Break Down A Channel

You take one YouTube channel and hand back a teardown that answers a single question: how did this channel grow, and what can the owner borrow? Not a data dump. A short, honest read that points at the moment things changed and the one move worth stealing.

This is a study tool, and it is the deep-on-one companion to research-my-channel: that skill reads a whole niche to decide what the owner should make; this one goes deep on a single channel to reverse-engineer how it took off. The performance craft it draws on (the outlier multiple, reading a video against its own channel's baseline) lives in knowledge/youtube-packaging-method.md; read it before you write the "what changed" section so the method has one home.

YouTube only, keyless, local. The channel history comes from the local yt-dlp tool (drivers/yt-dlp/) — no account, no key, it runs on the owner's machine. There is no TikTok or Instagram path here; say so plainly if asked.

Work the gates in order.

Step 0: Get the one channel

You study one channel per run. Get a handle or URL (@<handle>, https://www.youtube.com/@<handle>/videos). If the owner names several, do the most relevant one and say the others are separate runs. Confirm which channel back to them before pulling anything.

Step 1: Pull the channel history (local, keyless)

Pull the channel's video list with yt-dlp in flat mode, via Bash:

yt-dlp --flat-playlist --dump-json "https://www.youtube.com/@<handle>/videos" > <slug>-dump.jsonl

This returns one JSON line per video with view_count (rounded) and playlist_index (reverse-chronological order). It returns no dates in flat mode, which is expected and fine: this skill works from upload order, not calendar dates (see the driver README). If yt-dlp is not installed, install the binary locally (it is keyless, no account); if the pull is blocked or empty, say so plainly and offer to read a channel-page list the owner pastes instead. Never invent a video or a number.

Step 2: Run the breakout engine

Run the deterministic engine on the dump:

python ~/.claude/bos-run.py tool channel_breakdown <slug>-dump.jsonl --window 10 --min-segment 5 --out <slug>-breakdown.json

It returns a timeline (each video's rounded views and its outlier multiple = views over the median of the previous ~10 videos in order) and a breakout verdict (ok with a trigger video, or no_upward_inflection). The engine is the source of the numbers; do not re-derive them by eye. Read <slug>-breakdown.json.

Step 3: Write the teardown

Write channel-breakdown-<handle>-<date>.md and show the highlights inline. Five parts, in order:

1. The performance timeline. Describe the shape across upload order (steady, climbing, spiky) and call out the handful of biggest outliers by their real multiple and title ("this one did 11x the channel's baseline"). This is the spine.
2. The breakout moment. If breakout.status is ok, name the trigger video and what its multiple and the before/after median views say: the channel durably stepped up here. If it is no_upward_inflection, say so honestly: there was no durable step in this window; the wins were spikes that settled back, which is its own useful finding.
3. Spike or step? Say plainly whether the channel leveled up (a durable higher base) or fired a hit and returned to normal. The engine's minimum-segment rule means a lone viral spike will read as no durable step; trust that and say it.
4. What changed. For the trigger video (or the top outliers if there was no durable step), read what was different against knowledge/youtube-packaging-method.md: the topic, the angle, the title promise, the format. This is the transferable part, and the reason matters more than the number. Ground it in the real titles you can see; if a cause is a reasonable read rather than certain, say so.
5. The one move to borrow. Close with a single, do-it-this-week takeaway for the owner's own channel, drawn straight from what worked here.

Hard rules

• ❌ Keyless, YouTube-only, local. yt-dlp (kind local) is the data path; no account, no key, no other platform. Say plainly if asked for TikTok/Instagram.
• ❌ Real observed numbers only. Every view count and multiple comes from the engine's output on the real dump. Never invent a video, a number, or a multiple. "No durable step in this window" is a real, valid finding.
• ❌ No dates claimed from the flat dump. The timeline is upload order; do not present an invented calendar date as if observed.
• ❌ The funnel/monetization layer is out of scope (v1). No comment-to-DM, ManyChat, or offer/non-offer pillar analysis. Keep to the packaging-transferable core.
• ✅ This read is for the owner, not their customers. It is the assistant's own voice (the register is knowledge/research-method.md), so it carries no customer-facing copy obligation. It still uses no em dashes and invents nothing: the never-invent rule above is the load-bearing one here. Where a finding becomes a title or an angle the owner actually publishes, the skill that writes that copy (plan-my-youtube, script-my-video) owns the content rules for it.
• ✅ One channel per run. Bounded and finishable.
• ✅ Bands are relative, not quality. A video below ~1x is "below this channel's typical", never "under-performing"; roughly half a healthy channel sits there.

Output shape

channel-breakdown-<handle>-<date>.md with the five parts (timeline, breakout moment, spike-or-step, what changed, the one move to borrow), highlights shown inline, the full read saved to the file.