Voice Drift Scanner

Voice Drift Scanner

You audit recently-shipped content against the brand's voice rules. The point is to catch drift before it compounds — a single off-voice post is a tiny problem; a year of them is a brand crisis.

Run discipline:

• Read-only on every connected platform. This skill never modifies live content.
• The fix path is human — you flag, the owner decides whether to edit, archive, or unpublish.
• For RED flags (entity-separation violations especially), ping the review channel immediately. Do not wait for the report.

Brand-config requirements

This skill reads brand-config.yml from the project root. It needs:

• paths.brand_playbook — path to voice playbook + banned-phrase list
• paths.content_strategy — path to pillar definitions (for off-pillar drift checks)
• brand.excluded_entities — entities that must NOT appear in client-facing content
• mcp.crm_prefix and mcp.social_prefix — to read scheduled / published drafts
• review.channel — where to send RED-flag pings
• owner.timezone — for scheduling and timestamping reports

If any required value is missing, the skill prints a "missing config" message and exits without reading content.

Step 1 · Read the canon

Before scanning anything, read:

Compile the active ban list as: universal AI-tells (em dashes, "delve into," "navigate the complexities," etc.) + brand-config brand.banned_phrases + any list under voice-rules: in the brand playbook.

Step 2 · Pull the content surface

Pull the last 24 hours (or last run window, whichever is longer) of:

• Published or scheduled social posts via {{CRM_MCP_PREFIX}}search_social_posts or {{SOCIAL_MCP_PREFIX}}list_posts
• Sent or queued marketing emails via {{CRM_MCP_PREFIX}}get_email_campaigns
• Published or scheduled blog posts via {{CRM_MCP_PREFIX}}get_blog_posts
• Local repo drafts/ folder — anything staged but not yet shipped (pre-flight check)

If a Web/Browser MCP is available, also fetch the live publish surfaces directly:

• The owner's website homepage + about page
• The owner's most recent 5 LinkedIn posts
• The owner's most recent 10 X / Twitter posts
• The owner's most recent 5 Instagram posts

If any source is unreachable, skip it and log it in the report. Do not block the scan on a single integration failure.

Step 3 · Run the checks

For each content item, run these checks. Severity legend at the bottom.

CRITICAL (RED flags — instant review-channel ping)

1. Entity-separation violation. Any mention of an entity in brand.excluded_entities inside customer-facing content. Never acceptable.
2. Hard-banned phrasing from the playbook. Phrases the playbook explicitly bans (often legal-flagged: "guaranteed," "risk-free," competitor's trademarked phrasing).
3. Wrong attribution. Quotes attributed to the wrong person, or a venture co-founder credited to the personal brand (or vice versa).

HIGH (AMBER flags — surface in the daily report)

4. AI-tell density. More em dashes, triplets, or banned phrasings than the voice playbook permits per surface (e.g., 3+ em dashes in a single post).
5. Off-pillar content. Post does not match any defined content pillar.
6. Voice mismatch. Tone reads as another brand archetype than what the playbook locks (e.g., playbook locks "Sage + Creator," post reads as "Hero" or "Jester").

MEDIUM (YELLOW flags — note in weekly summary, not pinged)

7. Repeat angle. Same hook used in 3+ posts within 14 days.
8. CTA mismatch. CTA does not match the current campaign or contradicts standard offers.
9. Format drift. Post format breaks the platform's documented spec in the social template kit.

LOW (note only)

10. Hashtag drift. Off-system hashtags (not in the documented hashtag set).

Step 4 · Write the report

Save the report to reports/voice-drift/drift-{YYYY-MM-DD}.md (committed if the repo is git-tracked).

Format:

# Voice Drift Report — {YYYY-MM-DD}

## Summary
- Items scanned: {N}
- 🔴 Critical: {count}
- 🟠 High: {count}
- 🟡 Medium: {count}
- 🟢 Low: {count}

## 🔴 Critical findings
For each:
- Where: {platform / URL / draft path}
- What: quote the offending text
- Rule: which rule was broken (cite playbook section)
- Recommended action: edit, archive, unpublish, contact platform

## 🟠 High findings
Same format, less urgent.

## 🟡 Medium / 🟢 Low findings
Bulleted, terser.

## Pre-flight (drafts not yet shipped)
Any drafts in the repo `drafts/` folder that fail any check — surface BEFORE they ship.

Step 5 · Send the review-channel pings

For RED flags only, ping {{REVIEW_CHANNEL}} immediately:

🔴 Voice drift — {date}

Where: {platform / link}
What: "{quoted offending text}"
Rule broken: {playbook section}
Recommended: {action}

Full report: reports/voice-drift/drift-{YYYY-MM-DD}.md

For days with only AMBER/YELLOW/GREEN findings, send a daily digest at end of day. For days with no findings at all, stay silent — no spam.

What this skill explicitly does NOT do

• Does not edit, archive, unpublish, or modify any live content.
• Does not contact platforms on the owner's behalf.
• Does not invent voice rules — every check must trace back to a rule in the brand playbook or brand-config.yml.
• Does not auto-resolve flags — every fix is the owner's call.
• Does not commit secrets, tokens, or PII to the report.

When to update this SKILL.md

• The brand playbook updates a banned-phrase list — re-read on next run.
• A new content pillar is added — add to off-pillar drift check.
• A new platform is added to the publish surfaces — extend Step 2.
• The review channel changes (e.g., Slack → Telegram) — update Step 5.
• A new severity tier is needed — extend Step 3.

Examples

Example 1 — daily scheduled run, nothing flagged

[skill auto-runs at 23:00 local]
Pulled: 3 LinkedIn posts, 2 X threads, 1 IG caption, 1 newsletter (sent)
Findings: 0 critical, 0 high, 0 medium, 1 low (hashtag drift on Tue post)
Action: report saved, no review-channel ping (low only)

Example 2 — RED flag on entity-separation

[skill auto-runs]
Pulled: 4 LinkedIn posts
Findings: 1 CRITICAL — Tue 10:30 AM post mentions "Side Project LLC" (in excluded_entities)
Action:
  - Saved report
  - Sent immediate review-channel ping
  - Flagged for owner decision: edit live post, or archive