tl-top-partnerships - SKILL.md Agent Skill

name: tl-top-partnerships description: External brand-user performance report. Ranks a brand's sponsorships by effective CPM once the sponsored videos went live, and compares live eCPM against the sold-date projection. Use whenever a brand user asks "which of my sponsorships performed best", "top partnerships this year", "best ROI deals", "effective CPM on my deals", "which sponsorships overperformed", "/top-partnerships", or any variation of "show me my best-performing sponsorships". This is the brand-side equivalent of internal performance reporting — fire it eagerly any time a brand wants to look back at their booked deals through a performance lens, even if they don't say the words "CPM" or "eCPM".

Top Partnerships (Brand-side)

Helps a brand look back at their sold sponsorships and see which ones delivered the lowest effective CPM (eCPM) once the videos went live, vs the projection at sale.

Triggers

/top-partnerships — defaults to calendar YTD
/top-partnerships <range> — e.g. /top-partnerships 2025, /top-partnerships "last 12 months", /top-partnerships "Q1 2026"
Natural language: "top partnerships this year", "best sponsorships", "which deals performed best", "effective CPM on my deals", "show me my best ROI sponsorships"

What this skill computes

For every sold sponsorship the brand has where the video has actually gone live (has a publish_date and a non-null live views count):

Sold-date eCPM = price / projected_views_at_purchase_date * 1000
- The projection captured on the adlink at the moment the deal was sold. This is the eCPM the brand "agreed to."
Live eCPM = price / views * 1000
- The actual eCPM now that the video has accumulated views.
View ratio = views / projected_views_at_purchase_date
- 1 means the video out-delivered its projection.
Delta = live_eCPM - sold_date_eCPM
- Negative delta = the deal got cheaper per view than promised (good for the brand). Positive delta = the deal underdelivered.

It also pulls future bookings — any sponsorship that is sold, or open with the brand having reviewed it (brand_approval pending or approved), with a send date strictly after today — and tags each deal and each channel with the earliest future send date, or "Re-book - no future spot" if none exists. This turns the report into an actionable list, not just a backward look.

Output

A Google Sheet with two tabs, owned by the caller's Google account:

By Deal — one row per sponsorship, ranked by live eCPM (best first). Columns: rank, channel, title, video_url, send_date, publish_date, price, promised_views, live_views, view_ratio, sold_date_ecpm, live_ecpm, delta_ecpm, measurable, next_booking.
By Channel — one row per channel, aggregated across all that channel's deals in range. Combined live eCPM is sum(price) / sum(live_views) * 1000 (volume-weighted, not an average of CPMs). Sorted by combined live eCPM. Columns: channel, deals, measurable_deals, total_price_usd, total_promised_views, total_live_views, view_ratio, sold_date_ecpm, live_ecpm, delta_ecpm, next_booking.

In chat: a short summary + top-10 channels table + the sheet URL.

Workflow

Step 1 — Resolve the brand

Run tl whoami --json and read the brands array.

One brand → use it silently.
Zero brands → tell the user this skill is for brand-user profiles and stop.
Multiple brands → ask which one. Don't guess.

Step 2 — Resolve the time range

Default = calendar YTD (Jan 1 of the current year through today).

Accept these forms in the user's input:

2025 or "2024" → that full calendar year
"last 12 months" → trailing 12 months ending today
"Q1 2026", "Q4 2025" → that quarter
"YTD" → explicit current YTD
Anything else → ask the user to clarify, don't silently pick

Convert to a send-date-start / send-date-end pair (YYYY-MM-DD strings). Use send_date as the time anchor because that is when the sponsorship actually ran for the brand — purchase_date can be months earlier.

Step 3 — Run the script

python3 <SKILL_DIR>/scripts/top_partnerships.py \
  --brand "<BRAND_NAME>" \
  --send-date-start <YYYY-MM-DD> \
  --send-date-end <YYYY-MM-DD>

<SKILL_DIR> resolves to this skill's directory at invocation time (same convention as tl-views-guarantee, tl-keyword-research).

The script does everything: pulls sold deals in range (paginated), pulls all future bookings, computes per-deal and per-channel metrics, creates a Google Sheet with two tabs, shares it back to the caller, and prints a markdown summary plus the sheet URL.

It uses tl for data and gws for sheet creation. Both must be on PATH and authed.

Step 4 — Present the result

Take the script's stdout as-is. It already contains:

Summary line — total sold deals, measurable count, median live eCPM, count overperforming.
Top 10 channels by combined live eCPM — markdown table with the Next booking column bolded when it says "Re-book."
Sheet URL — point the user at the two tabs.

If more than half the top-10 channels show "Re-book", call that out in one sentence as the headline action item. If most of the top channels already have follow-ups booked, congratulate briefly and stop.

Keep the writeup tight. No em dashes, no "just wanted to", no hedging. The data does the talking.

Brand-user mode notes

This skill assumes a brand-user tl auth. It uses only public CLI commands (tl whoami, tl sponsorships list) — no tl db pg and no Elasticsearch.
The tl sponsorships list endpoint already filters to deals the calling profile is allowed to see, so passing brand:"<name>" is a belt-and-braces filter rather than a privacy boundary.
View counts come from TL's own tracking on the views field returned by the CLI. They're the same numbers the brand sees in the TL dashboard, so the eCPMs are reconcilable with what they see in-app.
Don't include creators' contact emails, internal notes, or owner_* fields in the brand-facing output. The script already drops them from the CSV.

Edge cases worth mentioning to the user (only if they apply)

A deal that ran very recently (last 14-28 days) may show a misleadingly high Live eCPM because views are still accumulating. Mention this only if more than half the top-10 deals have a send date inside the last 28 days.
If the brand has zero measurable deals in the range, say so plainly and suggest broadening the range (e.g., last 12 months).