Youtube Channel Operator

YouTube Channel Operator

Build a channel-agnostic YouTube operations core with two interfaces:

1. typed raw API methods for trusted power users;
2. guided workflows that resolve exact resources, preview a change set, and apply it after approval.

Use slackbot-builder for Slack transport maturity and data-chatbots for proposal/apply semantics. Use this skill for the YouTube-specific domain layer that joins them.

The one rule

Raw and guided operations must share one deterministic executor.

The model may choose an operation and draft parameters. It must never construct authorization, call arbitrary URLs, or independently decide what was applied. The server resolves the account and resource, builds the provider request, persists the exact reviewed action, applies it once, and verifies the result.

Route the task

Do not load every reference automatically.

Distinguish the three execution surfaces

Never claim an MCP or private endpoint can perform a Studio-only action unless the owning UI is actually controlled and verified.

Non-negotiable invariants

1. Pin the account. Map an immutable account key to one expected channel ID and one credential namespace. Verify channels.list(mine=true) before drafting and applying.
2. Isolate channels. Use separate OAuth grants and secrets per brand/channel. Never infer routing from a title or reuse one channel's refresh token.
3. Expose methods, not arbitrary HTTP. Raw tools accept a registered method ID, typed parameters, body, and optional immutable media asset. Reject custom URLs, hosts, headers, tokens, or an unreviewed execute flag.
4. Use full metadata for content work. Treat catalogs, search hits, and descriptionExcerpt fields as discovery records only. After resolving exact video IDs, fetch videos.list(part=snippet,...) and use the complete snippet.description before extracting, comparing, or rewriting content. Never conclude text is absent because it is absent from an excerpt. YouTube detail reads should use 50-video pages to match YouTube's native request limit and continue larger sets with an explicit page or cursor. YouTube update endpoints use replacement semantics for included parts, so merge a patch into the full current mutable part and review the complete outgoing representation.
5. Apply exactly what was reviewed. Persist the exact resource ID, base ETag, before state, desired patch, normalized provider request, actor, and expiry. If live state drifts, conflict and create a fresh proposal.
6. Keep Slack thin. Verify/ack/dedupe/render in Slack; resolve, validate, persist, execute, and audit in a channel-agnostic core.
7. Make external writes durable. Use a workflow or queue for uploads, scheduled rotations, multi-step packages, retries, and analytics follow-ups.
8. Verify the owning surface. Refetch the resource after every API mutation. For Studio-only work, reopen the exact Studio/video record.
9. Keep immutable audit history. Record the actual human proposer and approver, provider/account, exact request, before/after, attempts, and outcome. Never store OAuth tokens or Authorization headers in D1, logs, traces, or Slack.
10. Separate change proof from performance impact. A successful write proves state changed. Analytics later describe correlation, not causality, unless YouTube's native concurrent experiment supplies the result.
11. Separate acquisition objectives and evidence. Cheap paid views can be a valid awareness outcome. Report paid reach, organic behavior, attributed subscribers, and repeat engagement separately; do not collapse them into a single quality score or dismiss inexpensive views by default.

Build the two capability layers

Raw power-user layer

Expose a compact typed surface:

youtube.data.call
youtube.live.call
youtube.analytics.query
youtube.reporting.call
youtube.transcript.get

Require accountKey, a registered method ID, exact IDs/parameters, a validated body, and an optional immutable asset ID. Show the normalized provider request before execution. Keep the method registry complete even when common workflows cover only a subset. youtube.transcript.get is a read-only convenience tool over registered captions.list and captions.download calls; it must preserve the selected caption track and timestamps rather than scrape watch-page HTML.