A prompt-writing tips thread posted by Japanese X user @pekotyan_12 at 6:00 a.m. KST on August 19, 2026 passed 850,000 views. The first post claimed that "if you use ChatGPT, Claude, and Gemini without knowing '#', '###', '*', '""', '[]', '—', and '|', you're throwing away 90% of AI's potential." The thread continues with eight replies explaining how to use each symbol.
Seven symbols and how to use them
According to the original thread, # is a heading symbol that signals a topic change, as in "# Your role" or "# What I'd like you to do." ## and ### create hierarchy within that, as in "## Task details" or "### Step 1." * and double quotes emphasize must-follow rules such as "Don't use jargon," or set off quotes and example text. [] marks a placeholder to be filled in later, as in "Write a proposal for [industry]." — works as a divider between the instruction section and the reference information section, and | is used to request a table-style response, as in "| Item | Details | Priority |."
Where it overlaps with Anthropic's official guide
The direction of these tips goes beyond personal anecdote. Anthropic's official prompt engineering guide states that "XML tags help Claude parse complex prompts unambiguously, especially when your prompt mixes instructions, context, examples, and inputs." In its "advanced" section, the original thread also says that using XML tags such as <role>Your role</role> and <instructions>What I'd like you to do</instructions> conveys structure more clearly than markdown symbols, which is exactly the direction the official documentation recommends.
Take the "90%" figure with a grain of salt
That said, the phrase "throwing away 90% of AI's potential" is the author's personal impression, with no experimental figures or official statistics behind it. The direction, that structured prompts improve how accurately AI interprets your request, matches the official guide, but how much it improves depends on the prompt, the model, and the type of task. Adding symbols to a prompt costs nothing, so the more reliable move is to add just a heading (#) and emphasis (*) to an instruction you already use and compare the responses yourself, rather than trusting the "90%" number.