Agentic AI meets the Archetypal Wavelength of Empiricism
Using LLM chatbots to level up your Agent-prompting strategy
Engineers famously love laziness, right? The following may be the ultimate in laziness. And yet, as with everything you’ll find on this blog, it’s at least a full step up from Vibe Coding. It’s a downright scientific approach to using the least energy possible.
This is how to use LLMs—regular chatbots—to come up with prompts that will get your Agentic coding assistants—eg Claude Code, Codex—to one-shot high quality, production-grade code.
Here’s the SOP:
Type out the requirements to your next feature, your next epic, your next entire app idea into ChatGPT or the Claude chat interface in a stream of consciousness ramble or a list of bullet pointed requirements
Ask the bot to create a prompt for an Agentic model to build what you are dreaming up. Ask it to deliver this prompt in a “single markdown codeblock, being careful to avoid triple tick (`) characters that will break the markdown” or, for Claude, simply “a single markdown artifact.” (Markdown is the language AI understands best; the language of its system prompt.)
Provide the prompt to your favorite Agentic model.
Do this at least 5 times. Notice similarities in the way that Claude and ChatGPT write their prompts. Notice the differences. Notice what produces higher quality code. Notice what results in slop.
Start writing your own prompts.
My format, after running this experiment is
```
## Goal: [Explain the bugfix, feature or epic in a single sentence]
### Repo Structure
[Provide a commented directory structure. Ask Codex to make one for you if you don’t have one. Don’t get too into the weeds naming every single file; the broad strokes will do fine.]
### User Story (For real this helps!)
As a [specific kind of user or role] I want to [accomplish some kind of task or have some kind of tool] so that I can [meet some strategic business or personal objective]
### Deliverables
- [Thing one]
- [Thing two]
- [etc]
### Guardrails
# (this should be in your AGENTS.md/CLAUDE.md files too, but repeat it here:)
- All checks, tests, builds, etc. must read all green before submitting. Continue to iterate until everything passes
- [Whatever else you want to say, which may be more specific e.g. “don’t disrupt current functionality x” or “don’t expand the scope”]
```I don’t claim to have the best understanding of prompt structure, though. Experiment for yourself!
An aside about APTITUDE
If you’re new here, APTITUDE is the very self-development philosophy that promotes agency that is described by the blog’s title, An Agentic Development. It’s a sort of Free Will “line of intelligence” in the language of Ken Wilber. In other words, it is a series of stages through which someone who wants self-sovereignty must pass.
And the software development I am doing with AI agents revolves around building a companion app or two for it.
The WavelengthWatch project, currently on GitHub, soon to be on the App Store, aims to help users track what Phase in the ever oscillating highs and lows of life they are currently experiencing.
Knowing what flavor and Phase of the Archetypal Wavelength you are expressing through helps you modulate your behavior so that you can surf human incarnation’s inevitable ups and downs rather than be tossed around by them, victim to them.
APTITUDE asserts that without a little temperance—without restraint and mindful awareness around the urge to amplify highs and lows—we can do little but serve potentially unhelpful, unskillful loops.
Relating it back to Software Development
When you are “Collaborating” as a part of a team, science, rationality, intellect and an explicit shared desire to achieve some outcome are some of the best guiding principles for success that humanity has come up with. This suite of related concepts, which all rose to prominence during the Enlightenment, are classified as Orange in Clare Graves’ Spiral Dynamics.
APTITUDE, borrowing from Graves’ arbitrary color coding, identifies an Orange “Mode” of the Archetypal Wavelength that we must learn to gracefully surf if we would eventually want to make our own choices.
How to surf it with grace?
Instead of Presuming, Question;
Instead of Asserting, Hypothesize;
Instead of Crusading, Experiment;
Instead of Overlooking Details, Collect Data;
Instead of Forcing it, Analyze;
Instead of Failing, Synthesize!
So here’s the scenario. Here’s the User Story.
As someone who is prone to getting carried away by my ideas I want the WavelengthWatch App to ping me every four hours and ask what Mode and Phase of the Wavelength I am currently expressing, so that I can detect, for example, when I am Overlooking Details and correct course toward Gathering Data, ultimately enabling me to be more aware of my internal feelings generally so that I can behave mindfully.
Going back to the process for uncovering the ideal prompting technique
Don’t Presume you know whether Agentic AI can be useful if you’ve never prompted it
Don’t Assert that this prompting technique or that one can be the one and only best way
Don’t Crusade (as many are currently doing) that based on the above biases, Agentic AI is a total dud
Don’t Overlook the Details as more and more people are building their dream projects. (I assure you it’s already happening.)
Don’t Force a belief or a technique to fit a world for which it is no longer optimal
Don’t Fail.
Instead
Question: What’s the best way to use this new tool?
Hypothesize: Can AI tell me how to use itself?
Experiment: Let’s try at least five prompts that ChatGPT generates
Collect Data: Let’s look at the Pull Requests generated as if we are the bot’s senior
Analyze: Which of these prompts actually produces production ready code?
Synthesize: Write your own combination of the best prompts.
After all, just as the Orange Wavelength reminds us. One of the best ways to Collaborate, even with an Agentic AI, is to follow the tried and true steps of the Scientific Method.