AI Built a Complete Game in One Session. It Was Technically Perfect and Painfully Boring.

Two AI agents finished their assignments this week. Neither should have shipped as-is — and the reason why is the same reason your next AI draft might not be ready either.

What happened

Claude Opus 4.8, running in a full agent harness with no human steering mid-task, was given free rein to plan and execute its own research and submit the result to a shadow NeurIPS review. Both papers came back rejected — not for broken code or sloppy writing. The model kept picking the safest, most well-trodden research question available and executing it competently, instead of finding a question worth asking (Import AI #467, Aug 3). The same week, Simon Willison asked Claude Fable 5 to build a browser game from scratch with no help. It wrote the code, generated the assets, ran its own tests, and pushed a finished, bug-free game to GitHub. His verdict: "very, very easy" to build and "pretty boring" to play (Aug 5). His conclusion — designing something people actually want to play is still a human skill.

Both agents did exactly what they were asked. That's the problem: nobody told them what "good" meant beyond "finish it."

Why this matters

This isn't a story about AI failing at execution. Opus 4.8 wrote real code, ran real experiments, and produced two documents polished enough to format and submit to a real venue. Fable 5 shipped a complete, playable game with no bugs a QA pass would catch. The part that used to be the bottleneck — the thing you'd hire a junior analyst or a junior developer for — is done, fast, unsupervised. What both agents got wrong sat upstream of execution: which problem was worth solving, and what would make the result good rather than merely finished.

That distinction used to be easy to miss, because execution was slow enough to hide it. When a first draft took a human two hours, "is this the right thing to make" and "did you make it well" blurred together — you only discovered the direction was wrong after seeing flawed output. Now the output arrives in minutes, complete and polished, and the direction question is the only one left that AI can't answer for you.

What to do differently Monday morning

Before handing off any non-trivial task — a strategy brief, a campaign concept, a data analysis — answer three questions first, not after you've seen a draft:

1. Can you define "good" specifically enough to fail? Not "make it professional" — that's not testable. "Cite three comparable deals from the last 18 months" is.
2. Do you know this domain well enough to catch a wrong answer that sounds right? If you can't, the model's confidence is the only signal you have, and confidence isn't accuracy.
3. Are you willing to discard a polished result that solved the wrong problem? A complete draft creates pressure to ship it — the exact pressure a NeurIPS reviewer, and a bored playtester, don't feel.

A marketing manager delegates a campaign brief

At a 40-person SaaS company, a marketing manager asks Claude for three campaign concepts ahead of a product launch. All three come back polished: clean headlines, consistent tone, complete asset lists. A few months ago she'd have picked the one that read best and moved on. This time she asks the question Opus 4.8 was never asked: is any of these actually differentiated, or did the model pick the safest angle and execute it well? Checked against a specific bar — would a competitor's launch page also work for this claim — two of the three concepts fail immediately. The third needs a new headline, not a new brief. The AI produced the drafts in minutes; catching that only one was worth building took the judgment she'd have needed regardless, just applied before the review instead of during it.

A solo consultant scopes a client analysis

A one-person strategy consultancy asks Claude to analyze a client's churn data and recommend fixes. The output is thorough: six drivers, ranked by impact, each with a suggested intervention. It's also generic enough to apply to almost any SaaS company with a churn problem — the same failure as Opus 4.8's safe research question, technically complete and not actually about this client. The consultant catches it only because she scoped the request narrowly first, naming the two customer segments that mattered and why churn there was different from the industry average. A generic answer stood out as a miss instead of blending in as plausible. Scoping the ask before delegating did more work than any amount of reviewing after.

The takeaway

AI can finish almost anything you give it now — deciding what's worth finishing, and what "good" means for this specific case, is still entirely yours.