AutoShorts — Daily Viral Clip Pipeline#

Pipeline tooling lives at ~/Documents/skill-autoshorts/. Each day this skill picks ONE long video from INPUT_FOLDER, extracts every viable short-form clip (Gemini 3 Flash decides), shows them to the user for approval, and publishes the approved ones via Upload-Post.

Setup (only if not yet configured)#

1. Python environment#

bash

cd ~/Documents/skill-autoshorts && python3 -m venv venv && source venv/bin/activate && pip install -r requirements.txt

2. FFmpeg#

Required system binary. Verify with ffmpeg -version. Install with brew install ffmpeg if missing.

3. .env#

File lives at ~/Documents/skill-autoshorts/.env. Required keys:

text

UPLOAD_POST_API_KEY=...
UPLOAD_POST_PROFILE=...
GEMINI_API_KEY=...
INPUT_FOLDER=/abs/path/to/long/videos
OUTPUT_FOLDER=/abs/path/to/clip/output
WHISPER_MODEL=medium
TIMEZONE=Europe/Madrid

If a required key is missing, ask the user for it before continuing.

4. Upload-Post account#

• Sign up at https://upload-post.com → dashboard at https://app.upload-post.com.
• Connect TikTok, Instagram (Business/Creator account linked to a Facebook Page), and YouTube via OAuth in the dashboard.
• In Manage Users, create a profile — its name is UPLOAD_POST_PROFILE (NOT the social handle).
• Generate an API key in Settings.
• Verify: curl -H "Authorization: Apikey $UPLOAD_POST_API_KEY" https://api.upload-post.com/api/uploadposts/me.

Orchestration model#

This skill is invoked daily by the openclaw harness, which also handles the messaging bridge (Telegram, WhatsApp, or whatever channel openclaw is configured with). The skill itself does NOT talk to Telegram or any messenger directly — it just runs the pipeline and presents the candidates as text + absolute file paths. openclaw forwards your output to the user's phone, captures the user's reply, and feeds it back into the conversation.

Concretely: at Step 5 you print the candidates table and ask which IDs to publish; openclaw delivers that table plus the clip files via the user's chosen channel; the user replies on their phone (e.g., "1, 3, 5"); openclaw injects that reply back; you continue with Steps 6–8. Same pattern for any other "ask the user" point in the workflow (metadata review, dry-run confirmation, etc.).

If the skill is invoked outside openclaw (e.g., user runs /autoshorts directly in Claude Code), the same prompts work — they just appear in the terminal instead of on the phone.

Daily workflow#

This skill is meant to run as a daily infinite loop. Every run picks ONE video and walks it through the pipeline. Pick semantics are round-robin per cycle: each video is picked at most once per cycle. When every video in INPUT_FOLDER has been processed in the current cycle, a new cycle automatically starts and the same videos become available again — generating fresh clips from already-clipped sources. The state file at state/processed.json tracks cycle_started_at, last_processed_at per video, and cycles_count per video. Inside a cycle, the next pick is the newest unprocessed-this-cycle (mtime DESC), so fresh material always jumps the queue.

Step 0 — Preflight (run on every invocation, do not skip)#

Before doing any work, check that the environment is ready and ask the user for whatever is missing:

1. venv — does ~/Documents/skill-autoshorts/venv/bin/python exist? If not, run setup step 1 from the Setup section. (You can do this without asking — it's mechanical.)

2. ffmpeg in PATH — if missing, ask the user to brew install ffmpeg (do not install yourself; system-wide installs deserve confirmation).

3. .env file — check that every required key is set and non-empty:

   • GEMINI_API_KEY → if missing, ask: "Falta la API key de Gemini. Pégamela (la generas en https://aistudio.google.com/apikey)."
   • UPLOAD_POST_API_KEY and UPLOAD_POST_PROFILE → if missing, ask: "Necesito la API key de Upload-Post y el nombre del profile (Manage Users en https://app.upload-post.com)."
   • INPUT_FOLDER and OUTPUT_FOLDER → if missing, default to ~/Documents/skill-autoshorts/input and ~/Documents/skill-autoshorts/output and write them to .env.
   • WHISPER_MODEL → default medium. TIMEZONE → default Europe/Madrid.

If the user provides an API key in the conversation, write it to .env immediately, never echo it back, and warn that the key is now in conversation logs and they should rotate it after testing.

Input video format: videos in INPUT_FOLDER are expected to already be 9:16 vertical and ready to post (1080×1920 typical). If the user has burned-in subtitles, those should already be on the source video. The skill does NOT reformat, crop, scale, or burn subtitles — it only cuts the chosen segment and overlays a hook text on top. If a video arrives in landscape or any non-9:16 ratio, surface that to the user and ask before processing — running the pipeline as-is will produce TikTok-incompatible output.

How videos arrive into INPUT_FOLDER is the harness's job, not the skill's. The canonical flow: the user forwards a video to openclaw / Hermes / their agent in chat (Telegram / WhatsApp / etc.), the harness downloads it and saves it to INPUT_FOLDER. The skill itself only operates on files that are already there. If the user passes a video path that is NOT inside INPUT_FOLDER (e.g. /autoshorts /Users/foo/Downloads/podcast.mp4), copy it in first (use cp, do not move — the original stays put). Otherwise pick will not find it.

Step 1 — Pick the video#

bash

python autoshorts.py pick

Returns JSON with the next video to process. Output fields:

• path, name, size_mb, mtime, duration_s — file metadata.
• previous_cycles_completed — how many cycles this video has already been through (0 means first time ever).
• remaining_in_cycle — how many other videos are still untouched in the current cycle.
• cycle_started_at — timestamp of the current cycle's start.
• new_cycle_started — true if THIS pick is the one that opened a fresh cycle (every video has been processed in the previous cycle, the loop wraps around).

If new_cycle_started is true, mention it to the user briefly ("starting a new cycle — already clipped this video N times before, going for fresh moments"). It's not an error — it's the expected wrap-around. Gemini will likely pick different moments since the prompt is non-deterministic and HOT.md priors evolve over time.

The pipeline does NOT hard-stop when it runs out of fresh videos. The only hard-stop case is INPUT_FOLDER being empty — surface that and ask the user to drop something in.

If the user explicitly says "reprocess video X right now" out of cycle order, remove that entry from state/processed.json first, then run pick. Do NOT bypass the cycle logic by other means.

Step 2 — Transcribe#

bash

python autoshorts.py transcribe "<VIDEO_PATH>"

Writes output/<video_slug>/transcript.json with sentence segments and per-word timestamps. Whisper auto-detects language. Default model is medium.

Step 3 — Analyze with Gemini 3 Flash#

bash

python autoshorts.py analyze "<VIDEO_PATH>"

Uploads the video to Gemini Files API and asks gemini-3-flash-preview to return EVERY viable short-form moment (20–60s each), with timestamps snapped to word boundaries from the transcript. Output: output/<video_slug>/clips.json. Read this file to get the candidate list.

Step 4 — Cut every candidate and add hook overlay#

For each clip in clips.json, run two commands:

bash

python autoshorts.py extract "<VIDEO_PATH>" \
    --start <START> --end <END> \
    --output "output/<slug>/clip_<ID>.mp4"

python autoshorts.py hook "output/<slug>/clip_<ID>.mp4" \
    --text "<HOOK_TEXT>" --duration 3 \
    --output "output/<slug>/clip_<ID>_final.mp4"

The hook is rendered TikTok/Instagram-style: each line of text gets its own black pill (78% opacity, rounded corners) behind it, with white Impact text + black stroke on top. The pill keeps the hook legible on any background — pure white, pure black, busy screenshares — without needing to inspect the underlying frame. Positioned at the top of the frame for the first 3 seconds. Hook text comes from clips.json (Gemini wrote it in the video's language).

Cut and hook ALL candidates upfront — the user will review the actual final files visually, not metadata.

Step 4.5 — Visual QA of the hook (you do this yourself, no Gemini call)#

You are multimodal. Use that. Before showing the candidates to the user, verify the hook overlay actually renders cleanly on each clip.

For every clip_<ID>_final.mp4:

bash

python autoshorts.py preview output/<slug>/clip_<ID>_final.mp4

This extracts a single frame at t=1.0s (mid-hook) to preview_clip_<ID>_final.png next to the clip. Open it with the Read tool — Claude / openclaw both view PNGs directly. No Gemini call needed; the agent running the skill IS the multimodal reviewer.

For each preview, evaluate:

1. Is the hook text fully visible? Any letter clipped at the left/right/top edges?
2. Is the pill background extending past the safe area (more than ~5% from any edge)?
3. Does the hook cover the speaker's face or other critical content?
4. Are accent marks / special characters ( á é í ó ú ñ ¿ ¡) rendering correctly?
5. Is the hook overlapping with the burned-in subtitle? (Subtitles at the bottom are expected — only flag if they collide with the hook itself, which lives at the top.)
6. Any rendering glitch: garbled text, missing pill, transparency issue?

Add a "QA" column to the Step 5 table with one of:

• ✅ — clean
• ⚠️ <issue> — flag the specific problem (e.g. ⚠️ último carácter recortado, ⚠️ pill desbordado a la derecha)

Do NOT silently drop flagged clips — show them to the user with the warning so they can decide. The QA pass is advisory: a "⚠️" is a hint, not a veto. If multiple clips fail in the same way (e.g. the hook is consistently overflowing), that's a signal to suggest the user shorten the hook style going forward.

Step 5 — Present to the user#

Show a markdown table:

IDDurationHookScoreQAReasonFile138s"..."9✅...output//clip_1_final.mp4227s"..."7⚠️ acento "ó" recortado...output//clip_2_final.mp4…………………

Always include the absolute file paths in the table — openclaw uses them to attach the actual clip videos when it forwards the message to the user's messenger (Telegram / WhatsApp / etc.). Without absolute paths the user sees only metadata and cannot review the clips visually. Then ask:

Which clip IDs do you want to publish? (e.g. 1, 3, 5, or none.)

Wait for the user's reply (it will arrive via openclaw from the user's phone).

If the user replies none (rejects all candidates), skip directly to Step 8 and mark-processed with --clips-published 0. This consumes the video so tomorrow's run picks the next one — otherwise the same rejected candidates would surface again. If the user wants to retry the same video later, they can manually remove its entry from state/processed.json.

Step 6 — Generate platform metadata for approved clips#

For every approved ID, generate platform-specific copy. This is YOUR job as Claude — write it directly, do not call a tool. Match the language of the video.

• TikTok ( tiktok_title, max 90 chars): punchy hook, 1–2 emojis, hashtag mix at end of the title. Sweet spot ~70–85 chars.
• Instagram Reels ( instagram_title, up to 2200 chars): long-form storytelling — first line is the hook, then 2-4 short paragraphs (use \n\n), CTA ("Guarda esto", "Etiqueta a alguien…", "Comenta X para…"), then 20-30 hashtags mixing sizes (large/medium/niche). Sweet spot 500–800 chars total.
• YouTube Shorts ( youtube_title, max 100 chars but keep ~40-60 chars so it doesn't truncate on mobile): SEO-friendly with keywords. Description focuses on searchability, 3–5 hashtags max.
• A general title and description for any platform that doesn't have its own override.

Length contract (verify before publishing): YouTube title is the most constrained — write it shortest and most direct. TikTok and Instagram can breathe — TikTok up to ~85 chars in tiktok_title, Instagram captions are long-form by design.

Show the generated copy back to the user and confirm before publishing.

Step 7 — Schedule publishing#

Schedule one approved clip per day starting tomorrow at 10:00 in TIMEZONE (default Europe/Madrid). Each next clip += 1 day.

For each approved clip:

bash

python autoshorts.py publish "output/<slug>/clip_<ID>_final.mp4" \
    --platforms tiktok,instagram,youtube \
    --title "<GENERAL>" \
    --description "<DESCRIPTION>" \
    --tiktok-title "<TIKTOK_TITLE>" \
    --instagram-title "<INSTAGRAM_CAPTION>" \
    --youtube-title "<YOUTUBE_TITLE>" \
    --schedule "<ISO_DATE>" \
    --timezone "Europe/Madrid" \
    --tiktok-mode draft \
    --clip-id <ID> \
    --hook-text "<HOOK_TEXT>" \
    --viral-score <GEMINI_SCORE> \
    --reason "<GEMINI_REASON>" \
    --video-source "<SOURCE_VIDEO_FILENAME>"

The --clip-id, --hook-text, --viral-score, --reason, --video-source flags are not optional in practice — they feed the learning loop. Without them, learn cannot correlate engagement metrics back to which hook patterns worked. The values come straight from clips.json (the Gemini output) and the source video filename.

TikTok mode: --tiktok-mode draft (default) sends to the TikTok inbox via post_mode=MEDIA_UPLOAD so the user can finish editing in-app before publishing. Use --tiktok-mode direct ( DIRECT_POST) only when the user explicitly wants immediate publishing.

Always run with --dry-run first and show the user the exact request payloads. Only execute the real publish after explicit "go".

Step 8 — Mark video as processed#

bash

python autoshorts.py mark-processed "<VIDEO_PATH>" \
    --clips-generated <N_CANDIDATES> \
    --clips-published <N_APPROVED>

This appends the video's hash to state/processed.json so tomorrow's pick skips it. Run this even if --clips-published 0 — a rejected video is still consumed. The only time you do NOT mark-processed is if the pipeline crashed mid-run (e.g., Gemini errored out before producing clips); in that case let the user retry the same video tomorrow.

Step 8.5 — Reflect (optional, fast, qualitative)#

After publishing, you can run a quick reflect to capture WHY the user approved the clips they approved (no engagement metrics needed — just the approved-vs-rejected signal):

bash

python autoshorts.py reflect --window-days 30

This compares recent candidates ( learnings/candidate-history.jsonl) against approvals ( learnings/post-history.jsonl) and asks Gemini to extract qualitative patterns ("approves hooks with concrete numbers, rejects question-form hooks"). Output goes to learnings/runs/reflect-YYYY-MM-DD-HHMM.md.

These observations are NOT auto-promoted to HOT.md. They're notes for the user to review and curate. Run reflect occasionally — daily is overkill, weekly is fine.

Step 9 — Final summary#

Print:

#FileDurationHookSchedulePlatforms

…and the source video name with how many candidates were generated vs. published.

Weekly learning loop ( learn)#

This skill gets smarter over time. Engagement data from past clips (views, likes, comments, shares, saves — fetched from Upload-Post analytics) is fed back into the clip-selection prompt for future runs.

Cadence#

Run learn weekly, not daily. Engagement metrics need time to mature; daily learn would chase noise.

bash

python autoshorts.py learn

Defaults: 7-day soak (clips younger than this are excluded), 90-day max age (older are stale), composite score = 0.6·views + 0.4·engagement_rate, top/bottom 20% as winners/losers.

What it does#

1. Reads learnings/post-history.jsonl (every clip we published, with its hook + Gemini score + Gemini reason + source video).
2. For each clip in the soak window, calls GET /api/uploadposts/post-analytics/{request_id} — same request_id we got back at publish time.
3. Computes a composite score per clip and picks the top 20% (winners) and bottom 20% (losers).
4. Sends winners + losers + the existing learnings/HOT.md to Gemini Flash with a meta-prompt asking it to produce an updated HOT.md (≤80 lines) listing patterns supported by the new evidence.
5. Writes the new HOT.md (backing up the previous one as HOT.YYYYMMDD-HHMMSS.md.bak).
6. Writes a full audit to learnings/runs/learn-YYYY-MM-DD.md so the user can see exactly which clips were called winners/losers and how the learnings changed.

How HOT.md feeds back#

cmd_analyze automatically reads learnings/HOT.md (if it exists and is non-empty) and prepends it to the Gemini prompt as "PRIOR LEARNINGS — apply when selecting clips and writing hooks". Gemini then weighs those patterns when proposing clips and writing hooks for tomorrow's video. You don't have to do anything to make this work — it happens on every analyze call.

When to run learn#

• Manually, on demand: python autoshorts.py learn
• Scheduled, weekly via cron / openclaw: 0 9 * * 1 cd ~/Documents/skill-autoshorts && ./venv/bin/python autoshorts.py learn
• Skip if post-history.jsonl has fewer than ~10 entries — the rule of "5 winners + 5 losers minimum" will short-circuit the run with a "not enough data" note.

Things to NOT do#

• Do not edit HOT.md by hand AND keep running learn — learn will overwrite your edits. If you want manual rules, put them in learnings/insights/ (manual notes, not used by the pipeline).
• Do not delete post-history.jsonl or metrics.jsonl — they're append-only memory. Without them every learn starts from zero.
• Do not run learn more than ~once a week — Gemini will just churn the same patterns.

Operating notes#

• Always confirm before Step 4 (heavy ffmpeg work — do NOT skip, but confirm if Gemini returned > 15 candidates — could waste time), before Step 7 (publishing is irreversible once scheduled), and after Step 6 (metadata copy).
• If Gemini returns malformed JSON, the raw response is dumped to output/<slug>/clips.raw.txt — read it and re-prompt manually.
• Hook text comes from Gemini in the video's language. Do not translate.
• The Upload-Post free tier is 10 uploads/month — one publish to 3 platforms counts as 3. Warn the user if scheduling would exceed the quota.
• All clip files are absolute paths under OUTPUT_FOLDER/<video_slug>/. Surface them clearly so the openclaw harness can attach them when forwarding to Telegram / WhatsApp / whatever messenger channel the user has configured.
• If pick says "all videos already processed", tell the user and stop — do not re-process. They need to drop a new video into INPUT_FOLDER.
• The state file at state/processed.json is the only memory between runs. Never edit it programmatically except via mark-processed. If the user asks to "reprocess video X", the right move is to ask them to confirm, then remove the matching entry from state/processed.json manually.
• The Whisper medium model (~1.5 GB) downloads on first transcribe call. Warn the user the first run will take longer — subsequent runs reuse the cached model.