@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/pixel2motion/references/ribbon-fitting.md

@@ -1,90 +0,0 @@
-# Closed / Self-Intersecting Ribbon Fitting — Centerline Scaffold + Recenter + Snap
-
-Field-tested on a calligraphic ∞ mark (variable width 2→28px, one self-crossing,
-one occluding bead). Use this recipe when the mark is a **closed** variable-width
-ribbon and/or **crosses itself** — there are no two open boundaries to fit
-directly, so the playbook's "fit boundaries from source pixels" rule needs this
-adapted form. The principle survives: the centerline is only *scaffolding*; the
-final edges are snapped to SOURCE pixels.
-
-Implementation: `scripts/fit_ribbon_centerline.py` (seeds + exclusions JSON in,
-outline/centerline path d + report out).
-
-## The recipe
-
-1. **Seed a closed Catmull-Rom centerline** through hand-read keypoints in
-   drawing order: loop extrema, caps, the crossing (listed once per pass), the
-   taper. 15–25 points suffice; accuracy ±5px is fine — recentering fixes it.
-   Put the START at a hidden spot (behind an occluding dot, inside a knot):
-   both outline caps land there.
-
-2. **Measure along normals; auto-recenter.** Sample the spline densely (~30/seg).
-   At each sample march ±along the normal: the mask interval containing the
-   point yields a *midpoint* (true center) and a *width*. Then rebuild the
-   control points from the smoothed midpoints and repeat. Converges in 2–3
-   passes (mean shift 1.6 → 0.3px in the field run).
-   - **Space rebuilt control points by ARCLENGTH** (~55px apart). Rebuilding by
-     sample stride inflates the control count every pass (22 → 56 in one run)
-     because each pass resamples denser.
-   - Reject measurements wider than a contamination bound (~40px): they mean
-     the normal crossed into a merged/parallel stroke.
-
-3. **Exclusion circles bridge occlusions and crossings.** Inside each exclusion
-   (covering dots/beads, every self-crossing) skip measurement entirely and
-   fill midpoints/widths by interpolation along arclength — **circularly**,
-   the gap may wrap the seam. Smooth lightly afterwards: σ≈2.5 samples for
-   midpoints, σ≈3 for widths. Heavier smoothing curve-cuts high-curvature caps
-   (~1px inward bias) and flattens fast width ramps near caps.
-
-4. **Sub-pixel edge snap (the honesty pass).** Offset edges = center ± width/2
-   · normal still carry ~1px of systematic smoothing bias. For each edge sample,
-   find the luminance-threshold crossing along the normal by bilinear
-   interpolation and shift the sample onto it (clamp ±2.5px; skip inside
-   exclusions and where width < 6px — on hairline tapers the two edges are
-   closer than the search radius and snap to each other; smooth the shifts,
-   σ≈2). This pass is what reconciles the centerline-scaffold approach with
-   the "boundaries from source pixels" rule.
-
-5. **Emit smooth cubics.** Fit one cubic per ~30-sample chunk per edge
-   (endpoints fixed, tangents from data, least-squares handle lengths — two
-   scalars, closed form). Path = left edge forward + straight cut + right edge
-   reversed + closing cut + `Z`, `fill-rule: nonzero` (a self-crossing stroke
-   outline traversed left-forward/right-back winds consistently, so nonzero
-   unions the overlap — no holes). Both cut caps sit at the hidden seam.
-
-6. **Report exclusion arc-fractions.** For every exclusion circle, record where
-   the centerline passes through it as fractions of total arclength. A
-   crossing's two passes (e.g. 0.357 / 0.835) are exactly the cut parameters a
-   downstream split draw-on animation needs (see pixel2motion's reveal-patterns
-   "Draw-On Across Self-Intersections").
-
-## Pitfalls measured in the field
-
-| pitfall | symptom | fix |
-|---|---|---|
-| stride-based control rebuild | control count inflates every recenter pass | space by arclength |
-| heavy width smoothing (σ≈6) | caps/fast ramps cut ~1.5px inward | σ≈3 + edge snap |
-| snapping hairline tapers | edge snaps onto the *opposite* edge | skip snap where width < 6px |
-| linear (non-circular) gap fill | seam-adjacent exclusion interpolates across the whole loop | circular interpolation |
-| trusting the dot-region width | occluded; garbage measurements | exclusion circle + bridge |
-| gradient-filled parts in IoU | binary-mask diff inflates while visually faithful | judge structurally; report the artifact |
-
-## Wordmark font matching (live `<text>`)
-
-When the lockup includes a wordmark and exact letterforms are not required:
-
-1. **Enumerate fonts actually installed** (`ls /System/Library/Fonts*`,
-   `fc-list`). Never trust name-based stacks: two "different" families scoring
-   *identical* region IoU means both fell back to the same font — a missing
-   "Garamond" can silently resolve while a real one never existed.
-2. Per candidate, **auto-tune `font-size` to the source cap height** (2–3
-   render-measure-scale iterations against the wordmark bbox), baseline pinned.
-3. Rank by wordmark-region IoU **and ink-weight ratio** (rendered ink px ÷
-   source ink px). Weight ratio ≈ 1.0 beats marginally-higher IoU with 1.2×
-   heavier strokes: weight mismatch reads instantly, letterform differences
-   read slowly. (Field run: Baskerville 1.04 ratio / IoU 0.535 chosen over
-   Times 1.26 / 0.540 — visibly better.)
-4. Lock geometry cross-platform: `textLength="<measured width>"
-   lengthAdjust="spacingAndGlyphs"`, `text-anchor` + measured baseline; stack
-   best-first with graceful fallbacks; document the substitution as a known
-   residual.

package/skills/pixel2motion/references/twelve-principles-for-logos.md

@@ -1,116 +0,0 @@
-# The 12 Principles, Applied to Logo Motion
-
-Distilled from Disney's 12 Principles of Animation (Johnston & Thomas, *The Illusion of Life*) for the specific job of animating a vectorized logo. Each entry: what the principle means for a logo, concrete parameter ranges, and the failure mode to avoid.
-
-## 1. Squash & Stretch
-
-Deformation shows weight and life. For logos, deformation is **personality-gated**: playful brands may squash 10–25%; professional brands use plain scale transforms (0%) instead. Scale deformation proportionally to duration — ~5% at 100ms, up to 30–40% only in deliberate 1200ms+ brand moments.
-
-- Dots landing: squash to `scaleY(0.8) scaleX(1.15)` on impact, recover over 30% of the bounce duration.
-- Volume preservation: if X stretches, Y compresses. `scale(1.2, 0.85)` not `scale(1.2, 1)`.
-- **Failure mode**: squashing a luxury/fintech mark. If the brand words are "precise / trustworthy / premium", skip deformation entirely.
-
-## 2. Anticipation
-
-Prepare the eye before the main action. The wind-up is **20–30% of the main action's duration** at any time scale: a 750ms assembly gets a 150–220ms anticipation; a deliberate splash can take a full 300–400ms dramatic pause or counter-movement.
-
-- Patterns: brief dim/scale-down before the pop (`scale(0.96)` → action), a dot pulling back before sliding into place, a 1–2px counter-translate opposite the travel direction, a beat of empty canvas before the first stroke draws.
-- **Failure mode**: elements just *starting*. Motion with no anticipation reads as a glitch, not an entrance.
-
-## 3. Staging
-
-One clear focus at a time. The logo reveal is a tiny film: establish the mark, then the wordmark, then the tagline — never all at once.
-
-- Choreograph a **reading order** and verify it in the frame strip: where does the eye go at t=300? t=800?
-- Background/secondary parts must be quieter (smaller amplitude, lower opacity delta) than the focal part.
-- Negative space is staging too: a mask wipe that reveals the counter-form of a mark stages the silhouette itself.
-- **Failure mode**: simultaneous equal-amplitude motion everywhere — the eye has nowhere to land.
-
-## 4. Straight Ahead vs Pose to Pose
-
-Logo motion is **pose-to-pose by design**: define key poses (hidden → anticipation pose → peak → overshoot → final), then let easing interpolate. A deliberate 1200–2000ms sequence wants 6–10 key poses; a 300ms header reveal wants 2–3.
-
-- The final pose is non-negotiable: it is the QA-verified static vector (the Final Frame Contract).
-- Straight-ahead thinking only suits continuous idle loops (drift, shimmer) where there is no destination.
-- **Failure mode**: keyframes placed by trial-and-error percentages with no named pose behind each one.
-
-## 5. Follow Through & Overlapping Action
-
-Nothing stops all at once. The single highest-leverage principle for logo builds.
-
-- **Drag hierarchy**: root (container/main mark) moves and lands first → primary parts (letters) → secondary (dots, accents) → tertiary (sparkle, underline). Offset each level by 2–4 frames (~35–70ms at 60fps).
-- Letters in a wordmark arrive with **staggered settling**: stagger delay = 10–20% of one letter's duration (300ms letters → 30–60ms stagger). Cap total cascade ≈ 500–700ms; with many letters, shrink the per-letter stagger rather than stretching the total.
-- Follow-through occupies ~30% of the timeline: overshoot the final pose slightly (`scale(1.02)`, `translateY(-2%)`), then settle with ease-out. Fast stops earn longer follow-through; slow arrivals settle subtly.
-- **Failure modes**: everything stopping on the same frame (robotic); excessive overshoot (disconnected, floaty); all parts sharing one delay (no hierarchy).
-
-## 6. Slow In / Slow Out
-
-Easing is mandatory; linear motion looks mechanical at every duration (the only exception: continuous rotation in a spinner-style loop, which must be `linear` to avoid pulsing).
-
-- Entrances: ease-out (`cubic-bezier(0, 0, 0.2, 1)` or the stronger `cubic-bezier(0.16, 1, 0.3, 1)`).
-- Duration < 150ms: ease-out only — there is no time for an ease-in.
-- 150–400ms: ease-out or custom; 400ms+: full ease-in-out becomes available.
-- Deliberate narrative arcs: `cubic-bezier(0.34, 0, 0.14, 1)` (sharp interest, soft resolution).
-- **Failure mode**: default `ease` everywhere — it's an opinion, not a decision.
-
-## 7. Arcs
-
-Natural motion curves. A dot flying to its position should travel a subtle arc, not a straight line — combine X and Y keyframes with different easing, or use `offset-path`.
-
-- Short reveals: subtle curvature (a few px of perpendicular bow). Deliberate sequences: designed sweeping trajectories that echo the logo's own curves (a swoosh's dot should arrive along the swoosh's implied arc).
-- Rotational entries (orbit-in) are arcs by construction; mind the transform-origin.
-- **Failure mode**: diagonal straight-line travel — reads as engineered, not alive (fine only if the brand is deliberately mechanical).
-
-## 8. Secondary Action
-
-Supporting motion that reinforces without competing: a soft glow blooming as the mark lands, an underline drawing after the wordmark settles, a 1–2px shadow tightening on touchdown.
-
-- Secondary action obeys staging: lower amplitude, often after the primary beat, never overlapping the focal moment.
-- Use sparingly and personality-gated: playful brands can afford a sparkle; minimal brands get at most a fade or shadow shift.
-- **Failure mode**: confetti on a B2B logo; or secondary action firing *during* the primary action and splitting attention.
-
-## 9. Timing
-
-Duration is the message. Pick the band from usage context, then refine per part:
-
-| Context | Band | Notes |
-|---|---|---|
-| Hover / micro-feedback | 150–300ms | ease-out only |
-| Header / in-page reveal | 300–800ms | standard reveal, stagger ≤ 3 beats |
-| Splash / brand intro | 1200–2000ms | full narrative: anticipation–action–settle |
-| Loading / idle loop | 1500–4000ms cycle | seamless, hypnotic not annoying |
-| Error / attention shake | ~400ms, 1 cycle | rare for logos; only on interaction |
-
-Universal laws: shorter distances → shorter durations; bigger/heavier parts move slower; enter fast, exit faster; first-visit tolerance is higher than repeat-visit (a 2000ms splash that plays on every route change is a bug, not a brand moment).
-
-- **Failure mode**: one duration for every part regardless of size or travel distance.
-
-## 10. Exaggeration
-
-Push past reality for clarity — within personality bounds. Exaggeration budget scales with duration and brand energy: subtle polish ≈ scale from 0.95; dramatic ≈ from 0.6–0.8 with blur; playful overshoot ≈ `cubic-bezier(0.34, 1.56, 0.64, 1)`; elegant brands cap overshoot at ~2%.
-
-- Exaggerate the *important* beat (the mark's landing), not everything.
-- **Failure mode**: uniform exaggeration — when everything is dramatic, nothing is.
-
-## 11. Solid Drawing
-
-Volume, perspective, and proportion stay consistent while moving. For vector logos this means: transforms must not distort the verified geometry (uniform scale unless squash is intentional and volume-preserving), parts must not visually detach from the group during motion, strokes keep their width profile (`vector-effect="non-scaling-stroke"` when scaling stroked parts), and nothing clips the viewBox mid-flight.
-
-- Check the mid-action frames: is the logo still *the logo* at t=50%?
-- **Failure mode**: a swoosh that thins as it scales; letters whose baselines drift during the cascade.
-
-## 12. Appeal
-
-The sum test: would you watch it twice? The animation should feel like the brand wearing motion, not motion wearing a logo. Clean = appealing: fewer, better-chosen movements beat many simultaneous effects. Replay it; show it without context — is it recognizably *this* brand? Could a competitor ship the identical animation? If yes, the choreography is generic — revisit the personality mapping.
-
-- **Failure mode**: technically flawless, emotionally anonymous.
-
----
-
-## Quick interaction matrix
-
-- Timing defines follow-through length (fast stop → long settle).
-- Arcs govern follow-through paths (settling swings, not slides).
-- Anticipation ratio is duration-invariant (20–30% of action).
-- Squash/stretch reappears at settling (slight compression on landing).
-- Staging decides which part gets exaggeration; appeal judges the whole.

package/skills/premiere-batch-delivery/SKILL.md

@@ -1,71 +0,0 @@
----
-name: premiere-batch-delivery
-description: "Validate, export, and inspect multiple Adobe Premiere Pro deliverables with explicit pending and unverified states. Use for YouTube, social, web, mobile, archive, audio-only, versioned export, delivery QA, caption validation, direct export, AME submission, or repeatable batch output through the Premiere Pro MCP server. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "ayushozha/AdobePremiereProMCP"
----
-
-# Premiere Batch Delivery
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-Define a delivery matrix first. Treat a queued job as pending, not as a successful export.
-
-## Required server profile
-
-Start or restart the MCP server with `MCP_TOOL_PROFILE=delivery` to load the focused export, caption-check, and media-probe surface used by this skill.
-
-## Workflow
-
-1. List every target with its unique sequence index and ID, aspect ratio, resolution, frame rate, codec/preset, audio spec, captions, destination, and filename. Resolve the index/ID pair from a fresh `premiere_get_sequence_list` response; do not select by name alone.
-2. Check project integrity, delivery specs, captions, and frame accuracy. Before every active-sequence-only caption check or sidecar export, call `premiere_set_active_sequence` for that target index, require `verified: true` with the exact target ID, and re-read `premiere_get_sequence_list`; stop unless `activeSequenceID` still matches. Pass the same ID as `sequence_id` to `premiere_export_captions`. Treat host-only loudness or color summaries as advisory unless a decoded-media analyzer or visual review verifies them. Stop on unresolved errors.
-3. Call `premiere_list_export_presets_disk` to find real `.epr` files in known Premiere/AME directories, or use an explicitly supplied preset path. Confirm destination directories and collision policy. Create versioned filenames; never overwrite unless explicitly authorized.
-4. Prefer one-at-a-time `premiere_export_direct` calls with a confirmed preset; the call blocks, which bounds concurrency and makes failures attributable. Use `premiere_export_via_ame` only when asynchronous queueing is required, and record its returned job identifier.
-5. The CEP bridge cannot poll individual AME jobs to a trustworthy terminal state. Keep AME jobs pending until an external filesystem check observes a stable, non-empty output; if that check is unavailable, leave the job unverified rather than guessing.
-6. Call `premiere_probe_media` for every completed output and compare duration, dimensions, frame rate, and audio/video streams with the delivery matrix. Produce a concise delivery report.
-
-## Guardrails
-
-- Do not equate `queued` or `started` with `completed`.
-- Do not use the legacy platform compound exporters or `premiere_batch_export_sequences` as proof of completion; their CEP responses do not provide per-output completion readback.
-- Keep archive masters separate from distribution encodes.
-- Validate sidecar caption paths and names alongside the video deliverable.
-- Never reuse a prior active-sequence assumption across delivery-matrix rows; activate and verify each target immediately before its caption sidecar.
-- Report successful, failed, skipped, and unverified outputs separately.
-
-## Source
-
-Rewritten for AICUT from [`ayushozha/AdobePremiereProMCP`](https://github.com/ayushozha/AdobePremiereProMCP)
-(MIT), skills/premiere-batch-delivery/SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 7 skills; this is the one its entry stands for.

package/skills/premiere-pro-mcp/SKILL.md

@@ -1,161 +0,0 @@
----
-name: premiere-pro-mcp
-description: "Install, verify, troubleshoot, and operate the Adobe Premiere Pro MCP server. Use when a user wants an agent to set up Premiere MCP, connect Claude Code/Codex/Claude Desktop, control Premiere, import media, build sequences, edit timelines, apply effects, or diagnose bridge issues. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "hetpatel-11/Adobe_Premiere_Pro_MCP"
----
-
-# Adobe Premiere Pro MCP
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-Use this skill when working with the Adobe Premiere Pro MCP server from `hetpatel-11/Adobe_Premiere_Pro_MCP`.
-
-## Core Rules
-
-- Use the MCP tools for Premiere operations; do not invent ExtendScript unless the MCP tool surface is missing the needed operation.
-- Default MCP `tools/list` is a small always-on set. Call `search_tools` (BM25 query or regex pattern), then `invoke_tool` with the exact name. `PREMIERE_MCP_TOOLSET=full` lists every tool.
-- Prefer read-only discovery first: `get_project_info`, `list_sequences`, `list_project_items`, `get_active_sequence`, and relevant resource reads.
-- Use real imported media. If the user asks to edit with assets, verify file paths exist, import them with `import_media`, then place the imported project item IDs on a sequence.
-- Keep the temp directory consistent across the MCP server and CEP panel: `/tmp/premiere-mcp-bridge` unless the user explicitly configured another path.
-- Ask before destructive or externally visible actions: deleting clips/media, overwriting exports, closing projects, saving over important project files, or sending files elsewhere.
-- For generated/demo edits, prefer creating a new clearly named sequence instead of modifying the user's active sequence.
-- If a tool returns `success: false`, report the exact error and run diagnostics before retrying blindly.
-
-## Install Workflow
-
-If the user asks you to install or set up the MCP:
-
-1. Check the OS. The automated installer is macOS-focused.
-2. Clone or open the repo:
-
-```bash
-git clone https://github.com/hetpatel-11/Adobe_Premiere_Pro_MCP.git
-cd Adobe_Premiere_Pro_MCP
-```
-```
-3. On macOS, run:
-```
-```bash
-npm run setup:mac
-```
-```
-4. For non-macOS or manual client setup, run:
-```
-```bash
-npm run build
-```
-```
-5. Register the MCP server in the user's client with:
-```
-```text
-command: node /absolute/path/to/Adobe_Premiere_Pro_MCP/dist/index.js
-env: PREMIERE_TEMP_DIR=/tmp/premiere-mcp-bridge
-```
-For Codex, prefer:
-```
-```bash
-codex mcp add premiere_pro --env PREMIERE_TEMP_DIR=/tmp/premiere-mcp-bridge -- node /absolute/path/to/Adobe_Premiere_Pro_MCP/dist/index.js
-```
-
-## Premiere Bridge Startup
-
-After installing:
-
-1. Restart the MCP client if it reads config only at startup.
-2. Restart Premiere Pro.
-3. Open `Window > Extensions > MCP Bridge (CEP)`.
-4. Set `Temp Directory` to `/tmp/premiere-mcp-bridge`.
-5. Click `Save Configuration`.
-6. Click `Start Bridge`.
-7. Confirm the bridge panel says Premiere is ready before running editing tools.
-
-If Premiere is not running, you can install/build/register the MCP, but tell the user live tool verification requires Premiere and the CEP bridge panel.
-
-## Verification
-
-Run local checks:
-
-```bash
-npm run setup:doctor
-```
-```
-If Premiere is running and the bridge is started, verify with safe read-only calls:
-
-- `get_project_info`
-- `list_sequences`
-- `list_project_items`
-
-For deeper validation in a disposable project, create a test sequence with `create_sequence_from_clips` or `create_sequence` plus a real `.sqpreset` path, then call `list_sequences` and confirm it exists. Do not call blank sequence creation without a preset because newer Premiere versions can open a native dialog that blocks CEP.
-```
-
-## Editing Strategy
-
-- Start by understanding the project: project info, active sequence, existing media, tracks, markers, and selected sequence.
-- Build a plan in concrete Premiere operations before changing anything.
-- For rough cuts, import media first, then use `create_sequence_from_clips` so Premiere derives settings without a native dialog. Use `create_sequence` only with a real `.sqpreset`; use `duplicate_sequence` with `clearContents=true` when an existing sequence defines the intended settings.
-- For product or brand spots, prefer `assemble_product_spot` or `build_brand_spot_from_mogrt_and_assets` when the user's request fits those workflows.
-- For black-and-white looks, use `apply_effect` with `Black & White` rather than generic saturation-only changes.
-- For timeline cuts, prefer sequence-aware tools and include `sequenceId` when available.
-- Export only after confirming output path, format/preset, and overwrite behavior.
-
-## Troubleshooting
-
-If commands time out or report bridge errors:
-
-1. Confirm Premiere is open.
-2. Confirm `Window > Extensions > MCP Bridge (CEP)` is open and bridge is started.
-3. Confirm both sides use the same temp directory.
-4. Run:
-
-```bash
-npm run setup:doctor
-```
-```
-5. Ask the user to click `Run Diagnostics` in the CEP panel.
-6. Read `/tmp/premiere-mcp-bridge/premiere-mcp-diagnostics-latest.json` if it exists.
-7. Remove stale command/response files only if they are clearly old and the bridge is stopped or idle.
-
-Common fixes:
-
-- `ENOENT` on temp directory: create `/tmp/premiere-mcp-bridge`, save config again, restart bridge.
-- Tool succeeds in Premiere but reports failure: run `list_sequences` or the relevant list tool to confirm state before retrying.
-- Empty or malformed temp directory config: set the field to the path only, not JSON or an env assignment.
-```
-
-## Source
-
-Rewritten for AICUT from [`hetpatel-11/Adobe_Premiere_Pro_MCP`](https://github.com/hetpatel-11/Adobe_Premiere_Pro_MCP)
-(MIT), skills/premiere-pro-mcp/SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 3 skills; this is the one its entry stands for.

package/skills/promo-creator-skills/SKILL.md

@@ -1,209 +0,0 @@
----
-name: promo-creator-skills
-description: "产品宣传片制作总控 skill pack。用于从产品说明、官网、应用截图或 GitHub 仓库制作 60-90 秒宣传视频，按阶段完成 brief、storyboard、素材、AICUT 剪辑、BGM 设计和交付。当用户要做宣传片、产品视频、项目介绍视频、launch video、开源项目 promo、BGM 卡点或真实软件界面宣传片时使用。 Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "kangarooking/promo-creator-skills"
----
-
-# Promo Creator Skills
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-这是产品宣传片创作的总控入口。它必须能独立指导完整项目；子 skill 是阶段深水区，不是唯一的流程来源。
-
-默认目标：用真实产品信息、真实软件界面和可验证素材，制作一支 60-90 秒、可发布的产品宣传片 MP4。不要把它当成氛围视频、图片轮播或泛科技动效练习。
-
-## 总控判断
-
-当用户只说“做宣传片 / 产品视频 / launch video / promo”时，默认走完整流程：
-
-```text
-产品信息或 URL
-  -> 01-brief.md
-  -> 02-storyboard.md
-  -> 03-asset-plan.md + assets/
-  -> 04-edl.md + master-edit.html
-  -> 06-music-plan.md + assets/bgm/
-  -> final/promo.mp4 或 final/promo-with-bgm.mp4
-  -> 05-delivery.md
-```
-当用户明确只要某一段能力时，只执行对应阶段。例如“只要 BGM prompt”只用音乐阶段；“我已经有素材，直接剪辑”从 EDL 和 HyperFrames 阶段开始。
-```
-
-## 子 Skill 路由
-
-优先按总控流程推进；进入某个阶段时，读取对应子 skill 的 `SKILL.md` 获取详细规范。
-
-| 用户目标 | 详细规范 |
-|---|---|
-| 从零做完整宣传片 | `promo-workflow/SKILL.md` |
-| 只做产品分析和创意简报 | `promo-brief/SKILL.md` |
-| 已有 brief，要写分镜 | `promo-storyboard/SKILL.md` |
-| 已有分镜，要准备素材 | `promo-asset-producer/SKILL.md` |
-| 已有素材，要剪辑渲染 | `promo-editor/SKILL.md` |
-| 要配乐、BGM、卡点音乐 prompt | `promo-music-maker/SKILL.md` |
-
-## 阶段流程
-
-### Phase 1: Brief
-
-详细规范见 `promo-brief/SKILL.md`。输入可以是产品说明、官网、GitHub 仓库、应用截图、用户口述或竞品链接。
-
-产出 `01-brief.md`：
-- 产品定位和目标观众
-- 核心卖点和证明点
-- 推荐视觉风格
-- 叙事结构和时长建议
-- 风险、素材缺口和需要用户确认的问题
-
-暂停点：确认定位、风格、时长、主叙事。
-
-### Phase 2: Storyboard
-
-详细规范见 `promo-storyboard/SKILL.md`。不要在 brief 未确认前直接写完整剪辑。
-
-产出 `02-storyboard.md`：
-- 逐镜头时间码
-- 每个镜头的画面、文案、动效、转场、素材需求
-- 真实 UI / 真实产品信号应该出现在哪些镜头
-- HyperFrames 实现提示
-
-暂停点：确认镜头内容、节奏、文案密度和转场方向。
-
-### Phase 3: Assets
-
-详细规范见 `promo-asset-producer/SKILL.md`。优先真实产品素材，其次生成补充视觉。
-
-产出：
-- `03-asset-plan.md`
-- `assets/pack-a/`：AI 生成或补充视觉
-- `assets/pack-b/`：官网、截图、开源项目、文档或其他可验证素材
-- `assets/pack-b/pack-b-sources.md`：来源记录
-
-暂停点：确认素材质量和版权/来源可接受性。
-
-### Phase 4: Edit
-
-详细规范见 `promo-editor/SKILL.md`。用 HyperFrames/HTML/CSS/GSAP 构建可渲染视频，不只给剪辑建议。
-
-产出：
-- `04-edl.md`
-- `master-edit.html`
-- `final/promo.mp4`
-
-暂停点：确认成片结构、画面密度、节奏和需要返修的镜头。
-
-### Phase 5: Music
-
-详细规范见 `promo-music-maker/SKILL.md`。音乐必须服务剪辑节奏，不要默认生成 ambient 科技垫底音乐。
-
-产出 `06-music-plan.md`：
-- 音乐风格、BPM、情绪曲线
-- 和镜头转场对齐的 hit points
-- Mureka / Skywork Music Maker 英文 prompt
-- negative prompt
-- 至少 2 个有明显差异的方向
-
-用户确认或要求实际生成时，再调用脚本生成 BGM。
-
-### Phase 6: Delivery
-
-补齐 `05-delivery.md`：
-- 最终视频路径
-- 源文件路径
-- 素材和 BGM 清单
-- 修改方式
-- 已知限制
-
-## 工作目录
-
-每个项目创建独立 run 目录：
-
-```text
-outputs/promo-runs/YYYY-MM-DD-<product>/
-```
-如果用户给了明确项目名，使用用户项目名；否则根据产品名生成短 slug。
-```
-
-## 暂停规则
-
-必须暂停的节点：
-- Brief 完成后
-- Storyboard 完成后
-- 素材计划或关键素材完成后
-- BGM 方向确认前实际生成音乐
-- 最终渲染后
-
-暂停时最多问 3 个问题，并给出推荐选项。用户明确说“按你推荐来 / 继续 / ok”时，按推荐方案推进，不要反复追问。
-
-## 执行约束
-
-- 先做产品判断，再做画面。
-- 先写分镜，再批量生成素材。
-- 先确认音乐方向，再消耗音乐生成额度。
-- 真实软件界面优先于抽象装饰图。
-- 所有外部素材记录来源。
-- 不要把多个根 HyperFrames composition 留在同一渲染目录。
-- 如果图片、浏览器、音乐或渲染能力不可用，输出可执行 prompt、素材清单或命令，并明确缺口。
-
-## BGM 默认策略
-
-产品宣传片默认先给两个差异明显的方向：
-
-| 方向 | 用途 |
-|---|---|
-| `Commercial Product Launch` | 产品发布、商业展示、CTA 更强的广告片 |
-| `Clean UI Demo Groove` | 真实软件界面、SaaS、流程展示、UI demo |
-
-候选之间必须在 genre、rhythm、bass、hook、energy curve 上形成明显差异。避免多个 prompt 都写成 `minimal / ambient / soft pulse / glass ticks`。
-
-详细音乐规范见 `promo-music-maker/SKILL.md` 和 `references/product_promo_bgm_prompting.md`。
-
-## 最小可执行流程
-
-当只能读取本文件时，也按下面步骤工作：
-
-1. 根据用户输入创建 run 目录。
-2. 写 `01-brief.md`，暂停确认。
-3. 写 `02-storyboard.md`，暂停确认。
-4. 写 `03-asset-plan.md`，收集或生成素材，暂停确认。
-5. 写 `04-edl.md` 和 `master-edit.html`，渲染 `final/promo.mp4`。
-6. 写 `06-music-plan.md`，确认后生成或匹配 BGM。
-7. 合成 `final/promo-with-bgm.mp4`，写 `05-delivery.md`。
-
-## Source
-
-Rewritten for AICUT from [`kangarooking/promo-creator-skills`](https://github.com/kangarooking/promo-creator-skills)
-(MIT), SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 7 skills; this is the one its entry stands for.