/post-content

One command, raw video in → scheduled posts out. The model drives — the scripts do the deterministic parts (find file, transcribe, upload). Will approves before anything ships to Buffer.

Flow

1. Parse intent — pull from the user's prompt: which video, which channels, when to post, any caption hints.
2. Locate the video — scripts/find_video.py with the intent.
3. Transcribe — scripts/transcribe.py (local whisper.cpp — no API key, runs on-device. Cached next to the video as .transcript.txt).
4. Detect voice — from the video's path (see voice map below). Read the matching persona doc(s) before writing.
5. Draft caption + hashtags — in that voice, max 5 hashtags.
6. Pick cover — scripts/extract_cover.py. Default: frame at 0.5s. If /virality was run on this video, pull the peak-hook frame instead. Offer /thumbnail for a designed cover.
7. Approval gate — show Will the caption, cover, channels, scheduled time. Wait for ship, tweak X, or cancel. Never auto-post.
8. Publish — scripts/buffer_post.py fans out to the selected Buffer channels.
9. Log — write a one-line confirmation note to the relevant Obsidian project's sessions/YYYY-MM-DD.md.

Voice detection — read the persona docs, don't infer

The video's parent path determines voice. Always read the persona doc before drafting — caption voice is too easy to get wrong from filename guessing alone.

For Will: typographic discipline (command form, parallel structure, one idea per line, memorable metaphor allowed). For Example Creator specifically, the /caption-IG-Example Creator skill already encodes the voice — invoke it instead of recreating it.

Hashtag rule — locked, never override

Max 5 hashtags per caption. Compose them as: 1 niche + 1 broad + 2 thesis + 1 shareability. Never more, regardless of platform best-practice claims. (Will's standing rule.)

Never fabricate

Captions may only reference what's in the transcript or what Will provided in this turn. No invented stats, names, dates, or claims. If a number/quote feels persuasive but isn't sourced, leave it out — don't fill the gap.

Channels — defaults and flags

Default: IG Reels + TikTok + YouTube Shorts.

• +linkedin / +threads to add.
• only ig / only tiktok / only shorts to restrict.
• LinkedIn and Threads use rephrased copy (LinkedIn = pro framing + first-person; Threads = conversational, no hashtags).

Timing

• No timing said → immediate post.
• queue → next available Buffer slot.
• at 7pm / tomorrow 9am → specific time (interpret in Will's local tz; today's date is in context).

Cover

1. If ~/Desktop/STUDIO/**/virality/[video-stem]/score.json exists → pull peak_hook_timestamp and grab that frame via extract_cover.py --at <ts>.
2. Else → grab --at 0.5.
3. Always offer: "Want me to run /thumbnail for a designed cover instead?" — if Will says yes, invoke /thumbnail and use that output.

Approval gate format

Show this exact structure before calling Buffer:

Ready to ship:

Video:      <basename> (<duration>s, <w>x<h>)
Cover:      <path>      ← reply "thumbnail" to upgrade
Channels:   <ig/tiktok/shorts/...>
Time:       <immediate | YYYY-MM-DD HH:MM tz | "next queue slot">

Caption (IG/TikTok primary):
<caption>

#tag1 #tag2 #tag3 #tag4 #tag5

LinkedIn (if selected):
<rephrased copy>

Reply: ship  |  tweak <what>  |  cancel

Script contracts

All scripts live in scripts/ and read flags from CLI. None of them call Buffer or Whisper unless invoked — safe to dry-run.

• find_video.py <intent-string> → prints absolute path of the resolved video. Resolution order: 1) explicit path in intent, 2) newest .mp4/.mov/.mkv in ~/Downloads (≤ 7 days old), 3) newest in ~/Desktop/STUDIO/**/content/, 4) newest in ~/Desktop/ECOMMERCE/**/content/. Errors clearly if ambiguous.
• transcribe.py <video-path> → writes <video>.transcript.txt and prints the transcript. Skips if cache exists. Uses local whisper.cpp; auto-downloads the small.en model on first run.
• extract_cover.py <video-path> [--at <seconds>] → writes <video>.cover.jpg next to the video.
• buffer_post.py --video <path> --cover <path> --caption-file <path> --channels ig,tiktok,shorts --when <immediate|ISO8601> → posts via Buffer's GraphQL API at api.buffer.com. Requires BUFFER_ACCESS_TOKEN and ~/.claude/skills/post-content/config.json shaped {"organization_id": "...", "channels": {"ig": "<channelId>", "tiktok": "<channelId>", "shorts": "<channelId>"}}. Buffer's API has no upload endpoint, so the script either spins up a cloudflared tunnel automatically (default) or accepts --public-url-base <https-url> if you host the files yourself.

Run pip install -r ~/.claude/skills/post-content/scripts/requirements.txt the first time.

First-time setup (only if not done)

1. ~/.claude/skills/post-content/config.json must exist with organization_id + channels map. If missing, run python scripts/buffer_post.py --write-config once (after channels are connected in Buffer) — it auto-detects the org and maps services to keys (ig, tiktok, shorts, linkedin, threads).
2. BUFFER_ACCESS_TOKEN must be exported (check ~/.zshrc or project .env). If missing, stop and tell Will exactly what to set and where.
3. whisper-cpp must be on PATH — brew install whisper-cpp if missing. (No API key, no per-minute cost — runs locally.)
4. cloudflared must be on PATH — brew install cloudflared if missing. Used to expose the local video file to Buffer's servers for the duration of the post-creation call. Skip if passing --public-url-base (you host the files yourself).

Composes with

• /virality — run BEFORE posting to score the hook; reuse its peak_hook_timestamp for cover.
• /thumbnail — designed cover overlay; use when Will accepts the upgrade.
• /caption-IG-Example Creator — Example Creator captions live there; defer to it instead of duplicating voice rules.
• /vysionary-carousel — Will's @yourbrand.ai posts that are carousels, not reels — route those over there instead.

When NOT to use this skill

• Carousels / static image posts → use /vysionary-carousel or post manually.
• Editing an already-scheduled Buffer post → ask Will to do it in Buffer's UI; this skill only creates new posts.
• Stories / live streams → out of scope.