Generate Voice Guide

Generate Voice Guide

The guide itself is yours to write (the agent is the model). This skill provides the instructions plus a deterministic analyze_corpus.py that turns a post corpus into the mechanical evidence you reason over.

When to use

• "Build a voice guide for [my / someone's] LinkedIn or X."
• Whenever a sibling content skill needs a voice guide and none exists.
• NOT for corporate blog/landing-page voice (separate brand-voice skill) — this is individual.

How to run

1. Get the corpus. Capture the target's recent original posts (exclude replies, retweets, quotes, reshares) per platform with text + engagement + timestamps. Acquisition is agent-orchestrated:
   • X / LinkedIn-at-scale → Apify user-posts actor (APIFY_API_TOKEN): curl "https://api.apify.com/v2/acts/<actor>/run-sync-get-dataset-items?token=$APIFY_API_TOKEN" -d @input.json
   • LinkedIn → Phantombuster (PHANTOMBUSTER_API_KEY + LinkedIn session cookie), or web-automation + Robomotion Proxy for one-off public profiles.
   • Fallback: ask the user to paste 15–25 posts as plain text — fully keyless.
2. Analyze the corpus (deterministic stats + banned-phrase candidates):

python3 ${SKILL_DIR}/scripts/analyze_corpus.py --input ${WORKSPACE}/posts.json --platform x \
  --output ${WORKSPACE}/x-stats.json

Accepts a JSON array ([{"text","is_reply","is_retweet","engagement","created_at"}]) or a plain-text file (posts separated by a blank line / ---). Emits {post_count, length_chars, length_words, usage_rates, top_openers, banned_phrase_candidates}. Banned-phrase candidates are cliches absent from the corpus (the "derived from absence" rule), not a generic blocklist.

3. Author v1 of the guide from the stats + the raw posts: persona, the "meat" principle, dos/don'ts, 10–20 banned phrases (confirm the candidates), 5–8 hook patterns, format rules, tone calibration, 4–6 annotated real examples (quoted verbatim).
4. Sample-and-feedback loop (≥5 rounds): draft 3 samples per round on the user's own topics, show which guide lines drove them, collect feedback via the agent's channel, revise. 5 rounds is a floor — push for more.
5. Finalize & register: strip the iteration changelog, save the clean voice-x.md / voice-linkedin.md to Memory/workspace, and register the path(s) so create-x-content / create-linkedin-content / social-kit auto-discover them.

python3 ${SKILL_DIR}/scripts/analyze_corpus.py --help lists all flags.

Outputs

voice-x.md and/or voice-linkedin.md (persona, "meat" principle, dos, don'ts, banned phrases, hook patterns, CTA guidelines, format rules, tone calibration, annotated example posts), saved to the durable store with the guide paths registered.

Credentials / env

• Required: none. The agent authors the guide; the analyzer is keyless; the paste-text corpus path needs no key.
• Optional:
  • APIFY_API_TOKEN (paid, with a fallback) — If set → X (and LinkedIn-at-scale) corpus scraping via the user-posts actor (larger, automatic corpus). If not set → ask the user to paste 15–25 posts as plain text (the keyless default); the analyzer + guide work the same either way.
  • PHANTOMBUSTER_API_KEY (+ a LinkedIn session cookie) (paid, with a fallback) — If set → LinkedIn post scraping. If not → web-automation + paste-text fallback (the default).
  • Robomotion Proxy — recommended for web-automation one-off LinkedIn reads.

Notes & edge cases

• 5 iterations is a floor, not a ceiling — guides only lock in after several rounds.
• Use the user's own topics for samples so off-key lines are easy to spot.
• Quote real posts verbatim in examples; never paraphrase.
• Flag ghost-written/assistant-authored posts before baking them into the voice.
• Exclude replies, retweets, quotes, reshares — the analyzer already filters these.