Automation & Agents

Make Your AI's Content Get Read: Write "Attention" Into Its Rules File

AI output quality equals the goal function you define. Write only "be correct" and you get correct, and unread. Here is a rule block and a three-question self-check you can set up in about 20 minutes, applied automatically from then on.

09/21/2026 ｜ 2 min read

One Core Idea

AI output quality equals the goal function you define.

If your rules only say “be correct, be complete, cite sources,” the AI delivers exactly that, and nobody finishes reading it. It isn’t writing badly. You never made “being seen” something it cares about.

That is also why swapping in a stronger model usually doesn’t help: a stronger model just pursues the wrong goal more thoroughly.

You Need a Rules File First

Tools like Claude Code and Cursor support a rules document that is loaded automatically every session. On plain web chat, save the rules as text and paste them at the start of each new conversation.

The difference: pasting a prompt changes this one output; writing the rule into the file changes every output after it.

Steps

Step 1 | Paste this rule into your rules file

Attention before knowledge: no traffic means no revenue, and time is a cost. Anything published that nobody sees is wasted.
Every outward-facing piece (post, video, title, page) must pass the attention gate before completeness:
- A hook in the first 3 seconds (conflict / contrast / broken expectation)
- A reason to keep reading every few sentences (emotion or curiosity)
- An ending with an action the reader can copy directly
Pure knowledge piles count as unfinished.

Step 2 | Add hook formulas, or the AI writes filler

With only a principle, the AI answers “This article will explore…” and calls it a hook. Give it concrete patterns:

• Authority reversal: the official docs or mainstream practice is actually wrong
• Counterintuitive number: the safe option is 3x faster
• Concrete cost: one unread clause makes a whole product line unsellable
• Self-exposed failure: six hours of debugging, and the culprit was my own browser
• Scene inversion / personification: today I fired my AI employee

Shared structure: let the reader assume something more dramatic or odd first, then reveal the truth.

Red line: misleading is not deceiving. The reveal must land inside the same piece, no real people or events as bait, and whatever you promise must actually exist.

Step 3 | Add a post-check

Principles alone get read loosely. Add questions with a pass/fail answer:

1. From the first sentence or first three seconds alone, would I stop?
2. Is there at least one surprise in the middle?
3. Can the ending be copied and executed?

If any question fails, rewrite the opening or ending. Don’t just tweak. Put these three questions in the rules file too, and have the AI answer each with “pass” or “fail” plus one sentence of reason after every draft.

Step 4 | Don’t rewrite old drafts, only swap head and tail

Before publishing, do two things: replace the first sentence using Step 2’s formulas, and add a 30-second concrete action at the end. Leave the middle alone. Those two spots have the best return for the least effort.

Traps You Will Hit

1. Assuming a stronger model fixes it: check the goal before the capability.
2. Principles without checks: with no pass/fail question, the rule may as well not exist.
3. Hooks that become clickbait: if the content can’t deliver on the title, you lose account trust.
4. Overusing emphasis: more than three or four bold or highlight spots per piece and none of them mean anything.

One-Line Summary

Content nobody reads usually isn’t a writing failure. Your rules never had a line that cares about being seen. Put attention into the goal, add a pass/fail check, and the AI actually changes.