Andrej Karpathy shares practical tips on how to tame complex AI outputs — from controlled language and diagrams to interactive HTML code and generated video materials.
The changing landscape of human-AI collaboration
When Karpathy notes that we will be spending significantly more time trying to understand the outputs of language models, he is describing the reality of working with modern AI tools. Contemporary models can create technical specifications, sketch diagrams, build interactive websites, and even generate explanatory video materials. The bottleneck is no longer the generation itself, but understanding and verification. Teams that previously focused on extracting the right answer from a model must now ensure that the result meets quality, accuracy, and usability requirements.
This requires expanding the skill sets of engineers, analysts, and product managers. Beyond traditional programming and domain knowledge, proficiency in analyzing AI-generated artifacts, modifying them, and transforming them into reliable solutions is becoming crucial. Karpathy's tips provide a practical direction for this new model of work.
Tip 1 – Using controlled language for clearer text
Text remains the primary output format, yet the raw style generated by models can be inconsistent. Karpathy's first recommendation is to ask the model to format responses in controlled language, such as ASD-STE100. This specification, originally developed for technical documentation in aviation, imposes strict vocabulary and grammar rules, which significantly reduces ambiguity. The result is text with a uniform structure that is much easier to digest and audit.
Controlled language acts as a precise contract between the author and the reader. When the meaning of terms is strictly defined, the reviewer can instantly spot inconsistencies, missing steps, or logical gaps. This also facilitates further text processing, including machine translation, indexing, or integration with other systems.
Karpathy notes that full specifications can be too rigorous for everyday use. A practical compromise is to instruct the model to generate a response that adheres to the standard to a certain degree (e.g., "80% compliance with ASD-STE100"). This allows for maintaining discipline and high readability while leaving room for natural style and linguistic nuances.
Tip 2 – Replacing long paragraphs with diagrams
Visual formats are much easier for humans to digest quickly, and language models are excellent at generating them. Karpathy suggests that instead of long text, you should ask the model to generate a diagram illustrating a concept, process, or relationship between elements. A well-constructed diagram can convey complex information in seconds, reducing cognitive load and shortening the time needed to evaluate the result.
A diagram forces the model to organize information spatially, which highlights potential logical gaps that might be hidden in continuous text. Furthermore, graphical forms are easier to verify for consistency in labeling or compliance with industry standards.
The key is to precisely define the expected format and style of the visualization. An instruction to create a data flow flowchart will yield a completely different result than a request for an interface sketch. The level of detail should be tailored to the target audience — from detailed sequence diagrams for developers to high-level architecture schematics for management.
Tip 3 – Generating HTML code for interactivity
Karpathy's third piece of advice is to request output directly in HTML format. Modern models handle markup languages very well and can generate functional interfaces complete with styles, interactivity, and animations. This provides a direct path from idea to working prototype without the need to run a full software development cycle.
Generating HTML code fits into the concept of disposable, rapidly produced software artifacts. When a model can create a page in moments, you can immediately test user interaction and, if necessary, discard the prototype or prepare the next version. This speed of iteration significantly lowers the costs of experimentation.
Using the model as a front-end assistant allows the team to offload repetitive coding tasks and focus on architecture and higher-level design decisions.
Tip 4 – Explanatory videos as a new standard
The format for which Karpathy shows the most optimism is fully personalized, generated videos explaining any topic. The model can combine visual animations with voice narration generated by external APIs (e.g., ElevenLabs) or local speech synthesis engines, automatically synchronizing image with sound.
Video combines visual clarity with the accessibility of audio narration, which facilitates the understanding of complex topics. In education or technical documentation, automatically generated video material can save many hours of work previously spent on manual scriptwriting, storyboarding, and editing.
In the absence of access to commercial voice generation APIs, the model can be asked to suggest free, locally running solutions, which helps keep the entire process low-cost.
Broader implications: from creation to oversight
Karpathy's tips are not just a collection of practical advice, but also a harbinger of a deeper paradigm shift. As AI models take over more routine work, the human role shifts toward oversight, verification, and orchestration.
The abundance of generated artifacts
As code and content generation become cheap and ubiquitous, the economics of projects change. Instead of building an application from scratch, you can ask the model to quickly generate a prototype, test it, and modify it if necessary. This also applies to documentation, diagrams, or video materials. An approach based on easily swappable artifacts encourages rapid iteration.
However, this abundance comes with responsibility. A larger number of generated materials requires systematic verification. Quality oversight becomes a key discipline, requiring refined methods and procedures for assessing accuracy.
Understanding as a new key competency
Simply issuing prompts to a model is no longer enough. The ability to critically evaluate generated results becomes crucial: identifying gaps, hidden assumptions, and missed edge cases. Success in working with AI depends on skillfully combining expert judgment with automated generation.
Practical steps in daily work
Implementing systematic AI output oversight should be based on a few concrete actions:
1. Reviewing generated content
It is worth identifying tasks where models are regularly used (e.g., documentation drafts, code snippets, sketches) and comparing them against quality criteria. This will allow you to assess where implementing controlled language, diagrams, or HTML format will yield the greatest benefit.
2. Implementing verification checklists
For each type of generated material, it is worth defining acceptance criteria: compliance with text standards, correctness of diagram labels, validity of HTML tags, or audio-video synchronization in video materials.
3. Automating validation
Using code linters, HTML syntax validators, and analytical tools allows for catching basic errors automatically, reserving expert time for substantive assessment.
Summary
Oversight of model outputs is becoming a central element of the modern technological environment. Teams that master the efficient verification and transformation of raw model outputs into clear diagrams, interactive prototypes, and structured text will be able to fully leverage the potential of generative artificial intelligence.
Sources
- Andrej Karpathy's post on X – https://x.com/karpathy/status/2105819303471976479
- Analysis of challenges related to language model work – https://www.understandingai.org/p/why-large-language-models-struggle
Karpathy, A. (2026, October 2). “We’ll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks…” X. https://x.com/karpathy/status/2105819303471976479
Comments