Stylistic ContentVoice writer (multi-channel + fact-check)

A Claude Code skill writes posts in distinct voices per channel — a measured corporate channel, an "AI-native" builder channel, a deliberately raw "gonzo" channel with profanity, etc. — synthesizing each voice from a corpus of the founder's own past posts. Composable parameters select channel, optional folk/raw twist, and matching cover banner. Web-search MCP fact-checks claims before publishing. The skill knows which folder to save the draft to based on the chosen channel.

Founder content is the single biggest growth lever for early-stage B2B but the bottleneck is the founder's typing time. Synthesizing the founder's voice once and dispatching across 3–5 channels with channel-specific tone cuts raw writing time roughly in half while keeping each channel's audience contract intact. Fact-checking inline kills the "AI-written post embarrassed us with a wrong stat" failure mode.

1. Voice synthesis from the founder's corpus. Drop ~10 of the founder's strongest past posts into a folder (e.g., voice-samples/{channel-name}/). Ask the agent: "Read these 10 posts. Summarize the voice along these axes — register, sentence length distribution, vocabulary tics, opening patterns, closing patterns, emoji policy, taboo phrases, signature moves. Output a content-voice-{channel}.md file I can review." Have the founder edit it; the file becomes the canonical style rule.
2. Per-channel rules. Maintain content-voice-{channel-A}.md, content-voice-{channel-B}.md, etc., as independent files. Each file declares: audience contract, voice axes, length envelope, taboo list, signature opener / closer patterns, allowed media types, save-to folder. Optional folk:true substyle adds a raw / gonzo twist.
3. Skill scaffold. Create a Claude Code skill content-voice/SKILL.md with a single command that takes parameters: --channel=X, --folk=true|false, --topic="...", --length=short|long, and optionally --banner=true. The skill reads the matching rule file, drafts the post, runs the fact-check sub-agent, attaches a matching cover via the banner skill (case #26 — auto-banner), and writes the draft to the channel-specific folder.
4. Fact-check sub-agent. Before returning the draft, the skill extracts every numerical claim, named entity, and quoted phrase. For each, it spawns a web-search sub-agent: "Verify this exact claim with a primary source. Return 'verified + URL', 'contested', or 'unverifiable'." Any non-verified claim is replaced with [CLAIM NEEDS VERIFICATION] in the draft.
5. Founder review. Single human-review pass before publish. The draft contains the fact-check log appended so the founder sees exactly which claims passed and which were softened.
6. Publish. Optional: wire the skill to a publishing MCP (Telegram MCP for channel broadcasts; Twitter/LinkedIn MCP for social) so "publish to channel X" is a one-line follow-up.

Voice-synthesis prompt to bootstrap a new channel rule:

Read every file under voice-samples/{channel}/.
Produce a content-voice-{channel}.md file with these sections, filled from the corpus:
 - Audience contract (who reads this and what do they want)
 - Register (formal / informal / raw — with example phrases from the corpus)
 - Sentence length distribution (median, p90)
 - Opening patterns (top 5 ways posts begin)
 - Closing patterns (top 5 ways posts end)
 - Vocabulary tics (recurring words/phrases)
 - Taboo list (words/phrases never to use in this voice)
 - Emoji policy
 - Signature moves (3-5 distinctive structural patterns)
 - Length envelope (min, max words)
 - Save-to folder
For each axis cite 2-3 verbatim examples from the corpus.

Fact-check loop, verbatim instruction inside the skill:

For every numerical claim, named entity, and quoted phrase in the draft,
launch a web-search sub-agent with this instruction:
"Verify the exact claim against a primary source.
Return one of: VERIFIED + source URL, CONTESTED + alternative, UNVERIFIABLE.
Do not paraphrase the claim — verify it verbatim."
Replace any non-VERIFIED claim in the draft with [CLAIM NEEDS VERIFICATION: <original>].
Append the verification log at the end of the draft.

Skill invocation examples:

/content-voice --channel=mindset --topic="why we killed Confluence" --length=long
/content-voice --channel=ainative --folk=true --topic="claude code first hour" --banner=true

• Don't generate the voice rule by hand — synthesize from the corpus or it drifts toward generic AI-blog cadence. The corpus is the ground truth.
• One voice per channel; don't compose channels at runtime. Cross-pollinating voices destroys the audience contract.
• Fact-check MUST be a separate sub-agent call per claim, not a single "fact-check this whole post" pass — the latter hallucinates approval at the post level.
• Save the verification log inline in the draft; never strip it before the founder review.

• Lighter: skip the banner step and the publishing wire-up; use the skill only for drafting, paste into the channel manually.
• Heavier: add a per-channel performance feedback loop — pull engagement data weekly, identify the highest-performing openings/closings, regenerate the voice rule from the top-30 posts every quarter.
• Multi-author: maintain content-voice-{channel}-{author}.md so co-founders can each publish in the same channel without collapsing into one voice.