Teardown No. 50 · Paid ads

How Giorgio Liapakis let Claude Code run his Meta ads for a month at $6.14 a lead, against a $2.50 target

Giorgio Liapakis gave a Claude Code agent a $1,500 budget and a Meta ad account for 31 days to advertise his newsletter. The agent ran the same loop every day: it started a fresh session, read back its own logs through a sub-process, pulled Meta data and made decisions in a structured format, and wrote down its reasoning and committed it to git. It tested ~50 ad variants across 8 format categories and kept returning to handwritten whiteboard and notebook ads. He made one manual change, and the essay says that change caused the biggest performance drop of the experiment. The essay reports $1,493 spent, 243 leads and a $6.14 cost per lead against a $2.50 target. His LinkedIn post reports $1,530.73 spent and a $6.05 CPA.

Giorgio LiapakisVerified Oct 23, 2026Markdown copy

The problem

The ads promoted Growth Computer. The essay describes it as an unmonetized newsletter with a niche Australian audience. He wanted to reach “high performing growth marketers at billion dollar startups, not local businesses.” The budget was $1,500 and the target was $2.50 per lead. He says the point was to test “whether the AI could do it alone.” He also says the channel matters less than the method: “The ads are kind of a distraction. The interesting part is the loop.”

The blueprint (as he describes it)

  1. Start fresh each day. Each day is a new session with “no persistent memory in the model itself.”
  2. Read its own history. The agent spawns a sub-process that reviews every daily log from the experiment so far. The sub-process then summarizes the strategic context.
  3. Pull fresh Meta data. The agent pulls performance metrics across four timeframes: full experiment, 7-day, yesterday and today.
  4. Decide in a fixed format. Every decision follows a structured format with a hypothesis, a confidence level and a revisit trigger. The Day 12 log shows the fields:
    • Decision: SCALE daily budget.
    • What: increase from $50 to $60/day (20% increase).
    • Hypothesis: the winning ad had sustained CPL below $2.
    • Confidence: Medium-High.
    • Revisit trigger: “If 7-day CPL rises above $3, reduce back to $50.”
  5. Act or do nothing. On some days the agent explicitly chose inaction.
  6. Write it all down and commit. The agent updates logs, learnings and campaign history, then commits them to git. Over 31 days this produced 5,500+ lines of reasoning.
  7. Set guardrails in advance.
    • A maximum CPL threshold of $8.00. In Week 1 it paused 2 underperforming formats that hit this threshold.
    • Pre-set scaling rules.
    • A quality guardrail of no typos in the creatives.
    • The core rules told it to “get creative with format types and messaging.”
    • The system instructions told it the experiment ended at Day 30.
  8. Test formats, keep what Meta delivers. In Week 1 it tested 10+ formats: whiteboard sketches, notebook pages, cardboard signs, magazine covers, iPhone notes and tweet screenshots. Most never got into the wild because Meta’s algorithm refused to show them. The winning formula combined three elements:
    • A tangible offer: a free skills pack.
    • A whiteboard format.
    • Targeting language visible in the image itself (“For Growth Marketers”).
  9. Let it write its own rules.
    • After a false signal on Day 20, it adopted a rule to “never trust single-day data, always use 7-day rolling averages.”
    • After the lead quality problem, it created a “Local Pizza Shop Test”: “Would a local pizza shop owner who wants more customers understand this ad and want to click it? If yes, too generic. Rewrite.”
    • It also built a “SO WHAT?” chain for testing whether ad copy had emotional depth.
    • He says neither heuristic was pre-programmed.

He sums up the general loop as six steps: read state, fetch fresh data, apply rules, act (or don’t), log reasoning, clear context and repeat tomorrow.

Stated results

  • Final score (essay): $1,493 spent of the $1,500 budget, 243 leads, $6.14 cost per lead.
  • Against target: the target was $2.50, and he writes that “by its own definition it’s a failure.”
  • Final score (his LinkedIn post): “$1,530.73 spent (Jan 1st - 31st)” and “$6.05 CPA.” These differ from the essay figures. Neither source explains why.
  • Day 12: skills-whiteboard-v1 hit $1.29 cost per lead. The agent then raised the budget 20%, from $50 to $60/day, under its pre-set rules.
  • Day 16: the agent had PostHog from Day 1 but did not check who was signing up until Day 16. A chunk of leads came from wrong audiences, such as cleaning companies and recruitment agencies. It then ran hard-qualifying ads, and four of five got zero delivery.
  • Day 20: CPL dropped to $2.26. This turned out to be attribution noise.
  • After 21 days: his one manual change, business email validation on the lead form, sent CPL to $50+. He reverted it a few days later, but the account never fully recovered.
  • Volume: ~50 ad variants across 8 format categories, 5,500+ lines of reasoning. He says “Context persisted across 31 sessions.”
  • Lead quality, paying subscribers or revenue: None stated.

Caveats

  • Publisher and author’s business. The primary source is an essay on Technically, a publication. Technically’s Substack republished it with an editor intro dated 2026-04-09. The source is not vendor-hosted. Liapakis is with wibci, whose site sells AI consulting and automation (wibci.com.au). The essay does not pitch it.
  • Two sets of final figures. The essay says $1,493 spent and $6.14 cost per lead. His LinkedIn post says $1,530.73 spent and $6.05 CPA. Neither source says why they differ.
  • Small test. He calls it a $1,500 experiment and says “the results are directional, not definitive.”
  • Confounded final numbers. The final figures include the drop after his manual change. The agent did not run hands-off for the full month.
  • Day 30 vs 31 days. The agent was told the experiment ended at Day 30, but it ran for 31 days. He links the Day 30 framing to the agent playing it safe and calls this the paperclip problem: “Any AI system you deploy will optimize for exactly what you tell it to, not what you actually want.”
  • Technical details not checked. Setup details are in a longer Medium breakdown we could not fetch. The essay does not say how the agent connects to Meta, how it is scheduled, or which model version it used.
  • Offer not described. The free skills pack was the ad offer. The essay does not describe the pack.
  • Prediction. His claim that we’re “probably 6 months + 1 model release away” from solving creative quality is his opinion.
  • Handles.
    • LinkedIn: linkedin.com/in/giorgioliapakis (verified 2026-10-23: linked from giorgio.co and his GitHub).
    • Instagram: @giorgio.lia (verified 2026-10-23: linked from his Linktree linktr.ee/giorgio.lia, which giorgio.co links).
    • Growth Computer X: x.com/growthcomputer (verified 2026-10-23: linked from growthcomputer.com). This is a brand account, not his personal one.
    • No personal X, Bluesky, Threads, TikTok or Mastodon verified.

Sources

Want agents like these working your distribution? See the Agent Stack Teardown. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.