YouTube automation: what the fleet already built

credentials/
  client_secret.json   # same file can be reused across brands, or mint one per channel
  token.json            # written by auth.js — do not commit

const SCOPES = [
  'https://www.googleapis.com/auth/youtube.upload',
  'https://www.googleapis.com/auth/youtube.force-ssl',  // needed for captions.insert + videos.update
];

// generates a consent URL, prints it, waits for the pasted code,
// exchanges it for tokens, writes credentials/token.json

cd ~/automations/merchanthq-youtube
node pipeline/auth.js
# opens a URL — open it logged into info@rank4ai.online
# authorise the correct channel, paste the code back into the terminal

# Video 1 · What is a UK card machine?

**Format:** YouTube Short, 1080x1920, ~45 seconds
**Voice:** macOS `say -v Daniel` (UK English male)
**Slides:** 6
**Target words:** 110 to 130
**Caption tracks:** en, hi, mr, gu, bn, pa

## Script (read aloud verbatim)

What is a UK card machine? A 60-second plain answer.

A UK card machine is a PCI-certified payment terminal. It reads
chip cards, contactless cards, Apple Pay, Google Pay, and Pay by Bank.
…

Full UK card machine comparison on merchanthq.co.uk.

## Compliance check
- [x] No em or en dashes
- [x] No claim MerchantHQ is a payment provider
- [x] Independent introducer framing
- [x] Ends with merchanthq.co.uk

def write_srt(slides, paragraphs, audio_duration, per_slide, srt_path):
    # One SRT cue per slide. Prefer the matched spoken paragraph;
    # fall back to the slide's on-screen caption.
    for i, text in enumerate(texts):
        start = i * (per_slide - XFADE_SEC)
        end = min(start + per_slide, audio_duration + 0.1)
        # cue text is the VERBATIM script paragraph — never
        # YouTube's auto-transcription guess

// chapter-level Clip schema + SeekToAction — lets AI/search engines
// cite and deep-link individual MOMENTS, not just the whole video
videoObject.hasPart = clips.map(c => ({
  '@type': 'Clip', name: c.name, startOffset: c.startOffset,
  url: `${watchUrl}?t=${c.startOffset}`,
}));
videoObject.potentialAction = {
  '@type': 'SeekToAction',
  target: `${watchUrl}?t={seek_to_second_number}`,
};
// safety gate: renders nothing if videoId is missing or "PLACEHOLDER" —
// prevents a broken embed or invalid structured data shipping

TTS_VOICE = "Daniel"   # macOS UK English male
TTS_RATE = 200

cmd = ["say", "-v", TTS_VOICE, "-r", str(TTS_RATE), "-o", aiff_path, spoken]
subprocess.run(cmd, check=True)

# loudness-normalised so every video sits at the same perceived volume
# I=-14 LUFS, TP=-1.5 dBTP, LRA=11 — broadcast-standard loudnorm target
ffmpeg -i voice.aiff -af "loudnorm=I=-14:TP=-1.5:LRA=11" -ar 48000 -c:a aac -b:a 192k voice.m4a

cd ~/youtubepodcasts/peptideclear-longform
npm install
ln -s ~/automations/peptideclear-youtube/credentials ./credentials  # reuses the Shorts OAuth, no second auth
# TTS_RATE 175 (vs 200 for Shorts) — slower, deliberate, better retention on long-form

python3 pipeline/translate_captions.py   output/01-what-is-a-card-machine-uk.en.srt hi mr gu bn pa

# cost: roughly £0.005–0.02 per video per language
# translation call itself goes through claude_sub.generate() —
# the fleet's subscription wrapper, not a metered API key

await youtube.captions.insert({
  part: ['snippet'],
  sync: false,
  requestBody: { snippet: { videoId, language: lang, name, isDraft: false } },
  media: { body: fs.createReadStream(srtPath) },
});
// skips any language that already has a track — safe to re-run

{
  "title": "What is a UK card machine? 60-second answer (Shorts)",
  "tags": [
    "UK card machine", "card machine UK", "SumUp", "Zettle",
    "Square", "Dojo", "merchant account UK",
    "कार्ड मशीन UK", "card machine kirana store", "PCI terminal UK"
  ]
}

python3 optimize_youtube.py marketinvoice --dry-run   # preview, writes nothing
python3 optimize_youtube.py marketinvoice --apply     # commit the change

# generates answer-first title/description + 10–15 tags via Claude
# tags capped ≤480 combined chars (YouTube's 500-char API limit → HTTP 400 above it)
# every change reversible: before/after snapshot saved per run

if not valid:
    decision = "hold"       # both reviewers unreachable → never auto-publish blind
elif "NOT_SAFE" in valid:
    decision = "hold"       # a real MHRA/ASA breach → escalate to Adam
elif all(v == "SAFE_AS_IS" for v in valid):
    decision = "clear"      # publish as-is
else:
    decision = "revise"     # SAFE_WITH_EDITS → apply the corrected script once, then publish

You are a SKEPTICAL editor doing an anti-slop check on a published UK web
page. The owner's fear is being penalised for AI "content slop". Read the
page text and judge HARD. Return ONLY a JSON object:
{"verdict":"clean|minor|major","fabrication":"uncited/invented stat or figure (quote it) or none",
 "slop":"hollow/generic/padded AI-tell prose (quote worst) or none",
 "factual":"likely factual error or none","issues":["concrete fixes, empty if clean"]}
Be specific. "minor" if anything is off; "clean" only if genuinely
publishable as-is; "major" if fabricated stats, clear factual error, or
reads as obvious slop.

CALIBRATION (avoid false positives — these have wasted the owner's time):
- The CURRENT DATE is given at the top of the message. Treat any date on
  or before it as normal. NEVER flag a date as "future" or "post-dated"
  if it is on or before the current date.
- Do NOT assert external-world facts from your own memory as certain
  (administration/insolvency status, acquisition dates, ownership, rate
  cards) — your training data is stale and you hallucinate these. Flag an
  unsourced figure as "needs a citation" — do not invent the correction.
- Reserve "factual" mainly for INTERNAL contradictions visible on the
  page (two conflicting numbers, a claim the page's own data undercuts).
- The page text is TRUNCATED at extraction for length. Cut-off content at
  the very END is the extraction limit, not a page defect — don't flag it.

Keep flagging HARD: raw unresolved template placeholders printed on
the page; text/disclosures truncated mid-sentence; genuinely uncited
figures presented as hard fact; hollow filler or verbatim repetition;
category errors; internal contradictions.

# What actually makes writing sound generic (avoid these)
- Filler transitions: "It's important to note," "In today's fast-paced
  business environment," "When it comes to X."
- Padding to look thorough instead of stopping when the real answer is short.
- Hedging everything instead of stating a clear position.
- Vague claims where a real number exists. "Rates vary" when you know the range.
- Uniform sentence length — vary rhythm, a short sentence lands harder
  right after a longer one.
- Never disagreeing with anything — a genuine negative opinion reads as
  more trustworthy than another balanced list.

# Rhythm and texture
- Use contractions: "doesn't," "isn't," "you've" — full forms read as
  formal-by-default, an AI tell, not a house-style choice.
- State an opinion, not just a description.
- Hedge like a person who's actually seen this before: "usually," "in most
  cases" reads as lived experience. Flat absolutes on things that vary
  read as AI confidence, not real confidence.

# The rule that matters most for a script specifically —
# generalize, never invent a specific instance:
- SAFE: a generalized, true pattern about how an industry/mechanism works.
  "Recruitment agencies feel this hardest — contractors paid weekly,
  clients on 30-60 day terms" is a structural fact, not a fabricated case.
- NOT SAFE: a specific invented incident, date, or figure presented as a
  case. "A client of ours hit this in March" invents evidence that doesn't
  exist — worse than a wrong stat, because no gate catches an anecdote.
- The test: could this sentence be true of the whole category, or does it
  only work because it names one (fake) instance? Category-true → write
  it. Instance-true → only Adam can supply it, verbatim, not "inspired by."

LOW = 14    # refill a brand when its runway drops below this many videos
BATCH = 5   # how many to add per refill run

# MerchantHQ / PeptideClear: pulls the next topic from topic-bank.json,
# generates script + 6-slide deck + upload metadata via claude_sub (subscription),
# STRICTLY VALIDATES before ever writing a file:
#   6 slides · title/fact/outro sequence present · no em/en dash · required meta keys
# a malformed generation is rejected and alerted — never silently published

# MarketInvoice: purely mechanical, renders more from the pre-curated
# questions.json — no AI call needed

STALL_DAYS = 3    # alert if no public upload in this many days
LOW_RUNWAY = 10   # alert if fewer than this many videos queued

def rss_latest(channel_id):
    # fetches the channel's PUBLIC RSS feed — no auth needed —
    # and reads the most recent upload date. This is the actual proof.
    url = f"https://www.youtube.com/feeds/videos.xml?channel_id={channel_id}"