The week I made it impossible for Friday to sound like AI

ARTICLEGROWTH TACTICFRIDAY (AI CHIEFOFSTAFF)

Ronsley Vaz

4/24/20262 min read

Every draft Friday produced had tells.

The same phrases I'd struck out of twenty drafts before. The same punctuation I'd replaced every time. The long dash. The stock generic openers. The throat-clearing AI phrases.

I was the one catching them. Every time. That's not a chief of staff, that's a copy editor.

The Voice DNA doc had the list. Friday had read the doc at the start of every session. She still wrote the tells. Because asking an AI "please don't use these phrases" is not a rule, it's a request. And LLMs are polite about requests. They say yes and do it anyway.

I didn't want a polite Friday. I wanted a Friday who couldn't ship AI-sounding prose if she tried.

What shipped

Six things.

1. A linter that scans every draft. Looks for the long dash, the stock phrases, the giveaway openers, the corporate filler AI loves to reach for. Flags every hit with a line number. Zero tolerance, no judgement, deterministic.

2. A gate that blocks bad writing from being written. Before Friday can save any file in a content folder, the linter runs. If it fails, the write is blocked. Not flagged. Blocked. The agent cannot ship what can't pass.

3. An automatic review on every draft. Every time a draft saves, a review file appears next to it, scored 0-10 by the linter. No extra step, no asking for feedback. The report is there before I am.

4. A validator for every agent Friday uses. Each of the nine agents was checked against a required shape (name, description, tools, model). No agent could silently run without the right configuration.

5. Golden voice samples. Six tagged samples of my writing, each scored against both the linter and the deeper voice review. The fast path finishes in under a second. The slower path runs when I want the full score. Both have to stay green.

6. A banned-words skill. Every agent now loads the banned list at session start. Before the deeper review spends a single LLM token, the linter has already caught the obvious misses.

What it unlocked

  1. I stopped being the copy editor.

  2. Friday couldn't write the long dash. Couldn't reach for the corporate AI filler. Couldn't open a paragraph with a stock phrase. The tools said no before anything hit disk.

  3. The drafts that reached my inbox were already stripped of the noise. I could focus on whether the idea worked, not whether the prose sounded human.

  4. And when she slipped, the feedback loop was instant. A review file appeared with the line number and the fix. I didn't have to teach her twice.

Copy this if you're building your own

  • Don't trust your AI to avoid tells because you asked nicely. Write a linter. A plain list of phrases and patterns your AI must never use. A tiny script that scans every draft for matches.

  • Then wire it as a gate. Not a suggestion. Not a warning. A lock. If the draft fails the linter, the save fails. The agent cannot ship what can't pass.

  • Your AI will complain. It'll call your rules too strict. Ignore it. The first time you ship something and nobody writes "reads like AI" in the comments, you'll know why the gate is worth it.

Episode 2 of 19. I'm walking through every sprint I ran to build Friday. New episode every Monday, Wednesday, and Friday through Episode 19. Next: muscle-memory slash commands, the six I type a hundred times a day.

Want to start your own Friday? Grab it free at friday.amplifyais.com.