samin12/youtube-video-development-skill/youtube-video-development

YouTube Video Development

Turn one rough video idea into one traceable, review-ready project. A manager agent coordinates the work; the user owns creative decisions and every production or publication approval.

Read references/configuration.md before the first run. Read references/channel-context.md before judging channel fit, selecting a format, packaging, outlining, scripting, designing, editing, publishing, repurposing, or analyzing. Read references/thread-system.md when creating or reusing Buzz threads. Read references/clickup.md before creating or changing ClickUp records.

Before sending any owner-facing reply, specialist handoff, status update, or ClickUp prose, run the installed $unslop skill over the draft. Preserve exact commands, approval sentences, IDs, hashes, URLs, quoted source text, and required artifact markers while simplifying the surrounding writing.

Start or resume one project

In Buzz #Process, start or resume one bounded project when the user clearly intends to develop a YouTube or content idea. Natural triggers include:

• Take the latest idea from Capture and help me develop it.
• Take this from Capture.
• I want to develop this video idea: ...
• I want to make a video about ...
• /develop-video ...
• /develop-video capture to use the newest actionable idea in #Capture.
• /develop-video-images ... to explicitly request the full rendered package.
• /develop-video-plan ... to create titles and thumbnail directions without starting image jobs.

Do not require the user to type VIDEO:, name this skill, mention Grill Me, or trigger specialists and ClickUp separately. Treat /develop-video as an explicit alias for this same workflow, not as a separate implementation. Do not trigger on an ambiguous phrase such as develop this unless the message, referenced Capture item, or active thread makes the video/content intent clear.

1. Resolve the source idea. When the request says Capture, read the referenced item or the newest actionable idea in #Capture; never silently substitute an unrelated idea. Resolve a short stable label from that source and the user's wording.
2. Run scripts/clickup_video_project.py bootstrap --label "<label>". Before Grill Me begins, this must create or reuse both the exact ClickUp Doc and task, initialize a minimal managed root when the Doc is new or empty, link the Doc from the task at GRILLING, and verify both through readback. Treat a missing Doc receipt as a blocked bootstrap even if task creation succeeded.
3. Read references/video-grilling.md, then invoke the available Grill Me skill automatically with that video-specific interview contract. Michael must start from the Capture item, supplied references, available research, and channel context rather than interviewing from a blank page. He keeps grilling until the video's unique thesis, points, examples, stories, proof, asset map, and logical progression are concrete enough that another creator could not make the same generic video from the brief. Convert confirmed answers into a Video Decision Brief with the Shared Context Capsule from references/channel-context.md.
4. Resolve packaging mode. Natural requests, /develop-video, and /develop-video-images default to render; /develop-video-plan uses plan. Do not spend image credits until the brief is confirmed.
5. After confirmation, delegate distinct assignments from the same brief. Darryl [Packaging] owns concept/angle reconciliation, outlier evidence, titles, verified critical assets, Codex-first image generation, generation specs, and fit-check. Pam [Designer] produces divergent visual directions and critique. Andy [Outliner] and Creed [Asset Gatherer] work concurrently. Include the Shared Context Capsule in every handoff.
6. In render mode, invoke the installed $video-package workflow and run its renderer preflight. Use a renderer only when the current runtime exposes a working image tool that returns a durable file; the packaging skill chooses among available Higgsfield or GPT Image capabilities. Use Grok and Pam for concept or critique work, not as renderers. Keep the Thumbnail QA loop's three-attempt cap. In plan mode, return exact prompts but do not start image jobs.
7. Reconcile every reply against the confirmed brief. Reject unsupported claims, contradictions, duplicate work, temporary provider URLs, and packaging the video cannot fulfill.
8. Synthesize one page-marked artifact with references/clickup-doc-template.md. In render mode, finalize with --image LABEL=PATH for A-D and the comparison grid. The helper edits the bootstrap Doc in place, attaches the files, substitutes stable ClickUp URLs, creates or updates the page tree, and verifies readback.
9. Return the freshly verified task, Doc, Generated Thumbnails, outline, and research links plus the exact approval_requests emitted by finalize. They have the form I approve the package for ClickUp task <task-id> at digest <sha256>. and I approve the outline for ClickUp task <task-id> at digest <sha256>. Ask Samin to post each as a standalone Process message and record its 64-character event ID. Silence, a reaction, another author's message, an old approval for another project or revision, an arbitrary reference, extra clauses, a scoped rejection, or an ambiguous “looks good” is not approval.
10. After both approvals, tag Jim Halpert [Scriptwriter] in [03 WRITE]. Give him the confirmed brief, approved package and outline versions, Shared Context Capsule, proof/assets, runtime target, voice, CTA limits, and the exact full-script acceptance checks. Never claim Jim was dispatched or finished unless both the visible handoff and his reply exist in Buzz.
11. Reconcile Jim's reply against the approved outline. The Full Script must contain one matching H3 section per approved beat, in order, and every section needs matching [A-ROLL] and [/A-ROLL] markers around at least 40 plain-prose spoken words plus an explicit screen, demo, B-roll, or graphic cue; comments, headings, URLs, code, lists, and production notes do not count as spoken copy. Build the two-page script artifact described in the template, including the approved package and outline digests, then run scripts/clickup_video_project.py script with both verified Buzz event IDs. This phase refuses GRILLING and NEEDS_RESEARCH, cryptographically verifies that both owner approvals use the exact task-and-digest-bound sentence before any write, updates only 05 Script & Production and 06 Approval & Status in the existing Doc, advances the existing task to scripting / SCRIPT_REVIEW, and verifies exact readback.
12. Return the freshly verified task, Doc, outline, script, Generated Thumbnails, and approval-page links. Stop at SCRIPT_REVIEW for the user's next decision. Do not film, upload, schedule, publish, or approve the script for the user.

For an authorized CLI-only start, use scripts/start_video_workflow.py. Sending a Buzz message represents the user externally; require authorization for that message before running the command.

Grill before delegation

Do not dispatch specialists until the user confirms a shared understanding of:

• the intended viewer and their problem or desire;
• the video's single promise, unique thesis, and the exact difference from common coverage;
• the claims or teaching points, including the objection or misconception each one resolves;
• concrete examples, personal stories, demonstrations, comparisons, and receipts that make each point believable;
• the logical progression, including prerequisites and useful Level 1, Level 2, and Level 3 framing when the idea genuinely has levels;
• opening tension, escalation, payoff, and required proof;
• supplied references, reusable assets, source-to-claim links, and research gaps;
• creative boundaries and the next owner approval.

Research discoverable facts instead of asking the user to find them. Use those facts to ask sharper questions and offer grounded options. Mark unknown facts as gaps rather than filling them with guesses. The session is not complete if the brief could describe any competent video on the topic; it must capture why this is Samin's video and what viewers will get here that the source material and competing videos do not provide.

Keep one source of truth

Buzz contains the creative discussion and specialist threads. ClickUp contains the durable project status and the review-ready synthesis. The manager maintains one canonical status post and one exact task/Doc pair per video.

Every handoff names the source brief, packaging mode, Shared Context Capsule, required artifact, acceptance check, responsible agent, reply destination, unresolved questions, and next owner. The confirmed project brief overrides channel defaults when the user deliberately chooses a different format or direction. Every rendered Generated Thumbnails page must show Package A-D as four explicit title-thumbnail cards and retain one marker per card in the form <!-- package-pair: A | Exact title -->; the finalizer binds the A-D map into its receipt. Never claim an agent was called unless its message or reply is visible in Buzz. Never call a thumbnail generated unless a real file exists and passed QA. Never call it inserted until a stable URL and the matching package-pair map are visible in ClickUp readback. Never claim ClickUp completion until the helper returns a fresh readback receipt.

Stage gates

Use these states when they fit the host workspace:

GRILLING, NEEDS_RESEARCH, CONCEPT_REVIEW, PACKAGE_REVIEW, OUTLINE_REVIEW, SCRIPTING, SCRIPT_REVIEW, READY_TO_FILM, FILMED, EDIT_REVIEW, READY_TO_PUBLISH, PUBLISHED, BLOCKED, and STOPPED_NO_PROGRESS.

Each pass changes one current artifact, checks it against the current gate, records evidence, and then advances or stops. Stop when the gate passes, the next step needs the user, required inputs are missing, or another pass produces no measurable improvement.

Simulation boundary

A message labeled TRAINING SIMULATION may include mock answers and RUN FULL SIMULATION. Keep that label in every durable artifact, stop at CONCEPT_REVIEW, and never publish, schedule, upload, or represent the simulation as production work.

Never publish, schedule, upload, send metadata, or approve on the user's behalf without explicit authorization for that exact action.