更新日志 | NEXAGEN Skills

All notable changes to NEXAGEN Skills are recorded here.

Unreleased

website-builder: Design Director discipline (website type, business context, conversion strategy, design DNA, art direction, archetype mix, per-section layout grammar and spacing rhythm, media necessity per section, one self-critique pass) and a note on the Website Studio app's in-app Design Director and approval-gated visual generation. Packages not regenerated yet (build pending approval).

Requires NEXAGEN MCP server 2.2.0.

Website Studio (website-builder): website / landing page planning, design system, React / Next.js code, motion graphics, responsive / accessibility / SEO / performance review and build-ready handoff, with optional site videos and images through the existing NEXAGEN workflows (estimate and confirmation first). No deploy, preview, file or code execution tool exists in NEXAGEN, so the Skill delivers code in the conversation and never claims to publish. Catalog, docs (ko/en/ja/zh), routing tables and packages regenerated.
Magic: the default video engine (engine_key magic, first list_engines(kind=video) entry with is_default=true). When engine is omitted, Magic automatically selects a high-quality generation workflow from curated top-tier video models for the prompt, input assets, duration and options. A named engine is still used as given. video-creator, workspace-manager, GLOSSARY, GENERATION_POLICY, web catalog (video-creator features / flow / FAQ, ko/en/ja/zh).
estimate_video: optional prompt and aspect (Magic only) so the estimate follows the method Magic selects for the same input.

video-creator: no engine requested -> Magic; re-estimate when the inputs change; MAGIC_ROUTE_STALE / MAGIC_NO_ELIGIBLE_ENGINE handling; with Magic, report the submit response's estimated_credits (references are considered at submission).
Tool inventory unchanged (62). Snapshot tools-list.2026-10-15.json re-projected from the MCP 2.2.0 registry source: descriptions/schemas of list_engines, get_engine_capabilities, estimate_video, generate_video, generate_video_from_image, create_workspace_video, and the server's current wait_for_job timeout_seconds range (1-180).

Requires NEXAGEN MCP server 2.1.0.

estimate_video: free pre-submission credit estimate for canvas video (same server estimate as the credit hold). video-creator, COST_POLICY.
wait_for_job: server-side wait until a job is terminal; returns the get_result shape. JOB_POLICY, video-creator.
upscale_video (paid): 4K upscale of a finished NEXAGEN Video Lab video as a new job. video-creator, COST_POLICY.
engine_key: human-readable stable engine key in list_engines; accepted as engine alongside the legacy engine_id.
PLAN_CONCURRENCY_LIMIT error guidance (plan slots shared with the NEXAGEN web app). JOB_POLICY.

Job waiting: every Skill now waits with wait_for_job (call again on timed_out=true; the response carries still_running, job_not_cancelled, next_action, retry_after_sec). Short get_job polling loops are no longer recommended. NEXAGEN MCP has no push notification; this is stated in JOB_POLICY and video-creator.
Plan concurrency is documented as one account-wide pool (Video Lab, Canvas, AD Studio and MCP video together); MCP videos over the limit wait in the queue. JOB_POLICY, ERROR_POLICY, video-creator.
Canvas video cost: estimate_video replaces the former "no pre-submission estimate" note (video-creator, creative-director, CORE_SUMMARY; generate_video / generate_video_from_image manifests now name estimate_video as their estimate tool).
Tool inventory 59 -> 62 (snapshot tools-list.2026-10-14.json, re-projected from the MCP 2.1.0 registry source). The release scanners recognise the wait_ tool verb.

Public distribution hardening. Tool names, Skill names, workflows, paid confirmation, job polling and duplicate-submission rules are unchanged.

Polling guidance now follows the server's response hints (next_poll_after_ms, retry_after_sec, get_mcp_policy.limits) instead of fixed numbers in the Skills.
Uncertain paid submissions: check the existing job / recent results and honor NEXAGEN's idempotency response; no fixed time window is documented.
Error handling is driven by the returned error_code, retryable and retry_after_sec; availability errors are described as account availability only.
Upload guidance reduced to the minimum contract: prepare_upload -> follow the returned upload instructions -> register_uploaded_asset.
"Current project" guidance rewritten as client behavior (list recent candidates and ask).
Removed a reference to a manifest file that is not part of every package.
Public tool manifest: internal collection notes and rollout fields removed.
New release check scripts/public_skills_leak_scan.sh scans every public package before publication.

First release of the official NEXAGEN Skills pack.

12 Agent Skills: creative-director, image-creator, video-creator, cineflow-director, storyboard-creator, ad-campaign-creator, cinema-director, music-producer, voice-producer, canvas-assistant, workspace-manager, project-asset-manager.
Shared core policies: GLOSSARY, ROUTING, CONTEXT_POLICY, GENERATION_POLICY, COST_POLICY, JOB_POLICY, RESULT_POLICY, ERROR_POLICY, SAFETY, CORE_SUMMARY.
Tool inventory of the 59 public MCP tools with schema hashes, and a capability map (22 domains).
Skills catalog and web catalog (ko/en/ja/zh) with JSON Schemas.
Japanese and Chinese (ja/zh) catalog text, example requests and user docs (getting started, FAQ, troubleshooting, install guides, Skill docs).
Packages: Complete Pack, Claude, Cursor, Codex, ChatGPT (OpenAI) and 12 single-Skill ZIPs, with SHA256SUMS and reproducible builds.
QA scripts: tools, Skills, packages, drift, catalog, routing overlap, secret scan; 20-case scenario matrix.

Skills aligned with the public MCP tool behavior: workspace image jobs are polled without source, workspace results come from get_workspace, storyboard plans need a finished mosaic, split cuts are URLs only, Cinema saved direction is not applied by generate_cinema_video, AD element ids come from get_cinema_project, music/voice estimates are estimates (final charge on the job), voice scripts over 5000 characters follow split plan -> estimate -> confirmation.
Tool snapshot refreshed after public tool contract updates: create_workspace_image.count max = max_workspace_image_candidates, create_ad_project.ad_type points to get_ad_options, Cinema direction descriptions.
QA: runtime contract checks (17 no-credit paid tool cases, duplicate-submit guard, error handling, ID chain, overclaim scan) and 35 scenario traces.

See https://ai.nexagen.co.kr/skills for current feature status.