Hey friends,
Every week I share what I’m learning, testing, and building.
But lately something’s been on my mind: there are people in this community (and in the wider AI world) who have things worth teaching that I haven’t discovered myself.
So I’m starting something new.
Rather than always hearing from me, I want to showcase more of the community.
We can all learn from each other.
First up is someone I’d been wanting to have a conversation with for a while.
Eric Elliott.
When he shared his daily prompting technique with me, I had one reaction: why didn’t I know about this sooner? 🤯
Let me introduce him and his lesson.
👤 Meet Eric Elliott
Eric started coding at 6 years old.
He went on to help build Adobe Creative Cloud, and was there when the first versions of deep fakes were developed.
He later joined Adobe as AI Engineering Manager on Firefly, focusing specifically on Firefly Video.
He’s also the creator of SudoLang, a pseudolanguage built specifically for AI agent instruction, and the AIDD (AI-Driven Development) framework.
These apply real software engineering processes to working with AI.
Not vibe coding. Actual engineering.
As of early 2026, he’s working independently and sharing what he’s built.
→ SudoLang: github.com/paralleldrive/sudolang
→ AIDD Framework: github.com/paralleldrive/aidd
📌 The Lesson: Reflective Thought Composition (RTC)
AI defaults to the most common answer.
Not the best answer. The most common one.
That’s the core problem Eric has spent years working around and RTC is the framework he built to solve it.
Here’s the prompt, word for word:
🎯 restate |> 💡 ideate |> 🪞 reflectCritically |> 🔭 expandOrthogonally |> ⚖️ scoreRankEvaluate |> 💬 respond
Here’s what each stage does in plain English:
- 🎯 Restate: The model confirms it actually understood your question
- 💡 Ideate: It generates multiple possible answers or paths (not just one)
- 🪞 Reflect critically: It finds errors or weaknesses in its own thinking
- 🔭 Expand orthogonally: It explores angles it hasn’t considered yet
- ⚖️ Score, rank, evaluate: It judges and ranks what it came up with
- 💬 Respond: Only then does it give you a final answer
The result is that instead of getting the first plausible answer, you get the best answer the model is actually capable of.
Eric appends this to prompts every day across ChatGPT, Claude, Gemini, and his own AI software agents.
💬 In Eric’s Own Words
“The trouble with language models is they tend to do the most common thing. They default to average answers. But they also have in their latent space the knowledge of the brightest minds on earth. When you push them off the average path and get them to explore the best answers instead of just the common answers, they do much better. This prompt forces them to explore more than one path, and find the best one, not just the common one. Reflection helps them discover errors in their thinking and correct them.”
→ In other words: AI isn’t being lazy. It’s doing exactly what it was trained to do. RTC redirects that default behaviour.
“Whenever I type ::rtc the text gets replaced with the prompt. It turns any model into a thinking model that does better than most specifically trained thinking models.”
→ Eric has a text replacement shortcut set up on all his devices. More on that below.
“Prompt engineering is the science of exploring AI latent space and discovering which structures reliably activate its highest capabilities. Some people think prompt engineering will disappear as models improve. They have it exactly backwards.”
→ The people who understand this now will have a genuine edge as AI gets more powerful not less.
🛠 How to Use It (3 Steps)
- Copy the RTC prompt and append it to any hard question: Add it at the end of a prompt where the default answer feels flat, generic, or incomplete. ChatGPT, Claude, and Gemini try all major models
- Set up the ::rtc text shortcut (Eric’s own trick): On Mac: System Settings → Keyboard → Text Replacements. Add
::rtc as the shortcut and the full RTC prompt as the replacement text*
- Use Eric’s developer shortcut if you’re building with AI agents: Inside the AIDD framework, appending
-d 10 to any agent prompt triggers deeper thinking using RTC automatically. More at: github.com/paralleldrive/aidd
One thing to keep in mind: don’t use RTC for quick, simple questions. It processes more tokens and takes longer. It’s designed for the harder problems where you actually want the model to think!
*(On Windows: use AutoHotKey to set the same up. Now you never have to type or paste the prompt manually)
📈 Why This Matters
Most prompting advice is surface level.
Write clearer prompts. Be more specific. Add context.
RTC is different.
This one’s not about how you phrase a question. But rather it’s forcing the model to change how it reasons.
Eric didn’t build this as a clever trick.
It’s the result of years of A/B testing across tens of thousands of prompts.
It’s part of SudoLang, a pseudolanguage he built starting in 2020 with GPT-3, and it’s been running in production software ever since.
It works on every major model.
Even makes dedicated thinking models think better.
As he puts it: “Never underestimate the scientific method when it comes to building better prompts.”
🫡 Final Thoughts
What stuck with me most from this conversation wasn’t the prompt itself.
It was Eric’s framing of what AI actually is.
A model with the knowledge of the brightest minds in its latent space (and most of us never access it because we never ask it to go there).
That’s worth sitting with.
This is what I want to do more of in this newsletter:
Bring in people who are genuinely deep in this work, and share the specific things they actually use.
Rather than always hearing from me, we can all learn from each other.
Try the RTC prompt this week on a hard question you’ve been stuck on.
Reply and tell me what happened.
I’d love to know!
And if you’re interested in sharing a practical AI tip in a future newsletter, reply to this email and get in touch.
To better questions and smarter answers,
Nahid
