Once (topic)

A "Once" is a skill you freeze (save as a reusable instruction) after the AI has done a task well, so it can repeat that task in the same way later. In the videos, the idea shows up as baking a spec (a written description of what you want) into an AI after you and the AI work out the process together. You have the AI ask one question at a time until it extracts your process from your head, then you get a spec you "bake into an AI." The skill should be kept small, "just small enough so it achieves a task without bloating the AI's brain," and general, because you build it on topic A but want it applied to topics B, C, and D. This matters for AI development because it is how you stop re-explaining yourself. Instead of starting fresh every time, you freeze an example into a skill when it is okay to repeat that task on an ongoing basis. The videos describe a "build phase" where you spend most of your time creating skills for your specific tasks, and a "borrow phase" where you take skills from people inside your company. Without that context — your subject matter expertise, your brain, your intellectual property — the AI is "just a smart intern who's guessing." Once turns that context into something reusable.

When you work with an AI assistant, you often repeat the same setup: think about a topic, decide what to say, then produce the result. A once approach means you capture that work one time and reuse it, instead of redoing it every session. Start with the task. Wrap your context in XML tags (labels that compartmentalize text) so the model knows what material to use. Add examples when the output needs judgment or style. Three to five relevant, diverse examples reliably steer format, tone, and structure, and showing one normal case plus one edge case works better than repeated copies.

For anything involving money (payment and pricing decisions), tell the model to work step by step. First, extract key figures, then analyze trends, then write the recommendation. This visible sequence improves reasoning accuracy. On upgrade (switching to a newer model version), don't float automatically. Pinning plus evaluation turns the upgrade into a controlled decision you can inspect.

Keep a short summary at the top of each page and link related pages together. After running the workflow once, turn it into a reusable skill. Aim for leverage: let the AI handle the repetitive 80%, like topic research, while you do the 20% that adds value, placing human checkpoints there. Also bookmark earlier work so a long thread reuses what it already computed instead of re-deriving the same facts. Every extra example must earn its space.

Once you start making money with AI workflows, the temptation is to upgrade everything at once. Resist that. The bigger mistake is treating early one-time projects as the finish line. Once you have closed a few clients, the next step is turning those one-time jobs into recurring income, even if it is just small add-ons rather than a full monthly retainer. Those add-ons create stability on both sides.

A second pitfall is letting your project knowledge rot. Do not dump every document into project knowledge, because outdated material produces contradictory answers. Curate instead: keep authoritative current material and remove the rest. Vague or contradictory project instructions cause off-target output, so state a clear role, an evidence boundary, and a format rather than making the assistant reconcile conflicting rules. Never leave knowledge unchanged.

Third, do not upgrade a model before the evidence establishes a model-layer failure. If a document refresh precedes confident errors while latency and model versions stay stable, inspect retrieval and indexing first for a broken reindex or mismatched embeddings. Follow the evidence path, not the symptom alone.

Finally, money changes what you should protect. Early on, many people run client work under their own billing and invoice at month's end; that convenience shifts risk onto you. The best practice is to learn quickly and pivot when a project fails, because you learn far more by trying than by waiting.