Begins (topic)

Begins is a framework for building with AI that focuses on starting with the simplest, most effective solution for your specific task. Instead of jumping into complex systems, you first decide where the AI will sit: either in a project (a standalone task) or in a skill (a reusable process). For beginners, projects are recommended because they are simpler. The core idea is that you should have the simplest solution that achieves the task at the quality you have set.

What matters for AI development is that Begins forces you to understand your process before adding AI. One expert advises that you "need to understand the steps, what works and what doesn't before you ever introduce AI." This prevents the common mistake of letting AI write its own prompt or giving you an average answer. By starting with a clear, simple structure, you also avoid the hidden cost of letting AI run without oversight. The build phase, where you create skills (reusable instructions) for your specific tasks, takes about 80% of your time. Only after mastering this do you move to borrowing skills from others. This approach keeps you in control, ensuring the AI works for you, not the other way around, and helps you produce results that match your quality standards.

To use Begins, you start with the task itself. Wrap your instructions and any background material in XML tags (markers like `<task>` and `</task>`) so the AI knows exactly what content to process. This is the "begins" step: open with a clear, direct instruction, then place your context inside those tags.

Next, you "try" your prompt. Run it once to see if the output matches what you expected. If the AI understands the task but misses the format or style you wanted, add a one-shot or multi-shot example. A one-shot gives one concrete example; multi-shot gives three to five diverse examples, such as one normal case and one edge case. Examples turn the abstract instruction into a pattern the AI can imitate.

If the AI still confuses pieces of information, you "escape" the problem by adding more XML tags to compartmentalize instructions, context, and input separately. Use the symptom to choose the smallest fix: if context gets mixed up, add tags; if the pattern is wrong, add examples.

Finally, "lab" means testing and iterating. After your first attempt, review the output. If it's not right, refine—you don't need to start over. Ask the AI to ask you clarifying questions before it begins. Use a prompt like "Before you start, ask the five most important questions that would help you do this well. After I answer, then begin." This builds the right foundation and saves corrections later.

When you begin using an AI agent (an autonomous system that completes tasks for you), common pitfalls include skipping setup, failing to escape looping behavior, and ignoring safety caps. Many beginners try to start with a perfectly organized notebook, but modern tools like NotebookLM now allow you to begin with loose ideas or rough research directions, letting the agent help build the structure for you.

One major mistake is running an agent without a safety cap (a hard stop after a set number of turns). You should always add something like "stop after 20 turns" when you set a goal. Otherwise, the agent can enter an endless loop, running in circles without making progress. If it breaks, look at the skill files or instructions you gave it — a large file with many instructions often makes the AI worse. Start fresh with a small prompt and it behaves better.

Another best practice is to use the escape hatch effectively. When you identify an incorrect path, treat every failure as evidence and adjust your approach. Use prefilling (starting the AI's response with an opening marker) and closing the last marker to prevent the agent from adding unwanted commentary. You can also clear context every so often to avoid context run (when the memory buffer gets too full and degrades performance).

Finally, investigate why the agent gives you output you don't want. Examine your habit of coding first — clarify the requirement and define success before implementation detail. Ask who uses the system and in what environment, then adjust your agent accordingly.