Add this to your agents
|
Yo! I just spun up the analysis section of my new Meta Ads skill and analysed my last £10k of Meta ad spend. 666 ads. 56 campaigns. And it was pretty eye-opening. It;s made me realise how we have to ensure AI can improve on itself. Look, marketing has always moved quickly. And AI is making it change faster. A system that works today is outdated in a month if it doesn't adapt. ... new popular hooks ruin old winning ads If you're not staying on top of things, everything grinds to a halt. So I've built two things into every system I create now.
Think of this as a supervisor. Someone who checks the quality of everything before passing it off to the next stage. 2. A full feedback loop This is something that analyses success and not only optimises to do more of what's working. But also updates training based on how things are working now. This way, your agent systems will stay more up to date and adapt as the industry does. Case in point, I ran my analyses on the last £10k of ad spend in my account and it told me this. The Meta Ads performance analysis was just completed. Here's the summary of key findings from analyzing 666 ads across 56 campaigns (~£10,456 total spend):
Top findings:
- Landing page is the critical bottleneck — 87.6% drop from landing page view to view_content. This is where most ad spend is being wasted.
- Video direct-to-camera teaching format wins — consistently best CPA and ROAS across campaigns.
- Fewer ads with more budget each outperforms spreading budget thin across dozens of ads.
-Warm audiences produce best CPA (£33.99 warm vs £53.26 interest-based vs £89.66 broad). Lookalikes and broad should be killed or heavily restricted.
-"Freebie seekers" is the strongest hook angle — consistently outperforms other messaging.
- 1DPC AI product has better unit economics than the original 1DPC.
The natural next step is to seed these findings into the feedback loop system we just built. Let me log them into the Meta Ads learnings and proven patterns files.
I already knew most of this. But now it's confirmed, logged, and automatically feeding back into every campaign I build next. The best bit is I taught it how to weight the learnings so we can improve outcomes without sacrificing creative testing. At the end of the analyses, it created these guidelines for the next ad test I run. Done. Here's what I've seeded:
Learnings log (8 entries):
L-001: Video vs image format (early signal)
L-002: Warm vs cold audience types (early signal) L-003: Freebie seekers hook angle (early signal) L-004: Direct-to-camera teaching format (early signal) L-005: Landing page bottleneck at 87.6% drop-off (single data point) L-006: Budget concentration vs spread (early signal) L-007: Lookalike underperformance (single data point) L-008: 1DPC AI vs original 1DPC product (single data point) Proven patterns (4 early signals):
PP-001: Video outperforms static image
PP-002: Warm audiences best CPA; lookalikes underperform PP-003: Freebie seekers angle consistently wins PP-004: Concentrate budget on fewer ads All patterns have 50/50 exploration ratios since they're early signals. They'll tighten to 70/30 or 80/20 as more data confirms them. The Don skill will now reference these when reviewing new creative and campaign structures.
This has been fed back into the planning stage of the skill, so the next lot of ads I create should be more in line with what's working across my accounts. What I'm seeing is, with the right systems and feedback loops, AI is able to improve on itself. It;s able to analyse data far faster than you and, when guided, use it to make sure that each new run is better than the last. Make sure you're building feedback loops into your AIs. I'm finishing off this polish and a few other skills, then I'm releasing them all in GM+. If you want access to them, hit reply and let me know. Pete "Skynet creator" Boyle |