Any dashboard can hand you two hundred metrics. Knowing which one is lying, what it's costing you, and whether the fix worked. That's the whole job. Nine stages, thirteen specialist skills, yours to download and run. Underneath, every account is the same nine steps. Pick a business type below and watch what actually changes.
Free · 13 skills · runs in your own LLM · no card
What's connected, what it unlocks, what's blocked before anything runs
Your economics, your funnel in your own words, and the one event that means money
Eleven phases. Tracking is validated before performance is ever judged
Our rules run here: root causes, scoring, sequencing, before anything is touched
Draft the change. You approve it. Nothing goes live by surprise
Set it to run every week, on its own, without you
Three parts carry your business. Five carry the method, and those never change.
Switching pillars did not rebuild anything. It changed four things: the number we optimise toward, and the three places we go looking for the truth.
These load before Stage 01. They carry the judgment: the thresholds, the kill rules, the things we refuse to guess at.
If a skill can't be loaded, the system is told to stop and name it, never to quietly fall back on generic best practice.
Run the full nine-stage Meta loop on my account and give me one self-contained HTML dashboard at the end. Stop at each gate. Do not assume anything about my business. CONTEXT Business: [name] · [what you sell] Ad account: [act_id] · Market: [country] · Currency: [INR] What a win is worth: [average order value or closed deal value] The decision I am trying to make:Which things do I fix first, and what are they worth?Period: last 90 days SKILLS · LOAD THESE FIRST Before you run a single stage, load these from my account. They carry my judgment; without them you are guessing: · meta-growth-os: the staged pipeline and its approval gates · meta-audit: the eleven-phase audit and its scoring weights · meta-setup: the connector requirements and Limited Mode · meta-memory: what we already know about this account If any one of them is NOT available in this session, STOP and tell me exactly which skill is missing, before you run the stage that needs it. Do not silently fall back to generic best practice. STAGE 0 · READINESS (can this run honestly) Confirm the Meta connector and Google Drive are reachable. Detect by capability, not by connector name — names drift and contain typos. If either is missing, name the fix and wait. When I say it's connected, re-probe rather than taking my word for it. Ask which CRM or commerce platform holds my outcome data. Branch to a connector or a named export. If I say "we don't track that", record it as a finding, not a failure. STAGE 1 · DISCOVER (who this is) Check memory first. If a contract exists, show it and ask what changed — never re-interview. Do not pre-fill any declared truth from what you observe in the account. Ask me: the business model, the end goal event, my funnel stage by stage in my own words, the primary conversion, and the source of truth. If what my campaigns optimise toward differs from the end goal I state, show me both and ask whether it's deliberate. Do not conclude. Then ask for unit economics and derive my target. Do not proceed without them. → Write file: 01-Discover/Intake-Contract.md STAGE 2 · AUDIT (eleven phases, in order) Present each phase's status block and wait before continuing. Do not judge performance until tracking is validated. If the wrong pixel may be connected, events are duplicated, or CRM leads can't be reconciled — pause and say what is blocked. Segment before averaging: form type on Service, prospecting vs retargeting on E-commerce. Score 0-100 across eight weighted categories. Never score a category on guessed data. → Write file: 02-Audit/Audit-Report.md STAGE 3·4 · DIAGNOSE AND PRIORITISE (my rules run here) Sort every finding intoBroken(exists but wrong),Missing(should exist, doesn't), orLeaking(works, underperforms). Collapse findings that share a cause into ONE gap. Fourteen ad sets in Learning Limited is one gap, not fourteen. Count causes, not symptoms. Score every gap: Impact × Effort × Confidence × Time-to-Impact. Rank against the Leverage Ladder — offer beats creative beats targeting beats bid tweaks. → Write files: 03-Diagnose/Gap-Inventory.md, 04-Prioritize/Roadmap.md Wait for my yes before anything gets built. STAGE 5 · EXECUTE (drafts only, paused only) Everything is a draft. Anything built in-account is built paused, one item at a time, with my explicit confirmation per item. Never bulk-pause, never change a budget. If a recommendation resets the learning phase, say so and estimate what the reset costs. → Write files under: 05-Execute/ STAGE 6·8 · MONITOR AND SCHEDULE (it runs without me) Capture a baseline the day fixes ship. Then set a recurring weekly run from a cold start. Do not log a regression inside the learning phase or below ~50 conversions — that is noise. Notify me only when something changed, broke, or crossed a threshold. → Write files: 06-Monitor/Baseline.md, 06-Monitor/Drift-Log.md, 07-Report/ FINAL OUTPUT · ONE HTML A single self-contained .html file. Dark. Brand colours #0E0E0E obsidian and #FF5C1A signal orange. It must show the PROCESS, not just the answer: · the nine-stage flow at the top, with what each stage produced · the intake contract as I declared it, in my words · the gap inventory and the ranked roadmap · an approval checklist with real checkboxes · a closing "what this could NOT check" section No CDN, no external images, no server. It must open on a laptop with no internet. RULES · Never invent a number. · If a data source isn't connected, skip the check and mark it incomplete. · Never compare across attribution windows, objectives, or audience temperature. · Nothing goes live without my explicit yes. · Never declare a winner below ~50 conversions or inside the learning phase. · Lead with the answer. No jargon without a number next to it. · If a skill above could not be loaded, say so at the top of the output, not buried at the bottom.
Nothing here is aspirational. Each one is wired into at least one stage of the pipeline. Eight run the moment you download it. The rest need a data source connected, and two of those are accounts you already own.
runs all nine stages in order
eleven phases · fans out · scores
tier 1 blocks · tier 2 branches · Limited Mode
what we already know · saved per account
published medians with confidence marked · target derivation
what's actually being measured, before anything is judged
pixel · CAPI · dedup · event match quality
hook rate · hold rate · fatigue curves
learning phase · consolidation · budget
saturation · overlap · placement · exclusions
can you see the truth in your own reporting
what happens after the click
did anything break since the baseline
founder · client · internal · dashboard
produced in Execute, from the diagnosis
CAPI and offline mappings, written to spec
entities · field catalogue · datasets: required
account memory across runs: required
lead → close · no-show · deal value: free, you grant access
true revenue · margin · returns: free, you grant access
| The question a client actually asks | What's wrong | What it's costing |
|---|---|---|
| What's broken in my account right now? | ✓ | ✓ |
| Is my tracking actually intact? | ✓ | ✓ |
| Which campaigns are burning money? | ✓ | ✓ |
| What should I fix first, and in what order? | ✓ | ✓ |
| Did last week's change break anything? | ✓ | ✓ |
| Why did my cost per result go up? | Partly | ✓ |
| Which creative should I actually kill? | Partly | ✓ |
| What is a lead actually worth to me? | – | ✓ |
| Which campaign produces clients, not leads? | – | ✓ |
| Which campaign feeds my no-shows? | – | ✓ |
| What does my traffic return rate do to true CPA? | – | ✓ |
| Am I actually profitable after margin? | – | ✓ |
A dental clinic runs Lead → QL → Appointment → Visit → Treatment. An investment firm runs Lead → Booked Meeting → Completed Meeting → Client. Same pillar, different shapes. You draw it, in your words. We never hand you a template to confirm.
Written down one time. Applied identically to every account, and to every engagement we take on later. It's why two audits six months apart are actually comparable.
The account context is cheap and changes monthly. The judgment is expensive and took years. Only one of them should be retyped for every new engagement.
Every stage writes a file, and they cross-reference each other, so any recommendation traces back to the evidence that produced it, and forward to the metric that proves it.
Your declared truths, your funnel, in your own stage names. Re-confirmed every run
0–100 across eight weighted categories, evidence attached
Every problem as a root cause: Broken, Missing or Leaking
Quick wins, big bets, and what to deliberately not do
What ships when, and what each phase should move
Handed straight to an editor: hook, structure, timings, claims to verify
Dated record of what changed, so regressions can't hide
Founder one-pager, client narrative, or task list
Not the model, not the goal, not the funnel shape, not what counts as a conversion, not whether a setting was intentional. Observable facts get read from the account. Declared truths get asked as open questions and recorded in your words — never pre-filled from a guess, because a pre-filled guess is an anchor, not a question.
If a data source isn't connected, the check is skipped and marked incomplete, never estimated. Every benchmark carries its source, sample size and a confidence level — because published Meta benchmarks disagree with each other by up to 2×, and most posts don't say where their figures came from.
The audit pauses if the wrong pixel may be connected, events are duplicated, or CRM leads can't be reconciled. "Your last 90 days of reporting can't be trusted, and here's why" is frequently the most valuable sentence in the whole engagement — and most audits are structurally incapable of producing it.
Each carries the observation it rests on, what it depends on, an explicit "here's how we'd know this failed", and a leading indicator you can check yourself without commissioning another audit. Advice you can't falsify isn't advice. It's decoration.
Nothing is judged inside the learning phase or below roughly fifty conversions. Most of the damage done in ad accounts comes from decisive people acting on samples too small to have an opinion about — so the system is allowed to return "wait", and does.
Draft-first, paused-first, one item at a time. No bulk pauses, no budget changes, no audience edits. If a recommendation resets the learning phase, it says so and estimates what the reset costs — so you're choosing knowingly rather than discovering later.
Everything marked in terracotta changes with the account. Everything in mint is the method, and it took years to write down. Download it, run it in your own LLM, keep whatever it finds. The audit was never the hard part.
Free · 13 skills · yours to keep · no card
Free utilities built from our actual client work. Use them, share them, run the numbers yourself.
Model CPL, ROAS, and budget across 4 platforms × 11 verticals.
Run the numbers →Doing organic too? The SEO Growth OS is the same architecture for search — six nodes, thirty-two specialist checks. Meta Growth OS is its paid half.
Enter your email and the full bundle arrives in your inbox: 13 skills, a README, and a setup guide. It runs in your own LLM.
We send the file and nothing else unless you tick the box. See our privacy policy.
Rather not set it up yourself? Book a setup call.
The full bundle is downloading now: thirteen skills, a README, and a setup guide, as a single zip. It runs in your own LLM. We saved so we can send updates only if you asked for them.
Didn't start? Download it directly.
Want it set up properly instead? Book a setup call.