Animated Explainer Workflow (Light Producer)

You drive a 6-stage pipeline. You handle the creative stages in chat; the gen-ai CLI handles all media generation and rendering. State lives in ~/.gen-ai/projects/explainer/<slug>/ as JSON files.

Mode — interactive (default) or auto

Before anything else, decide which mode the user wants:

Interactive mode (default — pick this unless the user opts out)

The 4 creative stages each end with a hard STOP. You present, the user reviews, types continue / edits / picks. The full Rule One below applies. Right for first-time use, premium production, anyone who hasn't told you otherwise.

Auto mode (opt-in)

You execute all 6 stages end-to-end without STOPping. You still present each artifact briefly so the user can interrupt if they want, but you do NOT wait. You make the picks: best concept of the 3, best playbook for the concept, script as you'd write it. The user is signing up for "trust your judgment, go end-to-end."

Detect auto mode when the user's request includes one of:

• auto, auto mode, auto-approve, auto approve
• no approvals, skip approvals, don't ask, no questions
• just do it, yolo, full auto, end to end
• run it through, run all stages, no checks

If you see ANY of those, set mode = auto. Otherwise, mode = interactive.

If the user's wording is ambiguous, ask ONCE at the start: "Interactive (I pause at each stage for your review) or auto (I run end to end and you get the final video)?" — then proceed.

Even in auto mode, you MUST still:

• Announce the credit estimate before stage 5 (assets). One line:

"Spending ~1850 credits on assets now. Balance: 12,500 → ~10,650 after." Don't ask for permission, just announce. The user can Ctrl-C if they disagree.

• Read each director skill before its stage. Rule Zero is non-negotiable

regardless of mode.

• Surface real errors, not silently fail. If the CLI returns

{ "status": "error", "hint": "..." }, stop and tell the user.

• Stop on genuine ambiguity — if the user said "30s explainer" but the

topic obviously needs 3+ minutes to cover well, ask once before guessing.

Stage-by-stage behavior table

The 6 stages — FOUR are hard approval gates

Every creative stage is a hard stop. You do the work, present it, then STOP and wait for the user. Do NOT chain stages without explicit user approval.

1. research — you, in chat. Read references/research-director.md. STOP after presenting findings.
2. proposal — you, in chat. Read references/proposal-director.md. Pick a playbook with each concept. STOP until the user picks A / B / C.
3. script — you, in chat. Read references/script-director.md. Read the chosen playbook's audio.voice_style and reflect it in speaker_directions. STOP after showing the script.
4. scene_plan — you, in chat. Read references/scene-plan-director.md. Include the `playbook` field at the top of `scene-plan.json`. STOP after showing the scene table — this is the last gate before money is spent.
5. assets — CLI (~1500-3000 credits, 5-25 min wall time). Read references/asset-director.md. Write scene-plan.json (with playbook field) and script.json into the project dir. Then run gen-ai explainer:assets <slug> — the CLI auto-applies the playbook's image_prompt_prefix, image_negative_prompt, and music_mood.
6. render — CLI (~1-3 min). Read references/render-director.md. Run gen-ai explainer:render <slug> — playbook auto-flows from the asset manifest; ffmpeg uses its music_volume_db and narration_to_music_weight_ratio.

After stage 6: draft title / description / chapters / hashtags in chat. Then run gen-ai upload-to-drive <slug>/explainer.mp4 --name "<title>". Share the URL.

Rule Zero — Read the director skill before EVERY stage

Each of the 6 stages has a dedicated director skill at ~/.claude/skills/gen-ai-explainer/references/<stage>-director.md. You MUST read the director skill BEFORE executing each stage. Not after. Not skimmed. Read.

The director files are not "background reading" — they contain the exact process, query templates, schema shapes, self-evaluation rubrics, common pitfalls, and STOP gates for that stage. Skipping them produces lower-quality output that wastes the user's credits.

Skill-loading protocol (apply at the START of every stage)

1. Announce in chat: "Loading `references/<stage>-director.md`." One line. So the

user sees you're following the protocol.

1. Read the file with the Read tool. The full file.
2. Follow its Process steps exactly. Don't improvise — the directors

were written precisely so you don't have to invent the workflow each time.

1. When the director's STOP gate fires, STOP. Don't pre-load the next director.

Stage → director mapping (memorize this)

Do NOT (skill-loading violations)

• Skip reading a director "because you remember it from the last conversation."
• Read directors in batches "to save round-trips" — fresh context per stage.
• Improvise a stage from your general knowledge instead of following the

director's specific process.

• Carry stale director content from a previous stage into the current one

(e.g., applying scene-plan rules to the script stage).

• Silently drop director-mandated steps (web searches, self-evaluation,

pronunciation guides) "to be faster."

If you skip director-reading, the user will catch it: research will lack sourced URLs, the script will miss the narrative arc, scene plans will fail the 5-aspect checklist. Sub-quality output betrays the protocol.

Rule One — Approval gates in interactive mode

This rule applies in interactive mode only. Auto mode replaces STOP with "announce and continue" per the Mode section above.

In interactive mode, the four creative stages (research / proposal / script / scene_plan) each end with a hard STOP. After presenting your output:

• WAIT for the user to reply.
• If they say "continue" / "looks good" / "approve" / "go" — proceed to the next stage.
• If they say "edit X" / "rewrite Y" / "swap N" — revise, present again, STOP again.
• Iterate until they explicitly approve.

In interactive mode, Do NOT:

• Auto-advance to the next stage without user reply.
• Pre-draft the next stage "to save time" while waiting.
• Assume approval from silence.
• Skip showing the artifact to the user.
• Collapse multiple stages into one message.

If you skip a gate in interactive mode, the user pays for visuals they didn't sign off on. The whole point of interactive mode is human-in-the-loop control.

In auto mode the user has explicitly opted out of approvals — you go through all 6 stages and produce the video. You still announce each stage's output (so the user sees what you picked) and announce the credit estimate, but you don't wait for input.

Other rules

• The asset stage is expensive (~1500-3000 credits for 30-90s). Announce the

credit estimate before running, and confirm one more time at the start of stage 5.

• If a CLI call fails, read the error JSON's hint field and decide whether

to retry, re-plan, or surface to the user.

• Never override the default models unless the user asks. Defaults are:
• image: gemini-3.1-flash-image (Nano Banana 2)
• video: seedance-2.0
• voice: eleven-v3
• music: minimax-music

Resume protocol

If the user references an existing project, run ls ~/.gen-ai/projects/explainer/<slug>/ and decide which stage to resume from by which JSON files exist:

See pipeline.yaml for the machine-readable manifest.