Prompting tricks are losing their edge. Not because they stopped working — because the operators getting real output moved past phrasing and started managing AI instead.
Ethan Mollick said it plainly on July 7: prompting tricks are losing their edge, even before agentic systems fully arrive. His point wasn't that phrasing doesn't matter. It's that phrasing was never the actual lever. The lever is the one that makes a contractor or a new hire produce good work: a named outcome, a defined format, a stated quality bar, and someone checking the result before it ships.
That's not a metaphor for us. It's the fix we just made to our own blog.
We Ran Six Articles on Five Bullets and No Guardrails
Until the proposal we approved on July 21, every Framework Friday article came out of five inline bullets — a topic, an angle, nothing else. No required inputs. No rule about where a number could come from. No check before a draft reached review.
Six articles shipped that way. Two of them cited statistics that don't survive a fact-check: a misattributed "74%," and an unsourced "prompt fatigue" figure with no study behind it. Nobody caught either one at the time — not because nobody read the drafts, but because "does this sound right" was the entire review step. It took an outside benchmark pass, weeks later, to catch what a defined quality bar would have caught before publish.
That's the failure Mollick's point actually describes. We didn't have a bad prompt. We had no brief: no outcome beyond "write an article," no format spec beyond "LinkedIn-shaped," no bar an editor could check against, no loop except a read-through.
A Brief Fixes What a Better Prompt Can't
The fix wasn't a cleverer prompt. It was writing down, once, the four things Mollick names:
- The outcome — an angle has to clear a topic gate before it's allowed to become a draft: did we build or measure it, does it correct a named wrong belief, or does it name a pattern nobody's flagged yet.
- The format — Problem → Mechanism → Proof → Implication, 800–1,500 words, claim-shaped headings, one named framework per article, never a new coinage.
- The quality bar — a 16-item checklist, and every number in a draft has to come from a maintained evidence library or a logged first-party result. No "studies show."
- The review loop — the draft ships with its own checklist filled in, pass or fail, before a human opens it.
We haven't run a full batch through this yet — the next one is the first real test, not a finished result. But none of the four things above required a smarter model. Anthropic's own data on delegation puts a number on why that matters at scale: AI already does roughly 60% of developer work, but only an estimated 0–20% of that work is safely delegable without a human checking it first. That gap doesn't close by asking nicer. It closes the same way a stalled process gets fixed — the Handoff Audit we described for broken team handoffs applies just as well to the handoff between what you asked for and what the model returned. Not the model. Not the phrasing. The brief.
Management Doesn't Scale Down to Every Slack Message
This isn't a case for writing a brief before every AI interaction. A one-off summary, a quick rewrite, a draft you're deleting in ten minutes — none of that earns the overhead of a named outcome and a quality bar. Write the brief when the output repeats, ships externally, or is expensive to get wrong late. Everything else, prompt it and move on.
The Friday Diagnostic
Pick the last AI output that disappointed you. Don't rewrite the prompt. Write down four things instead: the outcome you actually wanted, the format it needed to arrive in, what would make it a pass versus a fail, and who — or what — checks it before it's used.
If you can fill in the first three but not the fourth, you don't have a prompting problem. You're missing the same brief we were missing three weeks ago. (See also: Stop Prompting - Start Looping for what happens once the brief is running unattended.)
