Case study

guardrail-skills

Three small checks that respect the reader: they strip the obvious AI-isms from copy and code and leave room for your own voice, and each one can run when the agent decides to or every single time.

Role
Author · maintainer
Timeline
Jul 2026 → present
☆ Star on GitHub
one line of proof
3 guardrails, each runs as a skill or a hook

90-second version

  • Readers spot stock AI phrasing in a sentence or two, and publishing it unedited tells them you didn't think their time was worth five minutes of yours.
  • Three guardrails, slop-scrub, comment-scrub and an orchestration recipe hook, each usable as a skill the agent calls or as a hook your tooling forces on every commit or prompt.
  • Open source on GitHub, and slop-scrub is the same 13-point check I run on this site's copy as a final pass before it ships.
The guardrail-skills repo on GitHub: three guardrails, each with install notes for skill and hook use.
Exhibit GS-1The guardrail-skills repo on GitHub: three guardrails, each with install notes for skill and hook use.
Stack
Markdown skills · Bash hooks · Claude Code · Codex · pre-commit and GitHub Actions
01

Problem

AI drafting is a fine place to start, and the problem shows up when the draft ships as is: stock phrases, dashes everywhere, numbers with no source, comments that restate the line below them, and multi-agent runs that start without a budget or a test plan. Each one is minor, but together they insult the reader's intelligence, and asking the agent to be more careful in a prompt doesn't stick past the next session.

02

Non-obvious insight

The fix is a short, repeatable pass that catches the obvious tells so the writer's own time goes into making the piece sound like them. Writing each rule down as a checklist lets the agent run it when it judges the moment is right, and the same checklist works as a hook that runs every time whether the agent remembers or not.

03

Bets I made

  • Every guardrail works two ways, as a skill the agent chooses to run or as a hook the tooling forces, and neither one depends on Claude Code, so it pastes into any assistant's instructions.
  • The banned-word list is a starting point that people are meant to fork and argue with, because a list tuned to your own writing beats a universal one.
  • The orchestration hook says plainly what it can't enforce: it hands the agent a recipe, and permissions and spend caps still have to come from the runtime.
04

What shipped

  • slop-scrub, a 13-point checklist for writing tells such as stock phrases, dashes, fake depth, unsourced statistics, engagement-bait endings and runs of short verdict sentences, with an optional voice file so a project keeps its author's sound.
  • comment-scrub, a checklist plus the comment-hygiene standard behind it, which strips comments that restate code, apologize for it or bury a commit message in the source.
  • inject-orchestration-recipe.sh, a prompt hook for Claude Code and Codex that supplies a cost and clock estimate, a shared contract, independent review and real tests for each chunk whenever you ask an agent to orchestrate a workflow.
  • Example pre-commit and GitHub Actions setups so any of the three can block a commit or a pull request.
05

Evals and instrumentation

This site is the test bench: its copy gets a slop-scrub pass as a build check that fails on the hard tells, with a test file that keeps the check honest, and the README carries a before and after paragraph that counts each tell the check caught.

06

What broke

The first version of the checks only warned on one kind of fake contrast line, and one slipped onto a live page of this site because a warning is easy to ignore. Those shapes now fail the check outright, which is the case for hooks in the first place.

07

What's next

More checks pulled from the agent fleet that runs MarginFront, Quarry and Ethical Eaters, starting with the review gates those agents already stop at before anything goes out.

Talk to me

Ask me how any of this works.

The fleet, the pricing, the closes. Thirty minutes, on the calendar.