The brand voice enforcer
If you’ve been publishing for years, your voice has drifted — it always does. New writers join, AI drafts slip through, and you write differently tired than fresh. The cumulative effect is an archive that reads like several different people, which quietly dilutes the brand. Style guides exist to fix exactly this, but nobody enforces one across hundreds of old posts because the manual work is impossible. An assistant that has actually read your guide can flag every deviation in an hour. The goal isn’t to sand every post into the same shape — it’s to catch the obvious lapses: passive voice in a brand that prides itself on active, jargon you’ve banned, stiff phrasing in a brand that’s meant to be punchy.
What you need
- Specter synced to your blog
- An AI assistant you already use
- A written style guide (one to four pages — rules to never break, rules to prefer, example rewrites)
The recipe
- Pull and stage. Run a Specter pull and commit a git baseline. Drop your style guide in alongside the posts.
- Run the analyzer (no edits). It produces a per-post report of deviations with severity and a proposed fix for each — no files changed.
- Triage the report. Mark each finding
accept,reject, ormanual. Aim for roughly a 50% accept rate; some “deviations” are personality. - Run the fixer, dry-run, push. The fixer applies only accepted findings; preview with a Specter dry-run and push.
The prompt
The analyzer is read-only:
Audit this archive against the style guide provided. For each post, find every
phrase that DEVIATES from a guide rule and record: file, line, the rule violated
(quoted), severity (high = breaks a "never" rule, medium = breaks a "prefer"
rule, low = defensible), the original phrase, a proposed fix, and a one-line
reason. Skip anything inside code blocks, blockquotes, direct quotes, or
headings. Write findings grouped by file, with a summary of totals by severity
and the most-violated rules. Do NOT modify any files.The fixer touches only what you approved:
Apply the accepted voice fixes from the annotated report. For each finding
marked accept: find the original phrase, replace it with the proposed fix, and
re-read the surrounding paragraph — make a minimal connective adjustment only if
the fix broke flow. Do nothing for findings marked reject or manual. Log every
applied change with the paragraph before and after. Change nothing else.Cost and time
| Blog size | Tokens (analyzer + fixer) | Cost | Wall-clock |
|---|---|---|---|
| 50 posts | ~150k + ~80k | $0.50 | 20 min |
| 200 posts | ~600k + ~320k | $2.20 | 75 min |
| 600 posts | ~1.8M + ~960k | $7 | 3.5 hr |
Most of the wall-clock is your review time — right for a judgment call like voice.
Pitfalls
- Over-application. Accept every finding blindly and you’ll homogenize the archive into something uncanny. Some deviation is personality.
- Quotes and code. Verify the analyzer actually skipped code blocks and direct quotes before pushing.
- Guest posts. Leave them alone unless the author has signed off — reputational risk.
Where to go next
It pairs well with the conversion copy sweep — one combined pass keeps new CTAs from breaking the voice you just enforced.