A scientific example currently shows very clearly where the strengths of prompt-driven work lie – and where its limits are.Harvard physicist Matthew Schwartz has described how, using the open-source tool BootLoops and the language model Claude, he produced 36 manuscripts in 18 subject areas within three months, from particle physics to linguistics.However, according to his account, the results only became scientifically valuable through human control.His conclusion is unmistakable: every result must be verified independently.

For companies integrating AI into development or content processes, this is the more relevant message than the number of manuscripts.For it does not describe a model that works autonomously, but a workflow in which humans and tools have clearly distributed roles.

What a harness actually does

In AI practice, the term harness – literally: harness or harness system – refers to a layer built around a language model.It organises what the model should do and when, hands over tasks, receives results and passes them on.BootLoops is exactly such a tool: it supports AI models with exact scientific calculations and is available as open source.

The point behind it is more important than the individual tool.A language model produces text that sounds plausible.Exactness is not a built-in feature, but something that must be produced from outside – through verification steps, through tools that calculate rather than formulate, and through people who hold the result against reality.A harness is the organisational framework that reliably repeats precisely these steps instead of leaving them to the chance of individual prompts.

The difference between quantity and value

36 manuscripts in three months is a remarkable throughput figure.In itself, however, it says nothing about quality.That is exactly what makes the described case instructive: the output was produced quickly, the value was produced more slowly – namely where an expert provided direction, discarded, refined and recalculated.

This pattern can be transferred one-to-one to projects.AI shifts the effort, it does not eliminate it.What used to go into creating a first draft now goes into specification, evaluation and correction.Anyone who does not plan for this effort gets a lot of material and little usable result.

Vibe coding needs the same framework

Vibe coding refers to development in which code is predominantly generated via prompt and only superficially reviewed.This often works surprisingly well for prototypes and small utility scripts.In client projects with data protection requirements, interfaces and multi-year maintenance, it is not enough.

The lesson from the scientific example applies here just as much: it is not the model that decides quality, but the framework in which it works.For software projects, this means concretely:

  • Automated tests as the first verification instance, so that generated code not only compiles but also does the right thing.
  • Binding code review by people who know the professional context – not as a formality, but as a decision step.
  • Clear specification before the prompt, because vague tasks lead to vague results, only faster.
  • Traceability: it should be documented which parts were created with AI support and who reviewed them.
  • Tools for exact tasks where accuracy matters – calculations, validations and migrations do not belong in the hands of a text generator alone.

What this means for the division of roles

Organisationally, this development amounts to a shift in competence profiles.What is needed less is the ability to produce lines quickly, and more the ability to assess results.Professional depth is not devalued by this – it becomes a prerequisite for being able to meaningfully evaluate AI output at all.

This also affects project planning.If review becomes the actual bottleneck, it is not a good idea to treat review capacity as a residual item.It makes more sense to plan verification steps early and firmly and to align throughput with what can actually be verified.

Realistic expectations beat big promises

The described case is not an argument against AI-supported work – on the contrary.It shows that with a good framework, surprisingly much can be achieved, including in demanding domains.But it also shows that the framework is not optional.The idea that one can commission a model and reuse the result unchecked does not stand up to practice.

For web and software projects, this results in a sober calculation: AI accelerates drafts, variants and routine work, but shifts the bottleneck to verification and control.Projects that plan for review, tests and clear responsibilities from the outset therefore benefit significantly more than those that rely on pure output.

Anyone introducing AI should therefore not first ask about the best model, but about the framework in which its results become verifiable.

Sources