Crea Tu "Segundo Cerebro" con Este Sistema de Archivos

Main ideas / lessons

• To get useful outputs, you must “talk to AI as if it were a person,” not as a search engine. Generic prompts lead to generic “brochure” answers.

• The key problem isn’t that the AI lacks knowledge of the world; it lacks knowledge of you and your specific situation. You must provide the relevant context and constraints.

• Use a consistent prompting workflow (“four rules”) plus a file-based system so corrections persist over time.

• If you implement the file system properly, you’ll eventually need to send only the task each time (the system already contains the rest), making the process highly automated.

---

Methodology: the “4 rules” for prompting AI (with examples and what to do)

1) Give context before instructions

Provide background information so the AI can tailor the response to your reality.

Include context such as:

• Who it’s for (audience)
• Tone of voice to use
• What should/shouldn’t be said (your house style, taboo phrases, etc.)
• What you already tried and what didn’t work
• Any distinguishing details that make the output relevant to your situation

Avoid the mistake:

• Asking only the instruction/task with no context (leads to generic results)

Example (dental clinic social post):

• Bad prompt: “Write me an Instagram post about the new invisible orthodontics service.” → Output sounds bland/generic and could fit any clinic.

• Better prompt (same task, with context):

  • You manage the clinic’s social media in a specific neighborhood
  • Target audience age range and behavior (families 35–55 who use Instagram for entertainment)
  • Rules like “never put prices in the caption”
  • Avoid certain phrases that signal “expensive clinic” tone → More differentiated, usable output.

---

2) Add limits/boundaries (“tell it what you don’t want”)

Prevent the AI from “filling gaps” with extra content, made-up details, or assumptions.

Explicitly forbid behaviors that would cause problems, such as:

• Inventing prices / made-up data
• Adding services nobody asked for
• Using a salesy tone or add-ons that you didn’t request
• Providing multiple proposals when you requested one

Core principle: If you don’t state a “no,” the AI will assume it’s allowed and may generate unwanted extras.

---

3) Ask for the plan/approach before writing

Require the AI to outline its structure first, then you approve and only then it writes.

Instruction workflow:

• Before writing anything, request:
  • How it will approach the task
  • The planned structure/sections
• Then confirm, and only afterward let it produce the final draft.

Benefits:

• Fewer mistakes
• Saves time (“correcting the approach” is quick; correcting a finished output is slow)
• Saves tokens/costs (especially relevant if using APIs)

Example (blog post about social media management costs):

• Bad: Send only the request → AI writes polished text for “big agencies,” not your real audience.
• Good: First request (in ~5 lines):
  • who it’s for
  • how it will approach it
  • what sections it will have → Then you approve, and the first version comes out closer to what you need.

---

4) Correct and store fixes in a file (not the chat)

Turn good prompting into a durable system that keeps improving.

Key rule:

• Corrections must go into your persistent documentation (a .md file in your working folder), not only the chat.

Why:

• Chats lose context when reopened/new chat starts.
• Documentation acts like a “system memory” the AI reads each time.

How the system improves:

• When you correct the same error once in the file, it doesn’t recur repeatedly.
• After several corrections, the documentation reflects real issues from your workflow—making outputs progressively better.

Example (VAT mistake recurring):

• Wrong method: Correct inside chat; next week the same mistake happens again.
• Right method: Add a line to the .md file, e.g.:
  • “Prices are always exclusive of VAT.”
  • “21% VAT is applied to the invoice, never to the quote.” → Mistake stops recurring.

---

The “twist” (what becomes different if you follow the system)

• Once you implement the file-based system and the four rules are captured in documentation:
  • You can abandon heavy prompting in practice
  • You only send the task in the chat each time
  • Everything else (context, limits, plan request behavior) comes from the stored file
  • This enables automation and consistency

---

How information is organized in the workflow (where each rule “lives”)

• Context: written once in the .md file in your folder
• Limits/boundaries: written once in the .md file
• Plan request behavior: written once in the .md file
• Task: included each time as the only thing in the chat message
• Bug/error corrections: documented in the .md file when they appear, so the system improves and doesn’t repeat the same mistakes.

---

Audience / use-case guidance mentioned

• If you use AI only occasionally, you may not need the full system.
• The system is recommended for people who:
  • work regularly with AI
  • already organize life/work using folders or structured documentation
  • want automation and reduced mental load

---

Speakers / sources featured

• Fernando (main speaker; narrator of the video content)
• Rubén (referenced indirectly at the end: “as Rubén always says…”)
• “Yoa training” (course mentioned as a link in description/comments; no separate speaker named)