The Night I Copy-Pasted My Style Guide for the Tenth Time

It was late, maybe eleven, and I was on the tenth article for the week. I opened my prompt template, scrolled past the git diff, and started pasting. Fifty-something lines of formatting rules: no AI template openings, no numbered conclusions, max two em-dashes, specific word swaps for Indonesian voice, TTS accessibility constraints. Same paste I'd done for every article that week.

The draft came back. I caught myself deleting an em-dash in paragraph three. Then another in paragraph six. The agent had used a numbered list as a closing section again, despite the explicit instruction not to. I rewrote the opening—it still started with a definition instead of a moment.

The prompt was fifty lines long, and the agent kept breaking the same five rules.

I'd assumed the fix was a better prompt. More specificity, stricter wording, maybe examples. So I made the instructions longer. Sixty lines. Seventy. The agent followed the new rules and broke the old ones. Context window is finite, and every token spent restating what the agent already "learned" from the system prompt is a token stolen from the actual task.

What editorial style actually is

Editorial voice is procedural knowledge. It's not a fact the agent needs to look up, it's a way of doing something, repeatedly, with consistency. The same way a junior developer doesn't learn code review by reading a one-page summary, but by sitting next to someone who says "no, don't do it like that" enough times until the pattern sticks.

Anthropic's engineering team describes Agent Skills as "organized folders of instructions, scripts, and resources that agents can discover and load dynamically" [12]. The metaphor they use is onboarding a new hire. You don't hand a new hire a seventy-line document in their first email—you give them a folder they can refer to as needed, organized by domain.

The best-practices documentation for skill authoring puts it bluntly: "The context window is a public good" [15]. Only add what the agent doesn't already know. My main prompt was spending most of its real estate on style rules the skill could carry.

Google's documentation style guide arrived at the same problem from a different angle. They mandate active voice [13], not because passive is grammatically wrong, but because it hides who's doing the action. Same insight: some rules aren't about correctness, they're about producing a consistent, recognizable voice. That kind of knowledge needs a permanent home, not a per-prompt append.

What changed

One file: blog-id-style/SKILL.md. All the formatting constraints, word swaps, TTS rules, and hard-earned anti-patterns live there now. My writing prompt shrank to three lines that reference it.

The main prompt dropped from seventy lines to about twenty. It focuses on the task: extracting narrative from a commit, structuring the argument, finding the specific moment that carries the article. The skill handles everything else. No more copy-paste marathon.

I wrote about this pipeline before, in the AI agent skills research stack and the avoid-ai-writing skill wiring. Verification didn't disappear. Automated checks still catch violations that slip through. But the skill is the first layer, and it's the single source of truth. When I update a style rule, I change one file, not seventeen prompts.

Why one file changes everything

EYD Fifth Edition launched August 18, 2022 [14]. The Indonesian spelling guide returned to the EYD name from PUEBI, and more than half the rules changed. Even official language evolves, expecting a static prompt to capture a living editorial voice was always wrong.

The lazier alternative would've been a shared text file with style notes, referenced in every prompt. But that's not discoverable. Agent skills are structured, loadable, and scoped. The difference matters when your context window is the bottleneck.

I used to think good writing prompts were comprehensive. Now I think they're short. The skill carries the weight. The prompt carries the intent.

Sources:

• Anthropic, Equipping agents for the real world with Agent Skills [12]
• Google developer documentation style guide, Active voice [13]
• Balai Bahasa Kaltim, Peluncuran EYD edisi V [14]
• Claude Docs, Skill authoring best practices [15]