AI
AI-Generated “Workslop” Is Destroying Productivity
- Workslop = “AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task”
- Instead of polishing good work, people use tools to generate content where the burden of evaluating quality is shifted to the receiver.
- The workslop “tax” is paid by the consumer — decoding the content, determining if context is false or missing, etc. You can spend more time trying to get the actual answer.
- There’s an interpersonal cost when you send your coworkers workslop (you could be viewed as less creative, capable, trustworthy, or intelligent).
- Solutions: Develop guidance for employees to use AI tools (instead of “AI everywhere all the time”), use AI purposefully rather than using it to avoid doing work, frame AI as a collaborative tool rather than a shortcut, uphold the same standards of excellence whether done by humans or AI.
Generative AI Foundations: Generative AI in Action
This course covers scenarios when common GenAI tools are useful and reasons they fall short for specific tasks. I learned about some sites (theresanaiforthat.com, futurepedia.io, producthunt.com, LMSYS) to help navigate the expanding landscape of AI tools and services. It also covers various factors to consider when evaluating such tools.
Assessing Data Readiness for Generative AI
This course covers how to assess and prepare data that will be used for AI solutions.
95% of People STILL Prompt ChatGPT-5 Wrong
- ChatGPT now has a router to pick which underlying model to use. Adding “think hard about this” will force the router to select a higher reasoning model. Use this for higher risk tasks.
- Specify output length to control verbosity. Shorter for critical information (100 words), medium for key takeaways (3-5 paragraphs), longer for comprehensive summaries (600-800 words).
- Use better prompts (i.e., meta prompting) by adding structure (instead of blocks of text), eliminating vagueness, adding error handling to handle conflicting instructions or missing data.
- Use XML to specify sections — e.g., <prompt>, <task>, <context>, <requirements>.
- Ask AI to create its own definition of excellence (rubric), grade its own work, and iterate until it creates the best result.
Software development
This Mental Model Changed How I Design Software FOREVER
- Acceptance test driven design (ATDD) is even more important with agentic AI.
- Code quality is still important even if AI writes it.
- Principles you can add to your AI tool instructions:
- Prefer simple, clean, maintainable solutions over clever or complex ones.
- Readability and maintainability are primary concerns.
- Use self-documenting names and code.
- Use small functions.
- Follow the single responsibility principle in classes and functions.
- Important distinction: When you ask AI to modify code, it’s not refactoring — it’s rewriting.
- A software engineer’s fundamental jobs are to manage complexity and learn about the problem being solved. This comes from modular code with loose coupling.
- Mental model:
- Start with user needs.
- Break down the problem into parts that can be solved separately.
- Take a part and specify the needed behavior in some detail.
- Specify coding standards.
- Get the tool to write the tests and code.
- Make improvements if needed.
- Move to the next part.
- This model is very similar to TDD — an empirical engineering approach for software design. It forces you to focus on a specific aspect which leads to limited WIP, reduced cognitive load, and smarter use of AI’s context window.