Most prompt-tip content is noise because prompts are context-collapse artifacts. The thing that worked for you worked inside a workflow — a specific model version, a specific task, a specific failure you were patching around. Strip that off and paste it into a LinkedIn carousel and what's left isn't knowledge. It's the residue of knowledge, the same way recounting a dream at brunch is the residue of having had one.
The steelman is real. Disciplined prompt-sharing exists. People publishing system prompts for legal review workflows, code-debug loops, structured extraction pipelines — that material transfers because it ships with the context it was built against. The model version is named. The failure mode is named. The task is constrained tightly enough that the reader can tell if their situation rhymes. I'm not arguing against that. I'm arguing that ~95% of what gets posted under the banner of "prompt engineering" isn't that, and the format is the giveaway.
Dream-telling and prompt-sharing fail in the same way. Both are intensely meaningful to the producer because the producer is reconstructing a lived experience as they narrate it. The producer feels the texture — the moment the model finally got it, the specific clause that fixed the hallucination, the surprise of the output. The audience gets none of that. They get the cleaned-up artifact, severed from the loop that produced it. A prompt without its task is a punchline without a setup.
This is why the genre has a signature shape. "Act as a Pulitzer-winning copywriter." "You are an expert in [field]." "Do not use AI-sounding language." These read like incantations because that's what they are — gestures at expertise rather than expertise itself. The persona prompts are the worst offenders. "Act as a senior engineer" tells the model less than three lines of the actual problem would. Roleplay framing is a substitute for specifying the task, and substituting is exactly what bad prompt content does. It substitutes vibes for constraints.
There's a meta-irony that the loudest commenters have already clocked. The viral prompts telling models to "sound human" and "avoid clichés" are themselves drawn from the same clichéd corpus the model trained on. The format is recursive. Derivative content asking for non-derivative output, optimized for a feed that rewards derivative content. The whole thing eats itself, and the people producing it can't see the loop because they're inside it.
The platform asymmetry matters too. The mockery is on Reddit. The producing is on LinkedIn and X. These are different audiences with different incentive structures — one rewards craft signaling, the other rewards reach. So the backlash never reaches the producers, and the producers never feel the pressure to change, which is why the volume keeps going up. The signal-to-noise ratio of public prompt content has gotten measurably worse over the last 18 months, not better, even as the underlying models have improved and the genuine craft has deepened. Those two things are happening on different surfaces.
What would actually be worth reading? Failure reports. "I tried this, here's the output, here's why it broke, here's what I changed." Concrete numbers — the prompt cut latency from 4s to 1.8s, or reduced hallucinations from 11 in 1,400 runs to 2. Prompts embedded in code, not formatted for a slide. The constraint is the content. If your prompt-tip doesn't survive the question "which model, which task, which failure mode," it's a dream — interesting to you, exhausting to me.
The test is simple. Could someone six months from now, on a different model, building a different product, get value from this? If yes, ship it. If no, write it in your notebook where it belongs.
Further reading
- Anthropic's prompt engineering docs — what production prompting actually looks like when it ships with constraints.
- Simon Willison on prompt injection and why prompts are programs — the strongest argument that prompts deserve the engineering label, made by someone who consistently shows their work.
- The chain-of-thought paper (Wei et al., 2022) — the origin of one of the few prompting techniques that actually transfers across tasks.
Top comments (0)