← All disciplines

Show before you tell.

A designer does not guess your style. You show them the thumbnails you like and the ones that earned clicks. Context engineering is handing that over before the work starts.

The idea, in one hire

You have hired a graphic designer for thumbnails. You do not say "make it nice". You show the ones you like, the ones that got clicks, the ones that flopped, and you say what your channel is about. Then the first draft is close, instead of the fifth.

Context engineering is exactly that, done deliberately: the material, before the work. The model cannot know your audience, your past work, your taste, or last month's numbers unless you hand them over.

What counts as context

  • Examples of good output — three or four, with a line on why each one worked.
  • Examples of bad output — more useful than the good ones, and almost nobody supplies them.
  • Real numbers — pulled through a connector, not remembered by me.
  • The audience — who they are, what they already believe, what they are tired of.
  • The constraints of the platform — length, format, what gets suppressed.

The prompt I use to test content before it goes out

This is the one I actually run. Claude writes the thing; Meta AI tests whether the audience would accept it. Two tools, two jobs.

What this prompt is for: judging a post before you publish it, as your audience would — not as a writer would. It is the cheapest way to find out that a script is boring without losing a week of posting.

Testing a script against a real audience
You are a social media growth specialist with more than five years of experience.

I will share my profile ID. Analyse every post and tell me which ones work and which do not,
look at my followers, and build a virtual simulation of my audience: what they care about,
what they scroll past, what makes them comment.

Then I will paste a script. Judge it as that audience would, not as a writer would.
Give me the top 10 problems with it, in order of how much they cost me.
If something is fine, say nothing about it.

Where it goes wrong

  • Context with no filter. Dumping forty pages into the window is not context, it is noise. Five good examples beat forty average ones.
  • Context the model cannot use. A screenshot of a table is worse than the table.
  • Stale context. Last quarter's numbers presented as current. Connectors exist so this does not happen.
My rule: if the work needs to sound like me, the model needs to see me — my old posts, my audience, my last month. Otherwise it writes for an average nobody.
← Prompt Loop →