Skip to content
Baseworks KB

Content Skills Voice Audit — Session Prompt

Created 2026-07-07
Status executed
Tags planpromptskillsvoicecampaignsnewsletterssocial-media

Status: executed 2026-07-07. Results: SKILLS-VOICE-AUDIT-2026-07-07 (findings + applied fixes). Follow-up to VOICE-AUDIT-2026-07-07.

The prompt (copy from here down)

Section titled “The prompt (copy from here down)”

I want you to audit our content-creation skills against the voice-drift findings from the 2026-07-07 voice guide system audit. Read these first, in this order:

  1. 03-resources/voice-guides/VOICE-AUDIT-2026-07-07.md — the audit. The core finding: voice rules exist but enforcement is attention-based, so AI-default patterns (contrastive scaffolding like “X isn’t Y. It’s Z” at high density, repeated sentence templates, meta-narration, unapproved metaphors) survive many review rounds, and fixes applied in one round regress in later rounds. Second structural finding: rules that live only in Claude’s local memory or in stale local copies drift, because other machines and future sessions never see them.
  2. 03-resources/voice-guides/VOICE-GUIDE-UNIFIED.md (v2.8 draft) — especially the new “Essay and blog register — contrast density” subsection and the delete-the-clause test.
  3. 03-resources/voice-guides/WRITING-PROCESS.md (v1.1) — especially the new “Final Voice Pass” section: after content settles, one logged pass purely for voice, separate from content review. This is the fix for cross-round regression.
  4. 03-resources/voice-guides/VOICE-GUIDE-PATRICK.md (v1.2) — the new “Blog Posts and First-Person Essays” section, as the model for what codifying accumulated session feedback looks like.

Then audit these skills and their companion guideline docs. We’ve noticed the same context drift in campaign work: wording and framing slipping off the per-platform trajectories and off the newsletter voice.

Skills (in .claude/skills/): create-campaign, create-newsletter, create-email, write-blog-post, write-communications-post, write-page-copy, post-to-group, draft-response, create-event-post, update-science-content, session-summary.

Companion docs: 02-areas/communications/campaigns/_campaign-creation-guidelines.md, 02-areas/website/blog-articles/_blog-post-guidelines.md, 02-areas/communications/newsletters/_archive-newsletter-guidelines.md, 02-areas/communications/social-media-tags.md, 02-areas/communications/communications-guide.md.

For each skill, check:

  1. Voice loading: Does it require reading VOICE-GUIDE-UNIFIED plus the correct personal guide BEFORE any drafting, including quick inline chat drafts? (This rule has been violated before; it’s now explicit in Patrick’s guide.)
  2. Reference, don’t duplicate: Does the skill restate voice rules in its own text? Local copies go stale and drift; skills should point at WRITING-PROCESS and the guides, and keep only skill-specific structure locally. Flag every duplicated or stale rule you find (e.g., anything still saying em dashes are allowed “sparingly” — the ban became absolute 2026-04-19).
  3. Final Voice Pass: Does the skill’s workflow end with the logged voice-only pass from WRITING-PROCESS v1.1? If not, add it.
  4. Per-platform trajectory: Is there ONE canonical place stating each platform’s format and tone, and does the skill point at it? Known rules that must be findable (verify current wording against memory/guides rather than trusting this list): Instagram ~150 words, “Link in bio,” hashtags inside the post content, always #baseworks_method and never #baseworks; Facebook collaborations get two separate posts (Patrick personal first, Baseworks page third) with the URL included; Facebook and LinkedIn links carry no UTMs; LinkedIn uses a personal-account article; positioning is niche-building, not mass-appeal funnel; newsletters have no Patrick/Asia sign-off; Brain Fodder sections use varied closing-prompt formats and no sensational language; group/participant posts use a neutral register with no praise adjectives; never link practice.baseworks.com publicly (use baseworks.com/practice/). If any of these live only in Claude’s memory rather than in a guideline doc a fresh session would load, codify them — that is exactly the drift mechanism the audit identified.
  5. Repetition scan: Campaign work produces sibling assets (newsletter + 3 social posts + blog). Check that the skills require a cross-asset repetition scan: recurring facts (e.g., the development-timeframe line) must vary in framing across assets, per the unified guide v2.8.

Process requirements:

  • Work on main only; pull first. Read the vault CLAUDE.md before starting.
  • Produce the audit as 03-resources/voice-guides/SKILLS-VOICE-AUDIT-YYYY-MM-DD.md first, findings before fixes; show me the findings, then apply the fixes I approve.
  • When updating skills, follow 03-resources/claude-code-skills/index.md protocol and update BOTH the table row and the detailed section in that index for every skill touched.
  • Add version-history entries where the docs have them; anything touching VOICE-GUIDE-UNIFIED is a draft pending Ksenia and needs an Asia inbox item.
  • Commit, push, and add a changelog entry in ~/Documents/baseworks-changelog/CHANGELOG.md when done.
  • Scope: this session is skills-and-guidelines only. Don’t rewrite any live campaign content or platform strategy; if you find published content violating current rules, list it in the audit for a separate decision.
Section titled “Related”