Hammadbaig-hub/daily-shorts-media

YouTube Daily Shorts

One Short per day, fully automatic, zero cost. The channel posts "Did you know?" style facts and short motivational pieces, one day in English and the next in Urdu.

Fixed settings (change here if the user changes them)

Setting	Value
GitHub media repo	Hammadbaig-hub/daily-shorts-media (public, branch main) — also holds the render tools in tools/
Metricool brand (blogId)	7592220, timezone Asia/Karachi
Channel	"Bright Life", handle @Bright-Life-m4 (put the handle in script.json; CTA mentions "Bright Life")
Publish time	19:00 Pakistan time (or the next free slot if that has passed; at least 15 min from now)
Privacy	public (user chose full auto-upload)

Why it is built this way

• Free: the user has no budget, so the video is drawn with Pillow + ffmpeg (tools/scripts/make_short.py in the repo) and the music is synthesized in the same script, which also means no Content ID copyright claims.
• Voiceover: narration comes from Kokoro, a small free neural TTS model that runs offline on CPU (tools/scripts/voice.py). Its ~350 MB model files download once from GitHub releases into ~/tts. English uses am_michael; Urdu is spoken by the Hindi voice hm_omega from Urdu-script text, with a few phonemes remapped. Each slide lasts as long as its narration, and the music is ducked underneath. Cloud TTS (Microsoft/Google) is blocked from the workspace, which is why it is offline.
• GitHub hosting: Metricool needs a public video URL. The cloud workspace can push to GitHub, and the raw file URL is public, so that is the bridge.
• Fact-checking: a facts channel lives or dies on accuracy. YouTube also suppresses mass-produced, low-value or misleading content, and a wrong "fact" invites dislikes and correction comments. Every fact gets verified before it is used — this is the part of the job that adds real value.

Daily workflow

1. Get the repo and plan the day

cd ~ && (git -C daily-shorts-media pull -q || git clone https://github.com/Hammadbaig-hub/daily-shorts-media.git)
pip install -q --break-system-packages kokoro-onnx soundfile pillow numpy
python daily-shorts-media/tools/scripts/plan_today.py daily-shorts-media/history.json

If the repo is not in the session, attach it first with add_repo (owner Hammadbaig-hub, repo daily-shorts-media, access push) and clone it.

plan_today.py returns today's lang, theme, whether a video was already posted today, and every topic used so far. If posted_today is true, stop and report that — never post twice in one day.

2. Pick a topic

Rotate across categories so the channel stays varied: space, animals, human body, history, geography, science, technology, food, ocean, inventions, language, and roughly one day in five a motivation piece (a lesson or habit backed by a real study, or a well-documented true story). Pick something not in already_done and not a near-duplicate of it.

Stay away from: medical or diet advice, politics, sectarian or religious debate, anything about real private people, and "facts" that are commonly repeated myths (e.g. goldfish 3-second memory, the Great Wall being visible from space).

3. Write the script and verify it

Write a hook (under 12 words, curiosity-driven), 3–5 facts (each one sentence, under 25 words), and a CTA.

For every fact, run a WebSearch and confirm it from a reliable source (NASA, National Geographic, Britannica, a university, a peer-reviewed study, a major encyclopedia). If a fact can't be confirmed, or sources disagree, drop it and use another. Keep the source URLs — they go in the description.

Urdu days: write natural, simple Urdu in Urdu script (not Roman Urdu), the way a Pakistani viewer speaks — not a word-for-word translation. Write numbers as Urdu words (بیس فیصد, ہزاروں) so the voice pronounces them correctly.

CTA examples — EN: "Follow Bright Life for a new fact every day". UR: "روزانہ نئے حقائق کے لیے برائٹ لائف کو فالو کریں".

Save as script.json:

{"lang": "en", "hook": "...", "facts": ["...", "..."], "cta": "...",
 "handle": "@Bright-Life-m4", "theme": 2}

4. Render

python daily-shorts-media/tools/scripts/make_short.py script.json daily-shorts-media/videos/<date>-<slug>.mp4 > render.log 2>&1

Rendering takes about 5 minutes on CPU, so run it with Bash run_in_background and wait for the completion notice. Don't wait with a pgrep -f make_short loop — it matches its own command line and never ends. Length follows the narration — usually 25–45 s for 3–5 facts. Keep it under 55 s so it stays a Short: if the facts are long, use 3 instead of 5.

Because the text is spoken, write it to be heard: spell out awkward symbols ("percent", not "%" in English; "فیصد" in Urdu), avoid brackets and abbreviations, and keep sentences short so the voice doesn't rush. Then grab one frame and look at it with Read to catch text overflow or broken Urdu:

ffmpeg -loglevel error -y -ss 5.5 -i <video> -frames:v 1 check.png

If text runs off the edge or looks wrong, shorten that line and re-render.

5. Push to GitHub

Append today's entry to history.json: {"date": "...", "lang": "...", "theme": n, "topic": "...", "title": "...", "file": "videos/...mp4"}

Keep only the 20 newest .mp4 files in videos/ (delete older ones — they are already on YouTube) so the repo stays small. Then:

cd daily-shorts-media && git add -A && git commit -qm "Short <date>: <topic>" && git push -q

Public URL: https://raw.githubusercontent.com/Hammadbaig-hub/daily-shorts-media/main/videos/<file>.mp4

Check it returns 200 before scheduling (curl -sI <url> | head -1); if the workspace can't reach that host, trust the successful push and continue.

6. Schedule on YouTube through Metricool

Call createScheduledPost with blogId 7592220, a top-level date (YYYY-MM-DDT19:00:00+05:00 — the call fails without it), and this info JSON (Metricool copies the video into its own storage, so the GitHub file can be pruned later):

{
  "providers": [{"network": "youtube"}],
  "text": "<description>",
  "media": ["<raw GitHub URL>"],
  "publicationDate": {"dateTime": "YYYY-MM-DDT19:00:00", "timezone": "Asia/Karachi"},
  "autoPublish": true,
  "draft": false,
  "descendants": [], "firstCommentText": "", "hasNotReadNotes": false,
  "mediaAltText": [], "shortener": false, "smartLinkData": {"ids": []},
  "youtubeData": {
    "title": "<title>",
    "type": "short",
    "privacy": "public",
    "tags": ["facts", "did you know", "..."],
    "category": "EDUCATION",
    "madeForKids": false,
    "isAiGeneratedContent": false
  }
}

isAiGeneratedContent is false because the video is plain animated text, not realistic synthetic footage; YouTube's disclosure is meant for realistic altered media. If a later version adds AI-generated realistic visuals or voices, set it to true.

Title: under 70 characters, curiosity hook + #shorts. Write it in the day's language. Example: Your Brain Can't Feel Pain?! 🧠 #shorts

Description: 2–3 lines expanding the hook, then Sources: with the URLs you verified, then 3–5 hashtags (#facts #didyouknow #shorts plus topic tags; Urdu days add #urdu #urdufacts).

7. Report

Tell the user in Roman Urdu, in a few lines: topic, language, title, scheduled time, and the Metricool plannerUrl. If running as a scheduled task with nobody watching, send that summary with SendUserMessage.

Weekly check (Sundays, or when asked "channel kaisa chal raha hai")

Use getAnalyticsDataByMetrics for the YouTube network over the last 7 days (views, watch time, subscribers gained, top videos). Summarize what worked — which categories and which language got more views — and use that to lean next week's topic choices toward what performs. Keep the alternating language unless the user says otherwise.

When something fails

• Metricool rejects the post: report the exact error; don't retry the same input. Common causes: missing top-level date, video URL not reachable yet (wait a minute, retry once), or the YouTube connection in Metricool expired (user must reconnect it in Metricool).
• Git push fails: the repo isn't attached with push access — call add_repo with access: push. If it still says Claude has no GitHub access, the Claude GitHub App must have daily-shorts-media in its selected repositories (github.com/settings/installations → Claude → Configure); the user confirms that step themselves because GitHub asks for 2FA.
• Fonts missing: they live in tools/assets/fonts/ in the repo; the script reads them from there.
• Voice model download fails: render with "voice": false in script.json so the day's video still goes out (music only), and mention it in the report.