clip-results

name	clip-results
description	Use when the user wants the closed-loop performance of an OpusClip clip batch across all the platforms it was posted to — "clip results", "how did the clips do", "cross-platform clip performance", "closed loop on the clips", "did the clips work". Pulls per-platform metrics natively (no API exists) and renders one consolidated report.
user_invocable	true

clip-results

Closed-loop measurement for a posted OpusClip batch. Given one or more OpusClip project IDs/URLs, gather each clip's performance on every platform it was posted to (YouTube, Facebook, Instagram, TikTok, LinkedIn) and produce one consolidated cross-platform report.

The reliability problem this solves

There is NO programmatic analytics path. The OpusClip API has no analytics/insights endpoints; the postIds it stores are OpusClip-internal (can't deep-link to the native post); and these posts bypass Buffer, so Buffer stats don't see them. The ONLY way to get the numbers is each platform's native analytics, read in-browser, matched to clips by the [opus:CLIPID] tag embedded in every caption.

Per the Primitive Test: browser navigation/extraction stays here in the prompt (UIs change — a script of selectors would rot; the Bitter Lesson says the model drives changing UIs/logins better). The deterministic work — the YouTube [opus:]→metrics join, aggregation, ranking, and report render — is the Go binary clip-results report.

When to Use

Use when the user asks how a posted clip batch performed across platforms.

Do NOT use when:

• The user wants to create/curate/schedule clips → that's the opusclip skill.
• Only YouTube numbers are wanted → /yt-analytics covers that (but note: YouTube is typically <⅓ of true cross-platform reach — see findings).

Inputs

$1... — OpusClip project IDs or clip.opus.pro/clip/<id> URLs. The 12-char id in that URL (e.g. P3051823ab0w) is the project id, not a single clip. One project = one source video's clips.

Steps

1. Refresh YouTube data (the join uses current metrics)

cd /Users/mikelady/dev/youtube_analytics
go run . fetch && go run . fetch-analytics --all   # skip if videos.json is fresh today

2. Clip inventory per project

The OpusClip CLI is bundled in the plugin cache:

CLI="$(ls -d /Users/mikelady/.claude/plugins/cache/opus-skills/opusclip/*/scripts/opusclip | tail -1)"
"$CLI" list --project <PROJECT_ID> --summary    # clip_id, title, score, duration

Capture clip_ids + titles. (Confirm OPUSCLIP_API_KEY is set.)

3. Scaffold the results file

Create data/opus_clips/<PROJECT_ID>-platform-results.json:

{ "project": "<PROJECT_ID>", "source_video": "<title (videoId)>",
  "measured_at": "<YYYY-MM-DD>", "rows": [] }

You'll append one row per (clip × platform) as you read each dashboard. Do not add youtube rows by hand — the Go tool joins them in step 5.

4. Browser pass — read each platform's native analytics

Load chrome tools first: ToolSearch select:mcp__claude-in-chrome__tabs_context_mcp,mcp__claude-in-chrome__navigate,mcp__claude-in-chrome__get_page_text,mcp__claude-in-chrome__browser_batch,mcp__claude-in-chrome__computer.

Use get_page_text, not screenshots — it returns the full caption (with the [opus:CLIPID] tag for exact matching) plus the metric columns, far faster and token-cheaper. Append rows {opus_id, title?, platform, views, likes, comments, follows}. Platform string ∈ facebook|instagram|tiktok|linkedin_org|linkedin_personal (youtube is auto-joined).

• Meta Business Suite (covers BOTH Facebook + Instagram) — business.facebook.com/latest/posts/published_posts/ → Content. Resize window wide (~1920). Columns → add Views + Follows, remove Reach (deprecated) + Shares. The DOM appends rows as you scroll; get_page_text returns all loaded rows. IG posts show the full caption + [opus:] tag; FB posts show a short title only — match an FB post to its clip by the same publish time-slot as the tagged IG post (they're scheduled together, ~2–4 min apart). FB ≫ IG on views, consistently.
• TikTok — tiktok.com/tiktokstudio/content → Views/Likes/Comments columns; captions carry the [opus:] tag. Virtualized list — scroll in modest increments and dedupe by tag (big jumps skip rows). Highest engagement of any platform.
• LinkedIn — EVC org — linkedin.com/company/<ORG_ID>/admin/analytics/updates/ → Content engagement table (impressions per post). Usually negligible.
• LinkedIn — Mike personal — Me menu → Posts & Activity → linkedin.com/in/<vanity>/recent-activity/all/; each post shows "N impressions". (An OpusClip hasConflict: true flag does NOT mean the post failed — verify.)

If a platform nav is permission-denied, ask the user before retrying; don't loop.

5. Consolidate (deterministic — the Go binary)

go run . clip-results report \
  --results data/opus_clips/<PROJECT_ID>-platform-results.json \
  --out     data/opus_clips/<PROJECT_ID>-platform-results.md

It auto-joins YouTube rows from videos.json by [opus:] tag, then computes per-platform totals + ranking, top clip per platform (the breakout view), cross-platform per-clip reach, and grand total — and writes the markdown. Never hand-compute the totals.

6. Report to the user

Lead with: cross-platform grand total vs the YouTube-only number (YouTube is usually a minority of true reach); platform ranking; the platform-specific breakouts (top clip differs per platform → cross-posting earns its keep); conversion (follows/subs — typically ≈0, clips buy reach not subscribers); and that OpusClip's per-clip score is a weak predictor of real views.

7. Commit

data/opus_clips/* is gitignored, so the results files stay local. Commit only code/skill changes.

Common Mistakes

• Treating the OpusClip API or Buffer as an analytics source — neither reports post performance. Native dashboards only.
• Using screenshots instead of get_page_text — slower, and you lose the clean [opus:] tags that make matching exact.
• Hand-summing platform totals — use clip-results report (it's tested).
• Forgetting fetch-analytics --all first — the YouTube join then uses stale views.
• Assuming a single platform's hero clip is THE winner — breakouts are platform-specific.

Known findings (baseline: P3051823ab0w, measured 2026-05-30)

Facebook was the #1 channel (beat YouTube); true reach ≈3.4× the YouTube-only figure; breakouts platform-specific; conversion ≈0; Opus score a weak predictor. Full data: data/opus_clips/P3051823ab0w-platform-results.md. See also the project memory project-opusclip-platform-reach.

This skill lives in the repo

Symlinked ~/.claude/skills/clip-results → /Users/mikelady/dev/youtube_analytics/skill/clip-results. Edit in the repo. The Go transport is cmd_clip_results.go (+ _test.go).