The End of the Perfect Prompt

Andrew Ng — the Stanford professor who co-founded Google Brain and Coursera, and has probably taught more people machine learning than anyone alive — put a number on something we have been building around for two years, and most people scrolled past it.

On the HumanEval coding benchmark — the numbers Ng published through DeepLearning.AI — GPT-3.5 asked once — one prompt, one answer — solved 48.1 percent of the problems. GPT-4 asked once solved 67 percent. Then GPT-3.5, the older and cheaper model, was wrapped in an agentic loop: draft, critique, revise, test, repeat. It solved 95.1 percent.

Read that again. The upgrade from architecture was bigger than the upgrade from a whole model generation. The loop beat the brain.

That is why Ng is now telling engineers that prompting, as a discipline, is on its way out. Not because prompts stop mattering, but because the prompt was never the unit of work. The workflow is. His four design patterns — reflection, tool use, planning, and multi-agent collaboration — are all the same idea from different angles: never let the model's first answer be the final answer.

One shot is a demo. A loop is a Cystem.

A single prompt is a slot machine. Sometimes brilliant, sometimes garbage, never accountable. There is no step where the output gets checked, no step where it gets improved, no step where the failure teaches anything.

A loop changes the physics. The model drafts. A second pass attacks the draft — what is wrong, what is missing, what would break. The draft gets revised against that critique. A verifier checks the result against reality, not against vibes. And only then does anything ship.

Each pass is cheap. Stacked, they turn a 48 percent machine into a 95 percent machine on HumanEval. That is not a trick. That is what quality control has always been — we just forgot to give it to the machines.

Then the loops get connected into graphs

The part most people skip is the second half of Ng's argument. One loop makes one task reliable. The real shift is connecting loops into a graph: a workflow where the output of one loop feeds the next, where the path can branch on what the work finds, and where state survives between runs.

That last part is the compounding engine. A graph that remembers is a graph that gets better every time it runs. The failure from Tuesday becomes a rule on Wednesday. The pattern that worked for one client becomes the starting point for the next. Prompting resets to zero every conversation. A graph never starts from zero again.

This is also where the honest caveat lives, because the industry is currently arguing about it. Loops are simple and cover most single-task work — one well-scoped job, retried until it passes, with a human reviewing the output. Graphs earn their complexity only when you have branching logic, multiple specialists, and state that must persist. Reaching for a heavy orchestration framework on day one is how you build a machine nobody understands. Structure first. Orchestration when the work demands it.

This is not theory to us. It is Tuesday.

Here is why this course landed on my desk with a note saying "this is essentially what we do already."

Every serious piece of work we put our name on runs through exactly this architecture. A draft is never the deliverable. It goes in front of a council of independent reviewers — separate seats, separate models, one whose only job is to attack the work, one whose only job is to argue against shipping it. The author never approves its own output. That is Ng's reflection pattern, made adversarial.

Verification does not trust the report. It reads the live artifact — the actual bytes on the wire, the actual page as served — because a green checkmark is a claim, not evidence. That is reflection pointed at reality instead of at the text.

The work is decomposed, delegated to specialists, and reassembled — planning and multi-agent collaboration, running as a daily operating rhythm rather than a research demo.

And underneath all of it sits memory. Every run writes back what it learned: the decision, the failure, the fix, the rule that prevents the repeat. The next run starts smarter than the last one ended. That is the graph that improves every time it runs — the exact missing piece Ng says most people skip.

None of this required waiting for a smarter model. It required deciding that the model's first answer is raw material, not product.

What this means if you run a business

Stop asking "which AI should we use" and start asking "what happens after the AI answers."

If the honest answer is "someone copies it into a document," you own a slot machine. You are paying slot-machine odds and hoping today is a good day. The teams pulling away from you are running the same models you are — through loops that check the work, graphs that connect the loops, and memory that compounds the results.

The practical order of operations has not changed since we started building this way. Pick one workflow that matters. Wrap it in a loop: generate, critique, revise, verify. Put a human gate where the stakes are real. Only wire loops together into a graph when a single loop genuinely cannot carry the job. Let every run leave the Cystem smarter than it found it.

The perfect prompt was always a myth — a hope that one clever sentence could replace a quality process. The loop is the process. The graph is the compounding. The models will keep getting smarter, and it will keep being true that a well-built Cystem around a decent model beats a genius model standing alone.

Build the Cystem. Watch it work.