## 1. Executive Summary

Andre CIA has written copy for ten years and says he was on the team behind the first launch in Brazil to hit seven figures in a single day. The decisive move in his career, he says, was deciding to build copy with AI. This episode is a working session on that system: how he programmed GPT starting from Cornell's AI paper feed, Discord experiments and a strict hypothesis-test-conclusion loop into a repeatable "document plus prompt" method that drafts an 80% webinar script in about 15 minutes. Around it he lays out the craft that makes the method work: a skill ladder from knowing to teaching, a four-variable model for choosing any sales format, an avatar-product-competition frame for new offers, and the difference between an urgent pain you sell on the promise and a recurring pain you sell on the symptoms.

## 2. Alpha / Sauce

- **Program the model like an intern, not a genie.** Every task gets four blocks: what the document is, the variables it must fill (product name, expert, promise, curiosity point), the copy structure to follow, and the prompt itself. The intern analogy decides how much instruction you give explain the route, not just the destination.
- **Put the prompt at the end of the document.** When GPT has memory and gets lost mid-run, a prompt appended to the uploaded file lets it recover its mission instantly; he versions the file (v2, v3 ) by pasting the prompt under each new draft.
- **Run it as an interview first.** Programmed correctly, the model asks you its variable questions (about three for the intro, seven or eight each for content and closing), then returns a draft 15 minutes to 80% of a webinar script, which he then records over.
- **Command loop: 1 = approve, 2 = revise.** "2" comes with a note on how to rewrite, and every writing command includes an order to re-read the document, the variables and the previously written sections otherwise the model drifts between benefit and pain angles and loses congruence.
- **AI does 80%, you do the 20% that converts.** He predicts two types of people leave the market: those who ignore AI entirely, and those who become dependent on it. The edge is knowing copy *and* AI either skill alone no longer compounds.
- **Multiply one recording into a week of ads.** He records nine intros and nine outros around lesson material, then swaps middles: 9 3 = 27 ads from one shoot, so a dead middle segment gets replaced without re-recording anything.
- **Dry test before you build.** Five product angles become five near-identical landing pages with different headlines and one ad pack; engagement picks the winner before a full funnel exists. One such test paid for a R$8,000 webinar that funded a project the client refused to bankroll.
- **APC: avatar, product, competition.** Fit the avatar first, push an off-target product into the avatar with bonuses, then choose red ocean (compete: stronger promise, price war) or blue ocean (differentiate: new mechanism). Sell urgent pain on the promise; sell recurring pain on the symptoms treat the disease or name the symptom.

## 3. Methods & Setup

1. Study the source, not the summaries: he read the Cornell/arXiv paper feed (roughly 30 new papers a day when he started) plus AI Discord channels, testing each claim in GPT; setup: paper feed + one sandbox chat.
2. Work hypothesis test conclusion; log each failed chat as "one more way that doesn't work," which is the only way programming skill accumulates; setup: running test log.
3. Choose one of four programming styles per task one-shot prompt, sequential prompts, complex single-pass programming, or iterative chat where each run improves the last; setup: style picked by task size.
4. Write the four-part document: what it is, variables, copy structure, prompt with an example block when the task is complex; setup: one uploaded file per deliverable.
5. Frame the instruction load with the intern test: profile, task, output format, example nothing more, nothing less; setup: four-block prompt skeleton.
6. Add "consult the document, the variables and everything written before you write" to every execution command; setup: congruence instruction baked into the prompt.
7. Let it interview you for the variables, answer, then command part one; revise with "2 + observation" until the draft holds; setup: 1/2 command loop on screen.
8. Keep long-form work inside one document; use custom GPTs only for repeat jobs (he uses one for presentation backgrounds with 16:9 or 9:16 output), chats for one-offs; setup: paid GPT tier + two GPTs minimum.
9. Reprogram by hand periodically so the skill doesn't rot dependency on your own GPT is still dependency; setup: monthly manual rewrite of a favorite prompt.
10. Feed transcripts in for content ops: record on Zoom, transcribe with tl;dv, push the transcript to Claude (200k-token window) for chaptered bullet points, then to a mind-map tool; setup: tl;dv Claude mind map.
11. Turn long video into ads with an auto-cutter (Munch-style): 10 12 captioned cuts per upload, Reels/Stories/square/YouTube shapes, and text-based editing where deleting a paragraph deletes the footage; setup: cutter on every class recording.
12. Validate packaging with engagement tests AI pulls the best moments for intros and thumbnails, a small Facebook campaign decides which thumbnail runs; setup: thumbnail/preview test before publishing.
13. Before building an offer, run the APC pass: avatar size and fit, bonus-based product adjustment, competitor map, red-vs-blue decision, then urgent-vs-recurring pain framing; setup: APC worksheet as the pre-build gate.
14. Keep a personal wiki of every learning tagged with "how can I use this?" his Notion wiki holds step-by-step playbooks he can execute without memory; setup: Notion + one note per learning.

## 4. Tools & Stack

| Tool | Role | In this episode |
| --- | --- | --- |
| ChatGPT (chat + custom GPTs) | Script programming and drafting | Four-part documents, 1/2 loop, webinar drafts, background GPT |
| Claude | Long-context reading | Takes hour-long transcripts, returns chaptered bullet lists |
| arXiv / Cornell paper feed | Learning the underlying science | His primary source, ~30 papers daily at the start |
| Discord AI channels | Testing and community signal | Where he cross-checked prompts and techniques |
| tl;dv | Zoom transcription | Captures every class, feeds the text pipeline |
| Munch-style auto-cutter | Video clip factory | 10 12 captioned cuts, text-based trimming, multi-format export |
| Notion | Personal wiki / second brain | Playbooks written as "how can I use this?" notes |
| Chat Farming / Full Funnel | His own AI sales + funnel stack | AI-driven sales, webinar pages, email, WhatsApp, CRM |

## 5. Actionable Checklist

- [ ] Write the four-part document (definition, variables, structure, prompt) for your next asset.
- [ ] Append the prompt to the end of the uploaded file and version the file on every edit.
- [ ] Program the model to interview you for variables before drafting.
- [ ] Add the "re-read document and previous sections" instruction to every writing command.
- [ ] Use only 1 (approve) and 2 (revise with note) as execution commands.
- [ ] Draft with AI, then hand-personalize the 20% that carries voice and proof.
- [ ] Reprogram one favorite prompt manually this month to keep the skill alive.
- [ ] Transcribe one long recording, extract bullets in Claude, publish the clips.
- [ ] Shoot 9 intros + 9 outros around your best lesson and assemble 27 ads.
- [ ] Dry test five headlines on five pages before building any funnel.
- [ ] Run the APC pass and label the pain urgent (sell promise) or recurring (sell symptoms).
- [ ] Keep one Notion wiki note per learning tagged "how can I use this?"