Video 25: Forked Subagents

Video 25: Forked Subagents

Published: 2026-04-23 Thumbnail: Face + "Subagents 2.0" folder icon (two small pixel-agent characters connected by dotted lines to a larger Claude flower tile). Same thumbnail across all three variants in R1.

Title A/B Test Round 1 (2026-04-23) — 2-hour snapshot

Same thumbnail across all (face + Subagents 2.0 folder icon).

Key takeaways (2-hour read, 13d 20h remaining):

• 6.5-point spread — real title-frame signal, not a thumbnail bottleneck. Big enough to act on but expect compression as sample grows (V23 R1 39% → R2 31.4% on the same control; spreads usually narrow).
• Version-label frame leads at 36.9% — consistent with the "X.0" track record: V20 "Plan Mode 2.0" (35.1% face, 38.5% faceless), V18 "Memory 2.0" (34.3%). The 2.0 framing works because it positions forked subagents as an upgrade-to-a-known-thing, not a new-thing-to-learn. Matches what the feature actually is (a generalization of V14's /btw + --fork-session prototype).
• Title-thumbnail echo is almost certainly amplifying the winner. Thumbnail text reads "Subagents 2.0" on all three variants — only variant #1 has a title that names the same thing. Variants #2 and #3 fight the thumbnail's own framing ("biggest subagent problem" and "10x more useful" don't land on what the thumbnail already tells you). This is NOT the redundancy anti-pattern — thumbnail names WHAT, title can name WHY/WHO/WHEN, but here only the winner has any alignment at all.
• "Fixed + Biggest Problem" at 32.7% — solid mid-tier but below V22's "Biggest Blind Spot" (34.3%). "Subagent problem" is more abstract than "blind spot"; the pain isn't sharp enough to feel nameable.
• "10x More Useful" at 30.4% is the bottom — two suspects stacked: (a) no "Anthropic" = documented 5-6pt authority tax (V12-V24 consistent), (b) "10x more useful" reads as vague evaluative hype, similar to V23 R2's "game-changer" comma-clause that flopped at 31.3%. The "specific numbers" rule from master summary refers to counts ("10 New Features") not magnitudes ("10x more") — different structure.
• 36.9% at 2hrs is mid-pack for a single-feature announcement. V14 peaked 46.8%, V15 49.4%, V22 44%. There's real headroom if R2 finds a stronger frame.
• New hypothesis — title-thumbnail anchor alignment matters more than frame strength when the thumbnail names the feature. Any R2 title should either (a) echo "Subagents 2.0" or (b) be run with a different thumbnail. Mixing frames that fight the thumbnail wastes slots.

Status: Early snapshot. Strong directional signal toward version-label frame. R2 run below.

Title A/B Test Round 2 (2026-04-24) — Superlative + upgrade vs version-label control vs segmentation

Same thumbnail across all (face + Subagents 2.0 folder icon).

Key takeaways:

• Narrow 3.6-point spread — tighter than R1's 6.5pt, and only 1.1pt between the new leader and the control. All three are legitimately strong; no breakout winner.
• Superlative+upgrade frame edges ahead at 34.9% — "Biggest Subagent Upgrade Yet" works because it implies a chain of upgrades and keeps the thumbnail-echo via "Subagent" + "Upgrade" (stays aligned with Subagents 2.0 concept). Below V22 "Biggest Blind Spot" (34.3%) territory — same family, similar scores.
• R1 winner regressed 36.9% → 33.8% — matches the now-documented early-round volatility pattern on every video where we've tested the same title across rounds: V18 (39.6%→37.2%), V23 (39%→31.4%), V24 (36.4%→39.2% went up), V20 (37.3%→42.5% went up). The 3.1pt drop here puts the version-label frame's steady state in the low-30s, not mid-30s.
• "Every Power User Will Use" at 31.3% — weakest of the three. Diagnosis: "Every" + "Will Use" reads as a prediction, not a fact. V11 R4 "Top 0.01% User's Guide" worked at 45.6% because it was specific statistical authority + owned by the speaker; "Every Power User Will Use" is third-person hearsay (same failure mode as V24 launch's "Top 0.01% Users Swear By"). Also reads weakest on thumbnail-echo: "Power User" introduces a new concept the thumbnail doesn't support.
• Major ceiling signal. Six distinct title frames tested across R1+R2 (version-label, problem-fixed, transformation+number, superlative+upgrade, segmentation+upgrade) all cluster in 30-37%. Per the skill's "Diagnosing Bottlenecks" rule: diverse frames producing a tight band = thumbnail is the ceiling.
• Untested at the 2.0 anchor: Pocock voice ("Claude Code Just Shipped Subagents 2.0") and insider/Internal frame (V15's 49.4% formula). The insider frame is the only remaining candidate with real upside potential at this anchor — worth one more R3 test before locking and switching to thumbnail rounds.

Title A/B Test Round 3 (2026-04-24) — Within-family word-choice test (fresh 14d window)

Same thumbnail across all. Round started fresh (13d 18h remaining at snapshot).

Key takeaways:

• 1.0-point spread — essentially flat. All three titles within 1pt. This is the definitive ceiling signal per the skill's "Diagnosing Bottlenecks" rule.
• Progressive spread compression across three rounds: R1 6.5pt → R2 3.6pt → R3 1.0pt. Each round, tighter band. The thumbnail is the ceiling and further title word-tweaks inside the "upgrade" family won't unlock upside.
• Eight title wordings tested across three rounds, none broke 37%. Peak was R1's 36.9% — well-documented as early-round optimism (same pattern as V23 39%→31.4%, V18 39.6%→37.2%). Steady state for this thumbnail is 33–35%.
• "Rebuilt" verb at 32.9% — fresh release verb didn't move the needle. "Rebuilt" slightly underperformed "Dropped" + upgrade/2.0, suggesting the verb isn't the lever. Drop verb-swapping as a future axis unless there's a specific hypothesis.
• Tie-breaker within 1pt: "Biggest Upgrade Yet" wins by 0.8pt but is abstract (no named feature). "Subagents 2.0 for Claude Code" exact-matches the thumbnail's "Subagents 2.0" text and names the feature concretely — strongest title-thumbnail echo.
• Insider frame still untested (V15's 49.4% "Internal" keyword). Skipped in favor of within-family word tests. Worth one targeted round ONLY if Ray wants to chase the 45%+ ceiling before moving to thumbnails — otherwise lock and ship.

Recommendation: Lock "Anthropic Just Dropped Subagents 2.0 for Claude Code" (maximum thumbnail echo, most concrete name, highest peak at 36.9%) and move to thumbnail testing. 1pt gap vs "Biggest Upgrade Yet" is within noise.

Thumbnail A/B Test Round 1 (2026-04-24) — early snapshot, ~13d 12h remaining

Title locked across all three: "Anthropic Just Dropped the Biggest Subagent Upgrade Yet" (R3 winner, not the alternative I'd recommended). Ray went with the leader by 0.8pt rather than the thumbnail-echo pick.

Key takeaways (early snapshot, expect ~3-4pt drift before final):

• Nate-style with /btw /recap /dream pills leads at 35.1% — first thumbnail in this video's testing to crack the title-rounds plateau (33-34%). The pills name three Anthropic-shipped uses of forked subagents straight from Ray's script — viewers who watched V14 (/btw), V18 (/dream), or know /recap will recognize them = trust-by-association. Same family-context principle that powered V22 command-list (41.1%) and V23 command-list (40.3%).
• 3.8-point spread — meaningful early signal but tighter than V22 thumbnail R1 (12.7pt) or V23 thumbnail R1 (7.9pt). Two reads: (a) the Nate-style execution has more upside but isn't fully tuned, (b) the title is the floor and thumbnails are doing surgical lifting at the margin. R2 with Nate variations would tell us which.
• Control at 33.6% essentially flat — same neighborhood as title-test rounds (33.1-36.9%). Confirms the original folder thumbnail was the floor that capped title testing. The 1.5pt gap to T2 is real but modest — the folder isn't dramatically broken, T2 is just slightly better.
• Faceless dark at 31.3% reproduces the V18 "It dreams." pattern (31.1%) and undershoots V20 "Plan Mode 2.0" faceless (38.5%). Diagnosis: T3 has the version label (Subagents 2.0) like V20 did, but the cloud-checklist icon is too abstract — doesn't say what the new thing is or why it matters. V20 worked because the version label + product framing did the work alone; T3 added an icon that didn't add information.
• Confirmed for the 3rd time on this channel — abstract cloud/system icons crash on faceless thumbnails. V18 (moon/brain icon: 31.1%), V22 R1 (faceless "Your new advisor.": 28.4%), now V23 (cloud-checklist: 31.3%). Retire abstract symbolic icons from the faceless thumbnail pool unless paired with a strong text-only hook.
• New insight — the /btw /recap /dream pills are functioning as a "family of slash commands" anchor, but for a NON-slash-command video. V22/V23 used command lists for slash-command videos; T2 uses pills naming three slash commands as proof-points for a feature-class video. Pattern generalizes: when the video is about a capability (forked subagents) rather than a command (/monitor), pill-list-of-features-that-use-it works as the family-context substitute.

Status: Early snapshot, T2 leads by 1.5pt. If Ray wants to push past 35%, R2 should test Nate-style variations (different pill combos, subtly-impressed closed-mouth + white oxford instead of open-mouth smile — V23 R2's 40.3% formula). If he wants to lock and ship, T2 is the call.

Files copied into references/thumbnails/v23-subagents-2-0/ — T1, T2, T3 all archived. uploaded.json manifest written.

Recommended thumbnail round when it's time

Three candidates that would each unlock different ceilings:

1. Command-list with /fork added to /plan /review /worktree /monitor — "this one is new" annotation + subtly-impressed closed-mouth + white oxford. Three consecutive wins on this format (V22 41.1%, V23 R1 37.2%, V23 R2 40.3%).
2. BEFORE/AFTER excalidraw — left panel: regular subagent with blank context losing nuance. Right panel: forked subagent inheriting the full conversation thread. Literally visualizes the video's central thesis. V22 R2 winner at 37.2%.
3. Nate-style /fork slash-command product shot — white browser window + /fork + flower icon + pill concepts (INHERITS / FORKS / REMEMBERS) + red arrow + subtly-impressed face. V24 winner at 39%.