My favorite tools
Wispr Flow: Talk, don’t type. Free, so much faster.
Granola: Meeting transcripts connected to Claude through MCP.
Claude: My daily driver for AI. The first 3 people to sign up through this link get a free week of Pro.
In the last 180 days we've spent a lot on Meta Ads.
We've launched 251 unique ads in that time and used 1,160 unique elements in combination with those ads like primary texts, display links, ad headlines, etc.
HOLY DATA!
I wanted to know what style of hooks perform better?
What styles of imagery correlate to better performance at the prospecting level?
What ad headlines lead to cheaper cost per SAOs? (SAO = sales accepted opportunity)
Along with finding the answers to these questions, I wanted to make it easy for my team and others to leverage these insights when creating that next ad, or writing that next headline.
What we built
Think of it in three layers.
Layer 1: The data. Every ad, every element, along with each one's associated performance
Layer 2: The context layer. The aggregation of performance by theme and emergence of "rules"
Layer 3: The Claude skills. Making it most convenient for my team to leverage these rules and insights. The cool part about these Claude Skills is that they're fed from my GitHub repo so every Monday when new ads/elements are processed and rules change, the skills are dynamically updated within everyone's Claude Desktop app.
Rules = DOs, DON'Ts, and TESTs.
DO: a rule that has been proven statistically significant that we should do more of in our ads.
DON'T: something we should do less of.
TEST: an emerging DO or DON'T that isn't statistically proven yet.
I made context files in Confluence for each element group:
Creative Headline
Creative Subheading
Ad Headline
Primary Text
Visual Patterns
These context files serve people and agents as a trustable source to guide creative decisions based on data. Not opinions. Not "I feel like the blue one did better." Data.
The stack:
Meta Ads: the source, 251 ads and 1,160 elements over 180 days and counting
Python scripts: re-evaluate every rule on a scheduled run each Monday based on new ad performance
Confluence: home of the context layer, one page per element
Claude skills: the copywriting and ad design skills the team actually uses
The cherry on top
Every Monday, python scripts run and re-evaluate the rules for each element based on performance and if changes are made the Claude skills are updated automatically with the new insights!
So the loop looks like this:
New ads run and generate new performance data
Monday's scheduled run re-scores every rule
A TEST that proves out gets promoted to a DO (or a DON'T)
A DO that stops working regresses to a TEST (or DON'T) or vice-versa
The context files update, and the Claude skills update with them
Nobody has to remember to go update the playbook. The playbook updates itself.
Is there a decision your team still makes on gut feel that you wish was backed by data?
Why this matters
It's good to challenge intuition. It's not wrong at all that your team may "feel" and "think" certain ways are the best way to do things. A system like this will challenge that intuition in a good way and shape it. I heard team members say "Wow, I didn't think that style was performing so bad", but I also heard, "Dang, I knew those type of headlines were working."
Your best insights stop living in one person's head. The rules are written down, in one place, for people and agents.
Every new ad starts from what's already proven. Nobody's starting from a blank page or repeating a mistake you already paid for.
The rules keep up with the market. What worked 6 months ago gets re-checked every Monday, so the team isn't following stale advice.
More shots at cheaper pipeline. When every headline is built on what drove cheaper cost per SAO, the whole account gets pulled in that direction.
Same system, different problem
Outbound email: sequence copy, reply rates, meetings booked → "Write a first touch to a VP of Ops using the subject line rules that book the most meetings."
Discovery calls: call transcripts, closed-won and closed-lost deals → "Which discovery questions show up most in deals we win?"
Landing pages: page variants, demo request rates → "Draft a hero section for our new pricing page that follows our proven DOs."
Proposals: sent proposals, discount structure, close rates → "Which offer structures close fastest without shrinking deal size?"
Cheers,
Matthew Shelton
P.S. Missed #033? Here's how I put Muse to work on FB Marketplace.
