Prompt engineering is dead
With enough budget, you too are a few typo-laden sentences away from a cryptographic breakthrough.
Anthropic, thankfully, included the actual prompts: they wanted something “worth publishing,” rejected “low hanging fruit,” and encouraged it to do “proper research.” They were naive and raw and almost comically short, and a billion output tokens later it had something worth publishing.
The guidance around the new models was headed the same way: strip out repeated instructions and examples, stop prescribing every step, give the model the outcome and let it cook.
That week my version of “the guidance” was OpenAI on leaner prompts, Anthropic’s new rules of context engineering and field guide to finding your unknowns, and Boris Cherny explaining why Claude Code cut 80% of its prompt.
Whatever prompt engineering meant, it was clear I was supposed to be doing something differently. What I couldn’t figure out was how to apply that to my own prompts. Do I just take whatever is in my head, spit it out, and trust the model to figure out what I was asking? Read the Anthropic post at face value and the answer was sort of yes?
Last Sunday at 12:01 PM, I was laying down when I voice noted a question I had:
Is there a way for an agent running anywhere on a Linux box, wherever, to ping a phone with a location fix? Meaning like, an agent running on a Cloudflare Worker has the ability to request, ideally without user intervention, a user’s location.
I never sent that note raw to a model. Conveniently, the whole thing was timestamped:
- 12:01: I recorded the voice note.
- 12:09: Went to my desk, the curiosity was too much, so I opened Claude, banged out a few hundred messy words, and asked it for a research brief.
- 12:15: Claude created a 1,418-word brief. I didn’t read it, I just pasted it into a fresh GPT session, which then ran for twelve minutes.
- 12:42: I came back to the finished report and glanced at it. The first-ranked option stuck out, so I pasted the whole thing back into the Claude brief chat with the note, “Holy shit, I think it may actually be possible.”
Claude verified the new notification automation, then suggested going somewhere else. I pushed back and told it to pivot fully onto Shortcuts and automations. The Claude session ended pretty quickly after that. Once I knew I wanted to probe it and build a real codebase, I switched over to Codex. Over the next two days I kept checking in, pushing it further, and eventually built the working probe.
On a drive it returned 32 of 32 location fixes; on a walk, 70 of 70. Median time was 4.38 seconds.
Once it worked, I started wondering how much of that 1,418-word brief had actually mattered. I had been making prompts this way for months because they seemed to produce better research, but I had never figured out what was “load-bearing” (hah). Boris said ablate, so I started ablating.
| Prompt | Words | Hits | Cost/run |
|---|---|---|---|
| Original voice capture | 43 | 0/3 | $0.56 |
| Asked for ranked options | 48 | 0/3 | $0.90 |
| Restated the whole intent | 77 | 3/3 | $1.09 |
| Full generated brief | 1,418 | 2/3 | $1.69 |
Same model, effort, and research harness; three runs each. Cost is the estimated mean per run. A hit means the response includes a Shortcuts automation that gets the current location and sends it back to the remote system. Prompts, transcripts, and code.
The original voice note fails all three times: each run orients around Apple’s Location Push Service Extension. LPSE is a niche iOS 15 API with its own entitlement and two sanctioned use cases. I had asked o1-pro a similar question a year earlier and declined LPSE outright. Had I stopped there, I would have mistaken LPSE for the limit of what was possible.
The problem was in the way I phrased the question: “is there a way…” invites the model to satisfice and find the first example (LPSE) that fits my criteria, but that’s not my intent (which I poorly articulated). And then there’s the qualifier “without user intervention,” which I intended to mean no involvement after a one-time setup, although some models seemed to interpret that as no user-invovlvement at all. Ridiculous!
So the 77-word version worked best, and I guess the extra ~1,300 words from the brief were ceremony. Maybe next year’s models will be more proactive and keep looking anyway, but that isn’t what I asked. Militaries have spent centuries on the nuances of communicating intent between intelligent people, and they still get it catastrophically wrong (and people die).
Maybe prompt engineering means communicating your intent clearly, if only it were easy.
Long live prompt engineering.