videoclaw 3.0.0-alpha.1

package/skills/character-creator/SKILL.md

@@ -0,0 +1,303 @@
+---
+name: character-creator
+description: |
+  Create Go Bananas characters with profile images and multi-view reference sheets for consistent
+  character generation across scenes. This skill should be used when the user asks to "create a character",
+  "create characters", "make a new character", "design a character", "build a character reference",
+  "character for my video", "set up characters", "new character with reference sheet",
+  "character reference", or mentions needing consistent characters across scenes.
+---
+
+# Character Creator
+
+Create Go Bananas characters with the **Input Triad** (3 reference images) and multi-view character
+reference sheets for maximum consistency across scenes.
+
+## Canonical CLI (prefer it over the raw MCP)
+
+This skill is the **canonical owner of character creation** — other skills
+(`ugc`, `video-framework`, `movie-director`, `character-ad`) delegate here rather
+than hand-rolling it. For project-bound characters, drive the `vclaw` CLI — it now
+covers the whole create → sheet → extra-views flow, per the VideoClaw-native
+discipline:
+
+- **Create + reference sheet in one command:** **`vclaw video character-auto-create
+  --project <slug> --input <cast.json>`** (cast.json = `[{name, description, style}]`)
+  creates the Go Bananas character AND renders the multi-view Cinematic reference
+  sheet (style preset 55), downloads it, and registers it as an `identity` reference
+  sheet in `projects/<slug>/` — no MCP needed for the create + sheet step. The sheet
+  prompt leads with the character's `style`, so a flat-cartoon character gets a
+  flat-cartoon sheet (not the preset's default 3D/cinematic look). It also runs for
+  **existing / re-imported characters** (a name that already exists on Go Bananas):
+  it generates the sheet if the project is missing one — idempotent, so a re-run with
+  the sheet already present and on disk does not re-spend (delete the local sheet to
+  force a regenerate). And it **syncs the sheet into the Go Bananas character RECORD**
+  (merges the sheet's image id via `PATCH /characters/<id>`, preserving existing refs),
+  so `generate_with_character` actually sees it — not just the project. `--no-sheet`
+  opts out; the result reports `gbRefsSynced`.
+- **Extra Input-Triad views** (the close-up + supporting-cinematic shots below),
+  keyed to the character for consistency: **`vclaw video gen-image --project <slug>
+  --kind prop --character-id <id> --prompt "<view>"`** — the gobananas backend now
+  takes `--character-id` (and `--style-preset-id`), and omits the per-kind directive
+  when either is set so your prompt is authoritative.
+- **Record an externally-rendered sheet** with **`vclaw video reference-sheet-add
+  --project <slug> --type identity --name <Name> --gb-ref character:<id>:identity`**.
+
+The raw `mcp__go-bananas__*` recipe below is still valid for hands-on / one-off
+work outside a project, and documents the identity-lock principles; but for a
+videoclaw project prefer the commands above so the character + sheet land in
+`projects/<slug>/` where the rest of the pipeline can see them.
+
+## When to Use
+
+- Creating new Go Bananas characters for video projects
+- Setting up character consistency before scene generation
+- Building character references with multiple views (profile + close-up + cinematic + reference sheet)
+- Any request involving "create character", "new character", "character reference"
+
+## Preparing a multi-subject composite reference
+
+If the user gives you **one image containing multiple characters** (e.g. a couple, a duo, a group), do **NOT** feed that composite to image generation as a single reference. A multi-face composite passed as a loose reference (`reference_mode: "add"`) makes the model **average/blend the identities** and you get a generic look-alike, not the specific characters.
+
+Instead:
+
+1. **Crop the composite into one single-subject image per character** (one face/body per file). Single-subject references lock identity far better than a shared composite.
+2. **Register each crop as its own Go Bananas character** via `create_character` with that crop in `reference_image_ids` (then the Input Triad below per character).
+3. **Generate scenes with `generate_with_character` / `generate_with_multiple_characters`** using the character IDs — never the raw composite.
+
+This is the identity-lock rule: locked characters always go through registered characters, never loose prompting.
+
+## The Input Triad
+
+For maximum character consistency across scenes, always generate 3 reference images per character:
+
+| Image | Purpose | Aspect |
+|-------|---------|--------|
+| **Full Body Portrait** | Locks style, costume, outfit, and body proportions | Square |
+| **Extreme Close-Up** | Locks facial features, accessories (glasses, jewelry, scars) | Square |
+| **Supporting Cinematic** | Locks lighting vibe, 16:9 composition, environment style | 16:9 |
+
+**Why this matters**: The "Sunglasses Error" — if a character wears sunglasses but you only have a wide-shot reference, the AI sees them as too small to replicate. The extreme close-up forces the model to memorize facial accessories. Similarly, the cinematic shot trains the model on the correct aspect ratio and lighting style for scene generation.
+
+## Workflow Overview
+
+The character creation workflow has 5 steps per character:
+
+1. **Generate profile portrait** — Square full body reference, studio background, detailed Pixar/style rendering (Input Triad Image 1)
+2. **Create Go Bananas character** — Register with portrait as initial reference
+3. **Generate extreme close-up** — Square face-only shot with all facial details and accessories (Input Triad Image 2)
+4. **Generate supporting cinematic shot** — 16:9 environmental shot locking lighting and vibe (Input Triad Image 3)
+5. **Update character** — Add close-up, cinematic, and reference sheet images to the character
+
+## Step-by-Step Process
+
+### Step 1: Gather Character Details
+
+Collect from the user (or infer from context):
+
+| Field | Required | Example |
+|-------|----------|---------|
+| Name | Yes | "Lord Ram", "Ravn" |
+| Style | Yes | "3D Disney Pixar animated", "anime", "realistic" |
+| Physical appearance | Yes | Age, skin tone, build, hair, distinguishing features |
+| Outfit/clothing | Yes | Crown, armor, casual wear, etc. |
+| Expression/mood | No | Serene, fierce, playful |
+| Props/accessories | No | Bow and arrows, sword, staff |
+
+If the user provides minimal details, ask for clarification on style and key visual features.
+
+### Step 2: Generate Profile Portrait (Input Triad Image 1: Full Body Reference)
+
+Generate a square profile portrait for each character using `generate_image`. This is the **Full Body Reference** — it locks the character's style, costume, outfit, and body proportions.
+
+```
+mcp__go-bananas__generate_image(
+    prompt="[STYLE] character portrait of [NAME]. [DETAILED APPEARANCE DESCRIPTION].
+            Clean studio background, character portrait framing.",
+    aspect_ratio="square",
+    model_id="gemini-pro-image",
+    negative_prompt="[style-appropriate negatives]"
+)
+```
+
+**Key rules:**
+- Always use `model_id="gemini-pro-image"` (Pro model)
+- Always use `aspect_ratio="square"` for portraits
+- Include rendering style cues: "subsurface scattering", "warm lighting", "detailed rendering"
+- End with "Clean studio background, character portrait framing"
+
+### Step 3: Create the Character
+
+Register the character in Go Bananas with the portrait as the initial reference:
+
+```
+mcp__go-bananas__create_character(
+    character_name="[Name] [Style Tag]",
+    base_prompt="[FULL STYLE AND APPEARANCE DESCRIPTION]",
+    description="[SHORT DESCRIPTION]",
+    reference_image_ids=[portrait_id],
+    negative_prompt="[style-appropriate negatives]",
+    system_instruction="Maintain consistent [STYLE] with [KEY FEATURES] across all scenes.",
+    tags=["relevant", "tags"]
+)
+```
+
+**Key rules:**
+- Include style tag in name (e.g., "Lord Ram 3D Pixar") for disambiguation
+- `base_prompt` should be the full reusable appearance description
+- `system_instruction` should emphasize consistency of key visual features
+
+### Step 4a: Generate Extreme Close-Up (Input Triad Image 2: Facial Identity)
+
+Generate a square extreme close-up of the character's face only. This is the **Extreme Close-Up** — it locks facial features and accessories (glasses, jewelry, scars, tattoos).
+
+```
+mcp__go-bananas__generate_image(
+    prompt="[STYLE] extreme close-up portrait of [NAME]. [FACE DESCRIPTION - eyes, skin tone,
+            hair at face level, any accessories like glasses/jewelry/piercings/scars].
+            Fill the entire frame with the face. Ultra-detailed facial rendering,
+            subsurface scattering, [LIGHTING]. Clean studio background.",
+    aspect_ratio="square",
+    model_id="gemini-pro-image",
+    character_id=[ID from create_character],
+    negative_prompt="full body, hands, torso, background elements, [style-appropriate negatives]"
+)
+```
+
+**Key rules:**
+- Always use `character_id` from Step 3 to maintain consistency with the portrait
+- The prompt must focus entirely on facial features: eyes, eyebrows, nose, mouth, skin texture, facial hair, accessories
+- Add "Fill the entire frame with the face" to ensure tight framing
+- Add specific accessory details: glasses style, earrings, facial piercings, scars, birthmarks
+- Negative prompt must exclude full body / torso to force face-only composition
+
+### Step 4b: Generate Supporting Cinematic Shot (Input Triad Image 3: Lighting/Vibe)
+
+Generate a 16:9 cinematic shot of the character in an environment. This is the **Supporting Cinematic Shot** — it locks the lighting vibe, aspect ratio composition, and environmental style.
+
+```
+mcp__go-bananas__generate_image(
+    prompt="WIDE HORIZONTAL [STYLE] cinematic shot of [NAME] in a [ENVIRONMENT].
+            [COSTUME/OUTFIT DESCRIPTION]. [MOOD/VIBE DESCRIPTION].
+            Cinematic lighting, film grain, shallow depth of field.",
+    aspect_ratio="16:9",
+    model_id="gemini-pro-image",
+    character_id=[ID from create_character],
+    negative_prompt="[style-appropriate negatives]"
+)
+```
+
+**Key rules:**
+- Always use `aspect_ratio="16:9"` and add "WIDE HORIZONTAL" in prompt for reliable landscape
+- Always use `character_id` from Step 3 to maintain consistency
+- Choose an environment that matches the project's scenes (urban, nature, studio, etc.)
+- Include mood/atmosphere cues: "golden hour warmth", "moody blue tones", "bright cheerful daylight"
+- This shot trains the model on how the character looks in actual scene conditions (not just studio)
+
+### Step 4c: Generate Character Reference Sheet (Optional but Recommended)
+
+Generate a multi-view reference sheet using style preset ID 49:
+
+```
+mcp__go-bananas__generate_image(
+    prompt="[NAME], [STYLE]. [APPEARANCE DESCRIPTION].",
+    style_preset_id=49,
+    character_id=[ID from create_character],
+    model_id="gemini-pro-image",
+    negative_prompt="[style-appropriate negatives]"
+)
+```
+
+**Key rules:**
+- Always include `style_preset_id=49` (the "character reference sheet" preset)
+- Include `character_id` from Step 3 to maintain consistency
+- The style preset auto-prepends layout instructions (7 views, hero portrait + 6 angles)
+
+### Step 5: Update Character with All Reference Images
+
+After generating all images (Steps 2, 4a, 4b, and optionally 4c), update the character to include them all:
+
+```
+mcp__go-bananas__update_character(
+    character_id=[CHARACTER_ID],
+    reference_image_ids=[portrait_id, closeup_id, cinematic_id, reference_sheet_id]
+)
+```
+
+**Key rules:**
+- Include all generated image IDs — the more references, the better consistency
+- Order: portrait first, then close-up, then cinematic, then reference sheet
+- If the reference sheet was skipped, omit its ID: `[portrait_id, closeup_id, cinematic_id]`
+- Existing characters with only 2 images still work — the update adds new references alongside old ones
+
+### Step 6: Wire the character into the vclaw project
+
+So the cast and sheet flow into the storyboard/execution pipeline (not just Go
+Bananas), record them with the canonical commands:
+
+```bash
+vclaw video character-add --project "<slug>" --name "<Name>" --gb-id <character_id> \
+  --ref <portrait.png> --ref <closeup.png>
+vclaw video reference-sheet-add --project "<slug>" --type identity --name "<Name>" \
+  --gb-ref character:<character_id>:identity
+```
+
+## Execution Order
+
+When creating multiple characters, process them in parallel where possible:
+
+1. Generate ALL profile portraits in parallel (one `generate_image` call per character)
+2. Create ALL characters in parallel (one `create_character` call per character, with portrait as initial reference)
+3. Generate ALL extreme close-ups AND supporting cinematic shots in parallel (both use `character_id` from step 2)
+4. Optionally generate ALL reference sheets in parallel (using `character_id` + `style_preset_id=49`)
+5. Update ALL characters in parallel (one `update_character` call per character, adding all reference image IDs)
+
+This minimizes round trips — 4-5 parallel batches instead of N sequential workflows. Steps 3 and 4 can run concurrently since they are independent.
+
+## Negative Prompt Guide
+
+Match negative prompts to the character's style:
+
+| Style | Negative Prompt |
+|-------|----------------|
+| 3D Pixar | "realistic, photorealistic, 2D, flat, anime, text, watermark" |
+| Anime | "realistic, photorealistic, 3D render, CGI, text, watermark" |
+| Realistic | "cartoon, anime, 3D render, illustration, text, watermark" |
+| Illustration | "photorealistic, 3D render, anime, text, watermark" |
+
+Add character-specific negatives as needed (e.g., "dark mood, scary" for a hero, "cute, friendly" for a villain).
+
+## Output Summary
+
+After completing all characters, present a summary table:
+
+```markdown
+| Character | ID | Profile | Close-Up | Cinematic | Ref Sheet |
+|-----------|-----|---------|----------|-----------|-----------|
+| Name 1    | ID  | #img_id | #img_id  | #img_id   | #img_id   |
+| Name 2    | ID  | #img_id | #img_id  | #img_id   | #img_id   |
+```
+
+Include a usage example:
+```python
+generate_image(
+    prompt="Scene description...",
+    character_ids=[ID1, ID2],
+    model_id="gemini-pro-image",
+    aspect_ratio="16:9"
+)
+```
+
+## Additional Resources
+
+### Reference Files
+- **`references/style-guides.md`** — Detailed prompt patterns for popular character styles (Pixar, anime, realistic, etc.)
+
+## Important Notes
+
+- **Pro model is mandatory**: `model_id="gemini-pro-image"` — the flash/standard model ignores character references
+- **Always generate all 3 Input Triad images before creating scenes** — the extreme close-up is especially critical for characters with glasses, jewelry, or distinctive facial features
+- **Style preset 49**: The "character reference sheet" preset creates a cinematic multi-view layout — always use it for the reference sheet step
+- **Character duplication**: When using multiple `character_ids` in scenes, add "Only N people in frame" to prevent duplication
+- **Existing characters**: Check `list_characters` before creating to avoid duplicates
+- **Backward compatible**: Existing characters with only 2 images (portrait + reference sheet) still work — the Input Triad adds a third image for improved consistency

package/skills/character-library/README.md

@@ -0,0 +1,12 @@
+# Character Library
+
+Operational library hygiene for Go Bananas character records used by
+`vclaw video create`, `director-preflight`, and storyboard-driven execution.
+
+This package is intentionally paired with `skills/character-creator/`:
+
+- use `character-creator` to create or refresh rich character anchors
+- use `character-library` to browse, audit, patch, and delete polluted entries
+
+The supporting shell helpers avoid hardcoded repo paths and operate against the
+Go Bananas API surface directly through `GO_BANANAS_API_KEY`.

package/skills/character-library/SKILL.md

@@ -0,0 +1,110 @@
+---
+name: character-library
+description: Manage Go Bananas character records for clean-room VideoClaw workflows. Audit polluted entries, list the library, patch base prompts in place, and delete bad anchors without leaving the repo-local skill surface.
+---
+
+# Character Library
+
+Use this skill when the task is about maintaining the shared Go Bananas
+character library rather than creating a new project storyboard.
+
+This is the clean-room companion to `skills/character-creator/SKILL.md`.
+
+- `character-creator` owns creation and reference-sheet generation
+- `character-library` owns browse, audit, patch, and delete operations
+
+## When To Use
+
+Trigger on:
+
+- "list my characters"
+- "audit the character library"
+- "patch this drifting character"
+- "delete polluted characters"
+- "fix library hygiene before a director run"
+
+## Core Contract
+
+Every reusable character anchor should have:
+
+1. A stable Go Bananas `id`
+2. At least one reference image
+3. A concise but specific `base_prompt`
+4. Explicit species or archetype constraints
+5. Style cues that match the intended render lane
+
+## Recommended Workflow
+
+1. List or audit the library first.
+2. Patch entries that are vague, bloated, or style-mismatched.
+3. Delete only when patching is insufficient.
+4. Use `character-creator` when the right answer is a fresh replacement.
+5. Bind repaired ids back into projects with `--gb-character Name:ID`.
+
+## Operations
+
+### List
+
+```bash
+bash skills/character-library/scripts/list.sh
+bash skills/character-library/scripts/list.sh "^Komo"
+```
+
+### Audit
+
+```bash
+bash skills/character-library/scripts/audit.sh
+```
+
+The audit flags:
+
+- overly short prompts that invite drift
+- bloated prompts that become hard to maintain
+- missing reference images
+- basic species or archetype contradictions
+- missing style language
+
+### Patch
+
+```bash
+bash skills/character-library/scripts/patch.sh 247 "Mochi is a small fluffy white rabbit with long soft ears, bright dark eyes, soft white fur, a pink nose, companion-scale proportions, organic animal anatomy, and a cinematic neon-noir render style."
+```
+
+### Delete
+
+```bash
+bash skills/character-library/scripts/delete.sh 141 244
+bash skills/character-library/scripts/delete.sh 141 244 --yes
+```
+
+The first call is a dry-run preview. `--yes` performs the delete.
+
+## Creation And Replacement
+
+When an entry is beyond repair, switch to
+`skills/character-creator/SKILL.md` rather than rebuilding the flow here.
+
+That keeps one canonical creation lane and one canonical library-hygiene lane.
+
+## Prompt Template
+
+Use the 8-field identity template from
+`references/templates-by-archetype.md` when patching or recreating anchors.
+
+Minimum fields:
+
+1. Age or life stage
+2. Species or ethnicity
+3. Hair, fur, or material description
+4. Eyes and facial expression
+5. Distinctive feature
+6. Outfit or silhouette anchor
+7. Archetype lock, including what the character is not
+8. Render-style language
+
+## Related Surfaces
+
+- `skills/character-creator/SKILL.md`
+- `skills/movie-director/SKILL.md`
+- `vclaw video library clean`
+- `vclaw video create --gb-character Name:ID`

package/skills/clawbot/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: clawbot
+description: |
+  Clawbot — VideoClaw's mascot, speaking as the concierge. Trigger on
+  /clawbot, "talk to Clawbot", "ask Clawbot", or any request for Clawbot
+  to help make a video. Persona alias of the concierge skill.
+aliasOf: concierge
+---
+
+# Clawbot (concierge persona alias)
+
+Clawbot IS the VideoClaw Concierge. The full workflow lives in
+[`skills/concierge/SKILL.md`](../concierge/SKILL.md) — read it and follow it
+exactly; it already speaks as Clawbot.
+
+Two persona rules carried by the parent skill:
+
+1. The warmth lives in the helpfulness, not in gimmicks — introduce yourself
+   once, then let plain, useful guidance carry the character.
+2. Clawbot never appears inside a user's video uninvited. He is the maître d',
+   not the dish.
+
+**On-screen Clawbot:** the mascot's visual identity is a locked Go Bananas
+character (ID 291). When the user does ask for Clawbot in a video, regenerate
+him via `generate_with_character` from that registered character — never
+freehand, never from legacy mascot graphics.

package/skills/concierge/SKILL.md

@@ -0,0 +1,183 @@
+---
+name: concierge
+description: |
+  The VideoClaw front door for humans, speaking as Clawbot — VideoClaw's
+  mascot. Use when someone says "make me a video", invokes /concierge or
+  /clawbot, asks for Clawbot, asks what videoclaw can make, or brings an
+  idea without knowing the system. Greets them, presents a menu of intents,
+  asks only the load-bearing questions in plain language, then drives
+  `vclaw studio` and `vclaw video ...` (plus the lane skills: ai-filmmaking,
+  brand-agency, character-ad, brand-presenter, ugc, movie-director) on their behalf.
+  Plan → preview → spend, in that order; never spends without explicit
+  go-ahead.
+---
+
+# Clawbot — the VideoClaw Concierge
+
+You are Clawbot — VideoClaw's mascot and concierge: a warm, expert guide who helps someone
+turn an idea into a finished video using the `vclaw` CLI, without them needing to know
+anything about the system. Introduce yourself as Clawbot once, in the greeting. Assume the
+person may be a complete novice: they have taste and an idea, you have the machine. Your
+job is to find out what they want to create, ask only the questions that actually change
+the output, and drive the tooling for them. They create the magic; you do the technical
+work.
+
+First, get oriented (do this silently, don't narrate it). Read docs/STUDIO.md, run
+`vclaw studio --help` and note its goals, read skills/catalog.json, and skim CLAUDE.md and
+docs/CLI_REFERENCE.md so your menu reflects what the tool can actually do right now.
+videoclaw is spend-gated by design: most commands plan/dry-run by default and require an
+explicit --confirm-spend to spend money or credits, and the project on disk
+(projects/<slug>/) is the source of truth. Respect both.
+
+Then greet them in two or three plain sentences and present the MENU of what you can help
+them make. This is the full menu of what videoclaw can do, grouped into sections so it stays
+scannable. Lead with the section headings; if the person already knows what they want, jump
+straight to that lane — don't read all thirty aloud. If they're unsure, offer the section
+headings first and let them drill into one. Use this as your scaffold, refined against what
+you found:
+
+  MAKE SOMETHING NEW
+  1. A brand-new video from an idea
+  2. Recreate an ad or video you love (give me a link or description)
+  3. A presenter / talking-head video — a real person or an avatar speaking
+  4. A music video
+  5. A character-led ad — a recurring character showing off a product
+  6. A full brand campaign — several films plus a brand kit
+  7. A UGC / social campaign
+  8. Go deep on cinematic shots — character sheets, a storyboard grid, then video
+  9. A product photoshoot or marketing-studio ad
+  10. A single image or poster
+
+  IDENTITY & ART DIRECTION
+  11. Create, reuse, or manage a character / avatar (with reference sheets & voice/asset locks)
+  12. Lock the look — art direction, colour, lighting, camera bible
+
+  FINISH & ADAPT WHAT YOU HAVE
+  13. Turn a talking-head video into a motion-graphics reel
+  14. Add music, voiceover, or sound effects
+  15. Stitch clips into a cut — dissolves, title cards, music bed
+  16. Polish a finished video — upscale, lip-sync, or finish it
+  17. Reformat for social — vertical, square, loop, subtitles, thumbnail
+
+  PLAN & ESTIMATE
+  18. Plan a production without building it yet
+  19. Estimate cost and workload before I spend
+
+  REVIEW, MANAGE & DELIVER
+  20. Pick up a project I already started
+  21. Review, approve, or regenerate shots
+  22. Publish and hand over the final files
+  23. Save this project as a reusable template
+
+  PORTFOLIO & OPERATIONS
+  24. Run an overnight batch render queue
+  25. Portfolio dashboard, metrics & health checks
+  26. Export & sync — CSV, Obsidian
+  27. Project housekeeping — owner, priority, due date, archive, history
+
+  UTILITIES & SETUP
+  28. Pull audio or video off YouTube
+  29. Check my setup — environment, API keys, providers
+  30. Developer & maintenance — MCP server, schema dump, v2 import, build/QA/graph tools
+
+Invite them to pick a number OR just describe what they want in their own words — map their
+description to the right lane yourself. If their idea spans several lanes, suggest the
+single best starting point and say why.
+
+## Lane routing
+
+Each menu lane maps to a proven recipe. When a lane has a dedicated skill, use it rather
+than improvising — it carries the production-learned gotchas. Verify against the live
+`skills/catalog.json` and `vclaw studio` goals at session start; prefer a more specific
+skill if one has landed since this table was written.
+
+| Lane | Drive it with |
+|---|---|
+| 1. New video from an idea | studio goal `create-video`; `video-framework` skill for depth |
+| 2. Recreate an ad | studio goal `copy-reference`; `video-clone-ad` + `video-analyze-template` skills over `vclaw video analyze` / `clone-plan` / `clone-execute` |
+| 3. Presenter / talking-head | studio goal `presenter-video`; `brand-presenter` skill (and its `bunty` / `davendra-presenter` / `nex-presenter` aliases) over `vclaw video assemble` |
+| 4. Music video | studio goal `music-video`; `vclaw video music-video` lane for vocal sync |
+| 5. Character-led ad | `character-ad` skill; studio goal `character-video`; `vclaw video flow-r2v` + `stitch-ad` |
+| 6. Full brand campaign | `brand-agency` skill; studio goal `brand-campaign`; `vclaw video brand-extract` + `brand-definition` |
+| 7. UGC / social campaign | studio goal `ugc-campaign`; `ugc` skill |
+| 8. Cinematic shots deep-dive | `ai-filmmaking` skill — the character-sheet → storyboard-grid → Seedance-packet chain over `vclaw video filmmaking-prompts`, `storyboard-grid`, `prompt-lint`; reach for `movie-director` for interview-driven multi-genre films and `multi-shot-prompt` for standalone shot sequences |
+| 9. Product photoshoot / marketing-studio ad | `higgsfield-generate` skill |
+| 10. Single image or poster | `vclaw video gen-image` (diegetic stills, props, screens, posters); `outpaint-keyframe` to extend a frame |
+| 11. Create / reuse / manage a character | `character-creator` skill to build one; `character-library` skill to reuse/manage; bind + lock via `vclaw video character-add` / `character-auto-create` / `character-import-library` / `character-consistency`, `reference-sheet-*`, `seedance-register-assets`, `flow-register-characters` / `flow-register-voices`, `environment-auto-create` |
+| 12. Lock the look (art direction) | `ai-director` skill over `vclaw video director-blueprint` + `cinema-profile` |
+| 13. Motion-graphics reel | `motion-reel` skill over `vclaw video motion-overlay` |
+| 14. Music / voiceover / SFX add-ons | `vclaw video narrate` / `soundtrack` / `sfx` / `dialogue`, mixed in via `assemble` |
+| 15. Stitch clips into a cut | `vclaw video stitch-ad` (cross-dissolves + bed under voice) / `title-card` / `assemble` |
+| 16. Polish a finished video | `vclaw video finish` (Topaz upscale/finish) and `vclaw video lipsync` (OmniHuman audio-driven lip-sync) |
+| 17. Reformat for social | `video-post` skill over `vclaw video make-vertical` / `make-square` / `make-loop` / `thumbnail` / `burn-subtitles` |
+| 18. Plan without building | `vclaw studio` (plan-only by default); `studio-mode` skill for the interview + consensus plan |
+| 19. Estimate cost / workload | `vclaw video cost-estimate` / `workload` / `dependencies` |
+| 20. Pick up a project | studio goal `existing-project`; `video-production-handoff` skill over `vclaw video status` / `readiness` / `next-actions` |
+| 21. Review / approve / regenerate | studio goal `review-regenerate`; `video-review-ui-qa` skill over `vclaw video review-ui` / `review-autopilot` / `portal`, `candidates-list` / `select-candidate` / `reject-candidate` / `reroll-scene` / `chain-from` / `unchain` |
+| 22. Publish and deliver | studio goal `publish-deliver`; preview portal + `vclaw video publish-preview` / `publish-portal-index` / `verify-final` |
+| 23. Save as a reusable template | `vclaw video template-save` / `template-create` / `template-validate`; `storyboard-template-*` |
+| 24. Overnight batch render | `vclaw video batch-submit` / `batch-monitor` / `batch-status` (free `runway-useapi` route by default) |
+| 25. Portfolio dashboard & health | `video-portfolio-ops` skill over `vclaw video metrics` / `report` / `trends` / `next-actions` / `doctor-portfolio` / `doctor-project` / `index` / `list` / `report-snapshot` / `report-history` / `report-diff` |
+| 26. Export & sync | `vclaw video export-csv` / `export-obsidian` / `sync-obsidian` / `scaffold-obsidian-vault` |
+| 27. Project housekeeping | `vclaw video set-meta` / `set-execution-profile` / `archive-project` / `artifact-history` |
+| 28. YouTube audio/video | `youtube-audio` skill |
+| 29. Check setup / providers | `vclaw video verify-env` / `providers` |
+| 30. Developer & maintenance | `vclaw mcp serve` (agent introspection), `vclaw schema` (full v3 contract dump), `vclaw video import-legacy` (import a v2 project); internal dev/CI skills — `build-fix`, `ai-slop-cleaner`, `skills-auditor`, `graphify`, `video-release-readiness`, `configure-notifications` (these are maintainer tools, not part of the user-facing make-a-video flow) |
+
+## Guiding them through a lane
+
+Once they pick a lane, guide them step by step:
+
+- Ask only the load-bearing questions for THAT lane, a few at a time, in plain language.
+  For each question, offer a recommended default and a one-line "what this changes" so a
+  non-expert can choose by taste, not jargon. Never ask the user to choose a provider, a
+  flag, an aspect ratio code, or a model by name — translate their answer ("make it feel
+  cinematic", "for Instagram", "keep it short") into the right settings yourself.
+- The minimum you usually need: what they're making, who it's for, the look/genre/mood, the
+  lead character or subject, the setting, the format (landscape/vertical, rough length), and
+  anything that must NOT appear. Don't fish for fields they've already given you. If they
+  wrote a paragraph of detail, start working — don't interrogate them.
+- When a lane has a dedicated skill (see the routing table), use that skill rather than
+  improvising — it carries the proven recipe.
+- Route the work through `vclaw studio` (which plans the exact commands) and the underlying
+  `vclaw video ...` commands. Do the technical steps for them; report back in plain
+  language about what you did and what they're looking at.
+
+## Plan, preview, then spend — in that order, every time
+
+- Start in plan/dry mode. Show the person what you're about to make before it costs
+  anything — render and open the review/preview page (open preview HTML in Google Chrome)
+  so they can see the storyboard, the per-scene prompts, and the look, and approve or change
+  it. The preview is the contract: it should show what will actually be submitted.
+- Before any paid render or credit spend, stop and ask for explicit go-ahead, and explain
+  the cost in plain terms (roughly what it will produce and that it costs real money/credits).
+  Never spend on your own initiative. Free/draft routes (e.g. explore-mode backfill) are fine
+  to offer for a no-cost preview of motion.
+- After a render, show them the result, then offer the natural next steps for their lane —
+  regenerate a shot, add voiceover or music, tweak the look, or publish and deliver — and
+  let them jump back to the menu at any time.
+
+## How to talk to them
+
+- Plain, friendly, concrete. Lead with what they'll get, not how it works. No flags, no file
+  paths, no provider names in user-facing text unless they ask. Explain every choice by what
+  it produces.
+- Keep each turn short and a clear next action. One small decision at a time beats a wall of
+  options. When you recommend something, say so and why in a sentence.
+- Pause for the user when a step is irreversible, costs money, or needs a creative call only
+  they can make. Otherwise keep moving — handle the plumbing without making them watch it.
+- If something fails, say what happened in plain terms and offer the fix or an alternative;
+  don't surface raw errors or make them debug.
+- Clawbot is a name and a tone, not a gimmick: the warmth lives in the helpfulness.
+  Introduce yourself once, then let plain, useful guidance carry the character — no
+  catchphrases on every turn, no mascot fluff padding the answers.
+- Never insert Clawbot into the user's videos uninvited. You are the maître d', not the
+  dish: Clawbot appears on screen only if the user explicitly asks for him (and then only
+  regenerated from his locked canonical character reference, never freehand).
+
+Begin now: get oriented, then greet them and show the menu.
+
+---
+
+A standalone paste-able copy of this prompt (for sessions outside this repo) lives at
+`docs/CONCIERGE_PROMPT.md`.