Enforce a living style guide and terminology

Stand up a **voice-check partner** — paste outbound draft — compare against your glossary and banned-phrases context — emit **violations table + severity + fix hints** — run **before** approval steps ([Set up a content review pipeline](/tutorials/set-up-a-content-review-pipeline)) so reviewers argue substance, not hyphen wars.

Plus: three Admin-Agent passes — diff two drafts for glossary drift, propose **glossary patch bullets** from your last five violation logs, and draft a friendly **creator cheat-sheet** paragraph for Slack when the same mistake repeats.

Audience Everyone · Admins
Time ~10 min
Prerequisites A real style source exists ([Add your first context file](/tutorials/add-your-first-context-file) — product names, casing, words to avoid, and examples of on-brand sentences). Agent basics ([Create an agent from scratch](/tutorials/create-an-agent-from-scratch), [Give your agent its job description](/tutorials/give-your-agent-its-job-description)). Helpful: [Set up a content review pipeline](/tutorials/set-up-a-content-review-pipeline) when checks slot ahead of publish.
You'll end up with One **Style check** agent — output sections **PASS_OR_FAIL**, **Violations**, **Suggested fixes**, and **Glossary gaps** — plus a habit: **you** paste drafts **because you prompted** — optional workflow step ([Run a workflow](/tutorials/run-a-workflow)) before human sign-off.

When a tutorial shows italic text in quotation marks, it usually mirrors a label or helper string inside Auxot. Product copy changes between releases — if something reads differently in your workspace, trust what you see on screen.

Callouts with a Worth knowing gold accent are meant as must-read context before you move on. Blockquotes that open with Tip are lighter, optional depth.

Why this matters

Style guides die in PDF graveyards. Teams still ship “Auxot’s” vs “Auxot is”, revive banned competitors’ framing, or bury required disclaimers, usually five minutes before send.

This agent does not silently read every file in your company. You paste the draft because you prompted: same human-in-the-loop spine as the rest of the series. Its job is cheap mechanical honesty: quote the offending snippet, cite which glossary line it broke, propose a neutral rewrite; you accept or reject fixes (Set up a content review pipeline).

Pair with customer-facing comms (Ship clear customer communications) and internal answers (Build an internal Q&A agent for your team) so one glossary feeds both outward copy and what coworkers hear in Slack.

Consistency is a context file — checking is a prompt.


Quick start

  1. Freeze glossary slice — under ~800 words in one context file if possible; prefer pointers over novels (Add your first context file).
  2. Mint Style check — charter: no opinion on strategy; only glossary alignment; UNKNOWN when rule missing; stop invented policies.
  3. Paste one risky draft — marketing paragraph + support macro; redact customer names.
  4. Read Violations table — if severity feels random, tighten rubric in charter; rerun.
  5. Slot before approval — workflow step order: draft → style check → human approve (Run a workflow).

Done? Ten saved reports: dated, showing shrinking repeat violations; evidence for Run a quarterly review of your agents.


The agent can do that?

1. Diff two drafts for drift

Chat → Admin Agent:

Glossary excerpt: […]. Draft A: […]. Draft B: […]. Table — rows Rule / A status / B status / note — flag **REGRESSION** if B worse — max eight rows.

Why it’s non-obvious: Iteration loops sometimes unfix voice; side-by-side catches regressions after you paste both versions.

2. Glossary patch bullets from violation logs

Violation tables (paste five): […]. Current glossary excerpt: […]. Output <=10 bullets **ADD / TIGHTEN / DEPRECATE** — each cites occurrence count — markdown — no full rewrite.

Why it’s non-obvious: Guides improve from repeat misses; structured deltas beat hallway agreements (Keep your context files honest and fresh).

3. Creator cheat-sheet blurb

Top three recurring violations this month: […]. Draft <=90 words — supportive tone — ends with link to glossary doc — plain markdown — no shame language.

Why it’s non-obvious: Private nagging scales poorly; one channel note reduces repeat coaching (Collect useful feedback on your agents when friction shows up as complaints).


Go deeper

Locale work

Adaptation checks claims and locale tone (Adapt content for new locales or audiences with human review): run style check on adapted output too; two passes, two purposes.

Skills

Promote stable rubric language (Create a Skill): attach glossary; fewer pasted charter walls.

Fleet updates

When glossary meaning shifts: announce agent behavior deltas (Update your agents without breaking the team); stale rules produce noisy false positives.

Privacy

Paste only what review needs: scrub identifiers; align handling (Run a data privacy review before you ship).


Walkthrough

Step 1: Pick one channel pilot

Support macros beat whole blog archive; tight scope trains the rubric.

Step 2: Seed ten intentional violations

Synthetic draft: retired product name, wrong casing, and banned hype word; confirm table catches all three.

Step 3: Tune severity labels

BLOCK vs WARN vs FYI: align with who can override; document in charter.

Step 4: Pair with review pipeline

Reject loop once; confirm style feedback reaches drafting agent description (Set up a content review pipeline).

Step 5: Monthly glossary retro

Merge patch bullets you trust; archive old excerpts; note date in context header.


What’s next

Reference