tl-views-guarantee

name: tl-views-guarantee tl-blurb: size a multi-video views guarantee description: Calculate the optimal views guarantee (VG) for a multi-video sponsorship buy with a YouTube creator. Given a channel ID or name, returns "video bundle size / views guarantee / likelihood to hit" based on bootstrap simulation of the channel's recent video performance (view counts measured at video age ~30 days). Use when someone asks "what VG should I push for with [creator]", "how many videos should I buy from [creator]", "calculate VG for [channel]", "what's a safe guarantee for [channel]", or anything involving setting views guarantees in a sponsorship deal. Triggers on "VG", "views guarantee", "views minimum", and any request to size a multi-video buy.

Views Guarantee (VG) Calculator

Runs a single script and returns one answer.

Usage

User says:

• "What VG should I push for with Sydney Watson?"
• "Calculate VG for channel 20107"
• "How many videos should I buy from Stephen Gardner?"

Default path — this is what you do almost every time. Run the script with just the channel ID or name, no flags (<SKILL_DIR> is this skill's directory):

python3 <SKILL_DIR>/scripts/vg.py <channel_id_or_name>

Print the script's four-line output back to the user verbatim — don't reformat it into a table, add commentary, or pull in extra metrics. Then, in one short line, offer to go deeper (e.g. "Want the full bundle-size sweep, a larger buy, or the per-video / cadence breakdown?"). Keep the first answer minimal; only reach for the flags below when the user actually asks for more. Don't volunteer them by running them speculatively.

Optional flags — only when the user explicitly asks for more

Add these one at a time, matched to what the user requested — never stack them by default:

• --sweep: the user wants to see every bundle size, not just the recommendation. Prints the recommendation followed by the full table.
• --max-bundle-size N: the user is considering a larger / long-term buy (5+ videos). Raises the ceiling from the default 4 up to 10.
• --target max: the user (or the brand) prefers the biggest bundle that still clears the threshold rather than the smallest. Pair with a raised --max-bundle-size to actually pick a larger buy. (--target min is the default and need not be passed.)
• --rich-output: the user wants the extra economics — Per-video VG, Upload cadence (longform uploads per month), and Campaign length (≈ bundle size ÷ cadence). In sweep mode it also adds Campaign length and VG/video table columns plus Projected views / Sample size / Sample window context lines.
• --full-range: only when the script itself printed an INCONSISTENT RESULTS banner suggesting it — forces the sample to the full 30–240d window instead of the staircase.

What the script does

1. Resolves channel ID from name if needed (via tl channels find <name>)
2. Pulls projected_views from the channel record
3. Picks a sample window matching the range used for the channel's projected views: primary is videos published 30–90 days ago; if fewer than 5 videos survive, widens to 30–120, then 30–180, then 30–240. Never samples beyond 240 days (the outer bound of that range).
4. Looks up each video's view count at age 28–32 days
5. Bootstrap-simulates 10k buys for each video bundle size from 1 up to --max-bundle-size (default 4, configurable up to 10)
6. For each bundle size: VG = max(0.75 × bundle_size × projected_views, p20 of bootstrap totals)
7. With --target min (default), picks the smallest bundle size whose likelihood ≥ 80%. With --target max, picks the largest such bundle, capped at --max-bundle-size.
8. If no bundle size hits 80%, falls back to the closest match: highest likelihood for --target min, largest bundle for --target max. The result is reported with its actual (sub-80%) likelihood as a negotiation signal.
9. If the sample came from anything other than the primary 30–90d window, prints a FALLBACK WINDOW banner — the bootstrap reflects older performance than the projected-views baseline, so treat the result as less current.
10. If the primary sample is thin (fewer than 10 videos) and the staircase didn't already use the widest window, the script silently re-runs the bootstrap against the full 30–240d sample and compares. If the two results disagree on the chosen bundle size, on whether they reach the 80% likelihood threshold, or by ≥10% on VG, it prints an INCONSISTENT RESULTS banner suggesting you re-run with --full-range to inspect the wider view.

Output format

This is the default and the only output you produce unless the user asks for more. Four lines — relay them as-is:

Channel: Sydney Watson (20107)
Video bundle size: 2
Views guarantee: 536,422
Likelihood to hit: 80%

Channel: Sydney Watson (20107)
Video bundle size: 2
Views guarantee: 536,422
Likelihood to hit: 80%

Bundle size            VG  Likelihood
          1       257,989        71%
          2       536,422        80%  ←
          3       841,851        80%
          4     1,166,196        80%

Channel: Sydney Watson (20107)
Video bundle size: 10
Views guarantee: 3,109,269
Likelihood to hit: 80%
Projected views: 343,986
Sample size: 14 videos
Sample window: 30-90d
Per-video VG: 310,926
Upload cadence: 6.2/month
Campaign length: ~1.6 months

Bundle size  Campaign length            VG    VG/video  Likelihood
          1      ~0.2 months       257,989     257,989        71%
          2      ~0.3 months       536,422     268,211        80%
          ...
         10      ~1.6 months     3,109,269     310,926        80%  ←

Edge cases

• <5 historical samples: skill exits with insufficient data — only <count> videos with 30-day view data (need 5+).
• 5-9 samples: prints ⚠️ THIN SAMPLE: only <count> videos — bootstrap variance is high, treat result as approximate. underneath the result.
• No projected_views on the channel record: usually means the channel hasn't published enough content to establish a baseline (not a data staleness issue). Statistical VG isn't available — fall back on your own judgement.
• Volatile/underperforming creators: the skill reports the floor VG with whatever likelihood the data supports (could be 60–70%). A lower likelihood isn't a failure — treat it as a negotiation signal. A missed guarantee typically just means the creator owes a make-good video, so a guarantee you're comfortable accepting at lower confidence can still be the right call.

Methodology notes

• Floor: 0.75 × bundle_size × projected_views (the 75% rule, aggregate)
• Skew adjustment: max(floor, p20 of bootstrap) — the bootstrap p20 will exceed the floor only when the channel has tight distribution and the floor would over-shoot, which is rare. In practice the floor binds.
• Bootstrap = resampling with replacement from actual historical 30-day view counts. No distribution assumptions.
• Sample window mirrors the same range over which the channel's projected views are computed, so the bootstrap reflects the same era of the channel that defined projected views. Staircase: 30–90 → 30–120 → 30–180 → 30–240. Always keep the bootstrap window inside the projected-views window — sampling beyond it mixes current projected views against a stale empirical distribution and produces misleading "channel under-delivers" results on recently-spiking channels.
• Per-video "30-day view count" = the average of all view-count snapshots recorded for that video between ages 28 and 32 days inclusive, cast to integer. Averaging across the 5-day window smooths single-snapshot jitter; alternative aggregations (min ≈ age-28 snapshot, max ≈ age-32 snapshot) would lock to one end of the window.
• Likelihood threshold for picking the bundle size = 80%.
• Default max bundle size = 4; raise via --max-bundle-size up to a hard ceiling of 10. 5+ video buys imply long-term campaigns — pair the raised ceiling with --target max to actually pick a larger bundle.
• Bundle-size sweep starts at 1 — single-video buys allowed when data supports it.
• Upload cadence (--rich-output) is derived from the videos already fetched for the bootstrap — no extra query. If the fetched set spans more than 180 days, only the most recent 180 days are used; if it spans less (high-cadence channel), the actual span is used. Cadence = (videos − 1) ÷ months so the rate doesn't leak into the blank intervals before the earliest and after the most recent video. Campaign length = bundle size ÷ cadence.