word "campaign" elsewhere in this skill refers to Meta Ads campaigns (the provider's term for its object) — not the Main Branch primitive.

When creating push.md, include the validator-required frontmatter and fill the values from repo truth or operator answers. When finished creative already exists outside git (a Drive folder, an export), ask "where do the finished files live?" and record media_location plus a media_backend hint (google-drive, r2, local, mb-media):

---
type: push
slug: YYYY-MM-DD-slug
kind: launch
status: planned
health: unknown
goal: { metric: "", target: "", by: YYYY-MM-DD }
owner: ""
audience: ""
offer: core/offers/<offer>/offer.md
promise: ""
media_location: ""   # optional — where binary creative lives
media_backend: ""    # optional hint; requires media_location
---

If the push ties to a bet, decision, research file, playbook, or outcome, add the typed frontmatter link (linked_bets, linked_decisions, linked_research, linked_playbooks, linked_outcomes) and mirror it in ## Related links as Markdown relative links, or preview mb doctor repair --plan and ask first. Use the connection matrix in docs/business-connections.md; never infer links from body-only references.

When an ad-adjacent workflow needs a resource-delivery plan, provider setup recipe, launch checklist, or external automation approval record, draft it as pushes/<YYYY-MM-DD-slug>/playbooks/<playbook>.md with type: playbook — plans and approval records only. Do not publish, schedule, DM, reply, mutate ad accounts, or claim a provider is supported unless mb and a shipped provider adapter prove that path with docs, tests, approval gates, and smoke evidence.

Step 0: Pre-Flight Readiness

Before triage, find the business repo, read deterministic Main Branch facts, then score ad-specific reference depth only when needed. This prevents generic ads from thin reference without duplicating repo-health or provider checks that mb already owns.

0a. Find Business Repo (REQUIRED — do this first)

NEVER search the filesystem. NEVER use Explore or Task agents to find repos. NEVER scan ~/Documents/GitHub/.

CWD-first: If current Main Branch markers exist in CWD, use it. If CWD looks like an old Main Branch repo, run mb status --json --peek and/or mb doctor repair --plan before saved config or discovery. Do not write to old repo structure.

If CWD is NOT a business repo, run:

mb status --json --peek

• If status identifies a usable business repo, ask the user to confirm it.
• If status cannot identify one, ask: "Which business repo should I use?"
• If /mb-ads was invoked from /mb-start, the repo is already identified — use it without asking.

Always confirm the repo before proceeding. Never assume.

After confirming, run mb status --json --peek from that repo. Use readiness, drift.items, integrations, measurement, and ranked_actions as the source of truth for setup, stale context, GitHub, provider readiness, and repair commands. Only run the direct scoring below when status is unavailable or lacks the needed ad-specific detail.

When the request depends on a new, thin, or underperforming offer, load .claude/reference/conversion/offer-sharpening.md before writing hooks or launch-plan claims. Route to /mb-think when audience, outcome, mechanism, proof, risk reversal, objections, reason to act, or next step is unclear.

0b. Score Reference Files (fast — direct Read only, NO agents)

NEVER spawn Explore or Task agents for pre-flight. Read files directly at the known repo path. Pre-flight should complete in under 10 seconds.

At the repo path, resolve offer context first with .claude/reference/business-primitives/offer-bet-push-proof.md, then check these files and count lines:

[resolved offer.md]          → 0 (missing), 1 (<20 lines), 2 (20-80), 3 (80+)
[resolved audience.md]       → same scoring
core/voice.md                → same scoring
core/proof/testimonials.md → same scoring
core/proof/angles/*.md  → count .md files EXCLUDING README.md: 0=0, 1=1, 2-3=2, 4+=3
core/brand/visual-style.md → same scoring (optional)

In multi-offer mode, score the offer-specific offer.md and audience.md (resolved via path resolution), not the brand-level core/offer.md.

Composite = sum of all 6 scores (max 18).

0c. Route on Score

Composite	Status	Action
12-18	GREEN	Proceed to triage
8-11	YELLOW	Warn user, show gaps, allow override
4-7	RED	Route to /mb-think with enrichment targets
0-3	BLOCKED	Route to /mb-setup

Display the readiness report, then proceed. See references/preflight-algorithm.md for gap guidance and smart mix recommendations.

0d. Check Image Generation Readiness

For Facebook image-ad generation, load the MAIN-374 decision and references/image-generation-workflow.md for ad readiness, source bites, playbook routing, review boards, and image-index.md.

mb connect doctor --json

Use status/connect facts first for Google/Workspace readiness. For provider rail smokes, use mb image smoke-openai --repo "$BUSINESS_REPO" --json (default rail) or mb image smoke-fal (explicitly selected fal.ai rail); see references/image-generation-workflow.md for the --generate, credential, media-storage, and image-index.md boundary. Never ask the operator to paste API keys into chat or public issue text.

0e. Paid-Traffic Measurement Gate

If this ad work is for Google Ads, paid traffic to a site/minisite/lander, retargeting, or any request that asks whether a campaign is ready to launch, check measurement readiness before saying "launch."

1. Load docs/google-ads-gtm-conversion-rubric.md.
2. Run mb connect plan or mb connect doctor --json from the business repo when provider readiness matters.
3. If a site repo is known, run:

mb site check "$SITE_REPO" --business-repo "$BUSINESS_REPO" --json

Use the returned state:

• blocked: do not recommend launch; list the blocked checks and exact next command/manual step.
• ready_for_preview: ads can be drafted, but traffic should not launch; tell the operator to run GTM Preview/Tag Assistant and finish provider metadata/approvals.
• ready_for_operator_review: ads can be prepared for review, but launch still needs explicit operator approval for GTM publication, conversion actions, consent posture, budget, billing, and spend.
• ready: local readiness checks passed, but do not mutate accounts or launch campaigns.

Do not invent ready_for_launch or treat ready as campaign launch permission. Main Branch can prepare and review; the operator launches manually.

Never ask the operator to paste Google Ads/GTM tokens, OAuth secrets, conversion uploads, or customer data into chat. Quote mb connect repair commands instead.

---

Meta Ad Account Readiness (Lazy)

Check Meta ad account access only when the user's request needs live ad account context. Do not duplicate provider setup or health checks in prose.

Readiness Flow

1. Read `mb status --json --peek` → integrations/providers/measurement facts.
2. If the operator needs setup choices, run `mb connect plan`.
3. If something looks broken, run `mb connect doctor --json`.
4. If `mb` says Meta Ads account context is `ready`, ask whether to pull a
   compact read-only account summary before generating.
5. Never block generation on missing ad account access.

Graceful degradation: If Meta Ads account context is not ready, mention the optional account context once, quote the mb repair command when useful, then continue from repo reference files, exported screenshots, or manual Ads Manager notes.

User-Facing Display

Describe the capability: connecting a Meta ad account, pulling a compact read-only account summary, or auditing active campaigns. Keep status capability-first and do not name connector vendor details. If the user asks, offer mb connect plan or continuing from reference files.

---

Triage (Flexible Intent Detection)

Detect what the user wants from natural language. Route internally to the right component pipeline. See references/entry-points.md for the complete entry point detection table and component composition.

Intent Detection

User Says	Entry Point	What Happens
"static ads", "full from scratch", "image ads"	Full Pipeline	Copy + compliance + images (classic flow)
"I already have images, just need copy"	Copy Only	Skip image gen, primaries + headlines
"Just need images for existing copy"	Image Only	Image prompts and optional provider generation
"creative variations", "hook library", "one-liners", "50 hooks"	Hook Library	Bulk creative variations (flexible quantity)
"video scripts", "ad scripts", "spoken word"	Video Scripts	Spoken-word script pipeline
"video ad script", "turn this offer into a video ad", "paid sales video", "long-form video ad", "VSL-style paid creative"	Long-Form Video Ads	Load references/long-form-video-ads.md
"I'm repurposing a video", "I shot a video"	Video Repurpose	Transcribe + extract hooks + copy variants
"I want ideas for an ad", "brainstorm"	Ideation	Account check (if available) + concept generation
"research winning ads", "mine reviews", "analyze competitors first"	Research / Mining	Route to /mb-think winning-ad research before generation
"launch ads", "paid traffic plan", "Google Ads launch", "$X/day for Y days"	Launch Plan	Provider-safe plan/check mode, no account mutation
"check launch", "how are ads doing", "continue or kill"	Launch Check	Read status/outcomes/operator exports, recommend continue/change/stop
"Check my ad performance", "what's working"	Account Check	Read-only Meta Ads context if mb connect and runtime tools are ready
"Give me 5 variations of this winning ad"	Performance Iteration	Pull winner + generate variants if account context is ready
"What's working before we create?"	Pre-Gen Account Check	Account overview + creative audit if account context is ready
"review", "audit", "compliance check"	Review	6-lens compliance review

Also accepts: "static", "static ads", "video", "video scripts", "one-liners", "review" -- these route to the same pipelines.

If unclear, ask: "What do you have and what do you need? (e.g., 'I have images, just need copy' or 'full from scratch')"

If research/mining is requested before generation: Route to /mb-think and load mb-think/references/winning-ad-research.md. Customer language, review mining, competitor gap maps, comment mining, and winning script teardown should save to research/ and codify into core/ before this skill generates ads. Do not paste raw review/comment dumps or copied prompt libraries into ad output. If the selected research file has brief_format: grok-8, use its downstream handoff first: business/offering for the promise, ICP for hooks, journey for funnel stage, competitive landscape for differentiation, brand story for voice, content/assets for proof and creative inputs, and metrics/constraints for the review bar. If the brief names a resource-delivery or provider workflow, draft or update a push playbook as a plan only; do not execute provider mutation.

If launch/check is requested for paid traffic: Load references/launch-plan-check.md. Prepare campaign materials, policy findings, measurement readiness, manual provider steps, and approval records; do not mutate ad accounts or start spend.

Proactive Account Awareness

If mb status --json --peek / mb connect doctor --json reports Meta Ads ready and the current runtime exposes the read-only ad account tools:

Before generating: Ask before pulling live account context:

"Meta appears connected for read-only checks. Do you want me to pull a compact account summary before making recommendations?"

If user says yes, run Account Check component (see references/meta-ads-integration.md):

• Say: "I'll pull a read-only summary and avoid saving raw account data."
• Pull only the compact account summary the current runtime supports
• Use safe readiness, activity, and naming-pattern context only

If user says no, proceed to generation with reference files only.

After generating: If the operator approved a read-only summary, use it to frame next steps without exposing raw account data or implying account mutation.

Account awareness is currently read-only. Write operations are on the roadmap and require explicit operator approval gates -- see references/meta-ads-integration.md.

---

Pre-Flight: Special Ad Categories

Before generating any ads, ask:

"Will this campaign run as a Meta Special Ad Category? (Housing, Employment, Credit, or Social Issues/Politics)"

If Employment (job training, career coaching, hiring, job boards):

1. Load additional rules: See Meta Policy lens → Employment section
2. Warn user: "Employment category has strict restrictions. I'll avoid salary assertions, 'if you've been...' patterns, and job-seeking status claims."
3. Tag the output: Add special_ad_category: employment to frontmatter

For Employment, avoid any assertion about current status such as job, salary, employment state, job-seeking status, or dissatisfaction. Use aspirational framing only.

---

Offer Context And Required Reference

Before loading reference files, resolve active offer context with .claude/reference/business-primitives/offer-bet-push-proof.md and the file list in references/preflight-algorithm.md. Use current core/ paths first; treat old paths as migration input only.

Before outcome-claim or substantiation recommendations, read money_path.objects.proof.quality. If quality.public_marketing.status is blocked, do not draft proof-backed public ad claims; route to /mb-think for permission collection before ads, pages, or public claims use that proof.

If content strategy layers or Skool surface notes exist, use them to decide the paid role and keep ads congruent with visible offer, pricing, and proof. Angles are additive; mix established angles with any newly codified ones.

---

Mode: Static Ads

Create campaign batches with image prompts + ad copy. Each batch = 5-6 angles, each angle = 3 image creatives (graphic, lo-fi, interrupt). Hook = 123-135 chars, no questions, no "you/your" in first 3 lines, no emojis. 5 ad styles (Deep, UGC, DR, Pattern Interrupt, Testimonial). Format pair: 1:1 + 9:16.

See references/mode-static-ads.md for the full workflow: campaign structure, hook formulas, copywriting batch sequence, ad styles by length, image prompt types, and the file save convention.

---

Mode: Hook Library (Creative Variations)

Generate punchy, truly diversified creative variations for static image ads (Andromeda-optimized). Users can request any quantity. Also called "one-liners" — same methodology, same pipeline.

The core rule: Every variation must include at least one specific anchor (role, niche pain, value prop, or proof point). The Specificity Test: if it could sell a gym membership, it fails.

See references/mode-hook-library.md for the full 6-step process, anchor rule, input modes, output file format with full generation context, and links to one-liner-methodology.md / one-liner-examples.md.

---

Mode: Video Scripts

Create diverse spoken-word scripts for camera delivery. 15-30 scripts across 3-4 buyer avatars, ~5th grade reading level, contractions, fragments. Each ad = a fundamentally different reason to buy.

See references/mode-video-scripts.md for the 6-step process, script structure (Hook / Body / CTA), spoken delivery optimization, and save convention.

For long-form paid sales videos, video ad scripts that need VSL structure, or VSL-style paid creative, load references/long-form-video-ads.md.

---

Mode: Review

Review ads through 6 compliance and quality lenses before shipping (FTC, Meta Policy, Copy Quality, Visual Standards, Voice Authenticity, Substantiation). Spawns 6 parallel Task agents (read-only), synthesizes a unified P1/P2/P3 report, presents proposed P2/P3 copy changes as a diff, and applies them only after explicit approval.

See references/mode-review.md for the full lens table, review process, severity levels, and status determination.

---

Automatic Post-Generation Pipeline

Every generation entry point (Full Pipeline, Copy Only, Hook Library, Video Scripts) runs this pipeline automatically after saving output. Do not ask the user whether to run compliance review -- it is automatic.

See references/post-generation-pipeline.md for the complete pipeline: checkpoint pre-review, lens tier selection, optional image-provider check, parallel agent spawning (compliance + image), synthesis, proposed-change approval gate, unified report, and post-review checkpoint.

Quick summary: Save a checkpoint for the pre-review state after operator approval, spawn 5-6 compliance agents + optional image agents in parallel, synthesize P1/P2/P3 findings, surface P1 to user, show proposed P2/P3 copy edits as a dry-run diff, apply copy edits only after explicit approval, present unified report, and offer a post-review checkpoint.

"While You Wait" pattern: When spawning parallel agents that take >30 seconds, show a brief note so the user knows what's happening:

"Running compliance review across 6 lenses + generating images in parallel. This takes about 2-3 minutes. These run as sub-agents so they won't eat into your session context."

---

Compliance (All Modes)

Never say:

• Cures/treats/heals [condition]
• Guaranteed results
• Will eliminate [problem]

Safe to say:

• Many have found this helpful
• Supports/complements existing approach
• Framework for understanding
• Education and guidance

---

Quality Checklist (All Copy Modes)

Before saving any batch, verify:

Check	Requirement
Anchor specificity	Every hook/variation has at least one offer-specific anchor
Cold traffic language	No insider jargon — would a stranger understand in 3 seconds?
Hook length	123-135 characters for static ad hooks
No questions	Hooks don't start with yes/no questions
No you/your	First 3 lines avoid direct address
Angle diversity	Each concept uses a genuinely different psychological entry point
Voice match	Copy matches voice.md tone (if available)
Compliance	No banned claims, proper testimonial attribution
Skool congruence	Claims match live about page + pricing cards (if skool-surfaces.md exists)
Save-ability	Would someone save this ad to reference later? Educational, actionable hooks drive purchase intent
Enemy framing	Does at least one concept use a named enemy from voice.md? Enemy-driven contrast creates identity alignment

---

Recovery from Compaction

If context was compacted mid-task, check:

1. Which offer? Use a future mb JSON active-offer field if present; otherwise ask the user to restore offer context. Do not silently route from .vip/local.yaml.
2. What entry point? Full pipeline, copy only, hook library, video scripts, review, account check
3. What stage? Planning angles, writing hooks, generating prompts, reviewing, pulling account data
4. What's done? Check pushes/ for partial work. If old campaigns/ records exist, run mb migrate campaigns --plan before relying on them.
5. Ad account status? Re-run mb status --json --peek; if needed, run mb connect doctor --json
6. Resume: Continue from the last completed step

For full pipeline: Did we finish image prompts (Part 1) before copy (Part 2)? For hook library: How many variations generated out of requested quantity? For video: How many of 15-30 scripts are done?

Quick Reference

Full/static: 5-6 concepts x 5 primaries x 5 headlines x 3 image prompts. Hook/video/review: default 30 hooks, 15-30 scripts, or 6-lens P1/P2/P3 review. Account/launch: read-only Meta summary or provider-safe paid plan/check; no account mutation without shipped adapter and explicit approval.