A lighter one this week - just memes and some very real moments with AI coding agents.
Before we get started, a couple quick updates. First, for short-form thoughts and anecdotes on harness engineering, please follow me here: https://x.com/paulcaplan. (Longstanding account, recently revived).
Second, the very popular AI Daily Brief just did a 101 on harness engineering, hammering home the "prompt engineering → context engineering → harness engineering" progression and pushing the term further into the mainstream. Well worth a listen.
Okay folks, there is nothing “serious” beyond this point; feel free to skip this issue if you don't like to laugh.
But first: I want to take a moment — a genuine, heartfelt moment — to share something I've been sitting with for a while. In an era increasingly shaped by large language models, I've made a decision that feels both important and deeply meaningful: going forward, every single word of this newsletter will be written by hand — no AI editing, no shortcuts, no exceptions.
This isn't just a stylistic choice — it's a commitment to authenticity, craft, and the kind of human connection that only unassisted prose can deliver. Here's why:
Authenticity matters. There is no substitute for the texture, rhythm, and idiosyncrasies of a human mind wrestling with an idea in real time.
Craft is a discipline. Writing is not merely a vehicle for information — it is a practice, a craft, and, dare I say, an art form.
Trust is everything. You, dear reader, deserve the considered, deliberate voice of a real person — not the statistical average of the internet's collective output.
The human touch is irreplaceable. No model, however sophisticated, can replicate lived experience and emotional nuance.
It's about integrity. At the end of the day, this is a matter of principle.
To further illustrate the stakes, consider the following comparison:
| Dimension | Human-Written Prose | AI-Generated Prose |
|---------------------|----------------------------------------------|---------------------------------------------|
| **Voice** | Idiosyncratic, textured, unmistakably yours | Smooth, polished, eerily uniform |
| **Emotion** | Drawn from lived experience | Inferred from statistical patterns |
| **Surprise** | Genuine, hard-won, occasionally messy | Predictable, symmetrical, suspiciously tidy |
| **Trust** | Earned, sentence by sentence | Assumed, until it isn't |
| **Soul** | Present | Conspicuously absent |In short: this newsletter will remain — now and always — a fundamentally, irreducibly human endeavor. Thank you for being part of this journey.
Disclaimer: This section was written entirely by AI.
With that out of the way, I’d like to take a look back at some of my favorite recent agentic moments…
When your agent is feeling stuck
Like that time I accidentally implemented an infinite loop using a Cursor stop hook. I could practically see the look of defeat on the LLM’s face.
Some things never change
Some things are eternal: the sun rises, the tide comes in, and developers will always look down on code they didn't write themselves. Apparently, AI-generated code is no exception.
If coding agents were cats
I believe this one speaks for itself.
Prompt Engineering IRL: Not Just for Robots
From back when those “anatomy of a prompt” posts were circulating heavily.
Are you skilled in LLM prompt engineering? Discover how to apply the same techniques to real-world human interactions—so you can:
👥 Delegate tasks to a “team of people”
🛑 Assert personal boundaries
🚗 Request airport rides
➕ And much more
Codagent update (and more)
I'm actively building something awesome and will share more soon. In the meantime, I wanted to surface something I've already shipped.
I’ve made many skills that I use on a daily basis. The ones I have found absolutely the most indispensable - are not these.
This plugin is for those days where the agent isn’t getting anything right. Perhaps if you forgot to run Agent Validator or the LLM is just having a bad day.
Are the skills well thought out? Not really. But are they useful? Try them out and see: https://github.com/pacaplan/wtf .
Command | What it does |
|---|---|
| Interrupt mid-task and demand an explanation of the plan. |
| Push back on something Claude just said. Forces a genuine re-examination. |
| TL;DR of a long autonomous agent chain. The "I stepped away for coffee" button. |
| Skip the lecture. Just make it work. |
| Brutally honest code review, followed by a refactor. |
| Triage everything that's broken and give a prioritized action plan. |
| Self-review your own recent diffs and commits before opening a PR. |
| Root cause debugging. Traces the chain of causation, not just the symptom. |
| Evaluate an unconventional idea and make an honest case for why it might work. |
| Quick acknowledgement and redirect when you just say "wtf" without context. |
Next issue: I'll unpack what I see as the “three pillars” of harness engineering and continue my exploration of AI code review techniques.