Prompting basics (general guide)

Practical prompting patterns for beginners: how to ask AI tools for what you actually need.

July 14, 2026 · By Bohol AI

This is a general guide for people new to LLMs. It does not describe a specific club workshop unless we publish a separate event recap with verifiable details.

Prompting is the practice of writing instructions that reliably produce useful output from a language model. It sounds simple — and sometimes it is. But experienced practitioners develop intuitions that take time to build: knowing when to add constraints, when to give examples, when to split a task into steps, and when the model is confidently wrong.

Start with role and format

Before describing the task, tell the model what role to take and what format you want back. "You are a copyeditor. Return a bulleted list of suggested changes." sets clearer rails than "fix this text." The model is less likely to improvise when you have closed the obvious escape hatches.

Patterns that transfer across tools

State the role and output format you want before the task
Paste exact error messages when debugging — not paraphrased summaries
Ask for a short plan before code when the task is multi-step
Say what the model must not invent (facts, libraries, contact info)
Use numbered steps for long instructions — the model follows them more reliably than dense paragraphs
Ask the model to explain its reasoning when stakes are higher — errors become visible

Why constraints matter

A prompt without constraints is an open invitation for the model to fill gaps with plausible-sounding invention. If you are asking about a real person, a specific API, or a current price — say explicitly that if the model does not know, it should say so rather than guess. This one instruction prevents a surprising number of confident errors.

Debugging your prompts

When output is off, narrow the problem before rewriting the whole prompt. Add one constraint at a time. Keep the version that worked as a baseline. Prompting is iterative; treating each run as an experiment is more productive than treating failure as a reason to start over.

Going deeper: follow thoughtful voices

Models change monthly; patterns age better than tool-specific click paths. Treat every output as a draft you still need to verify.

For people who want to go further, two practitioners worth reading regularly are Andrej Karpathy and Boris Cherny. Karpathy (formerly OpenAI, now building independently) writes and posts about how LLMs actually work under the hood — his explanations are technical but genuinely accessible, and understanding the mechanism makes you a better prompter. Boris Cherny, a software engineer who thinks carefully about AI tooling and developer experience, posts useful observations about how these tools behave in practice. Neither writes for Bohol AI — we list them as further reading because their public work is high signal. Links are in the sources below.

The skill is not knowing the magic words — it is learning to read the output critically and adjust.

Sources

All articles