SuperAPI AI Directory

artcraft-use

General & other

View repository

Updated 10-16 06:47

About this skill

ArtCraft Hybrid Creation

This skill is responsible for decomposition, tool selection, dependency handoff, and delivery verification. The standalone skills in the four domains are responsible for their respective native projects. The current platform is macOS arm64 and requires Python 3.11+; it does not depend on global Node and does not use plugin-private paths.

Execution implementation converges in the built-in artcraft-harness skill (installed bundled with the plugin; for standalone installation use npx skills add full-aigc-skills/artcraft-skills --skill artcraft-harness). Set HARNESS_DIR to the directory where that skill is actually loaded (absolute path); this skill's SKILL_DIR is only used to read its own examples/ and references/, and script execution paths always use $HARNESS_DIR/scripts/。

Installation and First Use

Locate the real directory of this skill. When the user has requested completion of creation or installation and authorization covers the necessary dependencies, run the public entry point; first install the Node and ArtCraft orchestration runtime with pinned versions and digests; when executing workflow.py, then install the required domain skill sources and locked CLIs according to the task graph. The installation scope is only the user data directory, no sudo is required, and PATH is not modified.

Set SKILL_DIR to the directory (absolute path) where the host actually loads this SKILL.md . User-level installations may be located in ~/.agents/skills/artcraft-use, project-level ones may be located in .agents/skills/artcraft-use, and plugins may be located in their skills/artcraft-use or host cache directory; use the actual load path as the standard, do not guess based on the current working directory, and do not search and then arbitrarily select a duplicate version. The skill directory and the CLI user data installation directory are two independent locations.

: "${SKILL_DIR:?请先设置为本 SKILL.md 的实际所在目录}"
python3 -I -B "$HARNESS_DIR/scripts/bootstrap.py" --runtime-only

This base installation returns JSON containing nodeExecutable、entryPoint,skills is empty; the installation receipt after task graph execution lists only the domain runtime identities actually used. The current lock file points to a pinned development release artifact; a missing artifact or digest mismatch will fail explicitly. Each domain skill maintains its own version and does not follow the ArtCraft runtime in a pseudo-upgrade.

Offline artifacts can be specified with --node-archive、--bundle-dir、--native-archive-dir ; these parameters do not bypass digest, path, or version verification.--runtime-home or CRAFT_RUNTIME_HOME specify an isolated installation directory. When reusing a version, re-verify the files; a corrupted version reports an error and is retained, not overwritten or silently upgraded.

Planning and Execution

First record the requirements, dimensions, duration, frame rate, assets, copy, delivery format, and modification boundaries. Choose tools that meet the native delivery requirements. Graphics -> layer design uses VectorCraft/PhotoCraft; motion graphics -> timeline uses EffectCraft/FilmCraft. CapCut is handled by its own standalone plugin and is not within ArtCraft's adaptation scope. Image Factory adaptation is still not connected to this runtime; the optional Video Factory public verification node must explicitly register installed plugins and media tools, see this skill's references/video-factory.md.

Seeplanning contract. Example examples/brand-campaign.json generates a logo, layered poster, motion graphics intro, and a short film with subtitles and audio. The example requires the user to provide voice WAV:

python3 "$HARNESS_DIR/scripts/workflow.py" \
  "$SKILL_DIR/examples/brand-campaign.json" \
  --output /absolute/path/brand-project \
  --authorization brand-project-authorized \
  --asset voice=/absolute/path/voice.wav

--authorization is a reference to an existing task authorization, not a credential or an additional approval process. When there is no audio, the plan should be explicitly modified and audioRequiredshould be canceled; fake voiceover must not be generated to satisfy the delivery check.

Each node specifies the domain pluginId, dependencies, public asset binding, and bounded domain planning. At runtime, real identities are added from the locked registry; the model cannot choose the interpreter, script, or shell. All inputs are verified before consumption, and any missing or changed input stops immediately. The project directory owner, plan revision, and invocation scope are fixed; repeated runs of the same revision reuse verified results.

Modification, Recovery, and Delivery

  • When modifying a semantic plan, use a new revision; content changes within the same revision are reported as conflicts. Only affected nodes are rebuilt, unrelated nodes are reused; old deliveries are retained. The same workflow authorization scope shares a budget; the first plan does not consume a revision round, and each subsequent new revision consumes one maxRevisions. The example allows one round; do not raise the limit or change the currency within the same authorization scope.
  • Development version 3 supports registering a source project revision: register the complete artifact and root of the previous result as node externalInputs, with payload.sourceProject.assetId pointing to that input and expectedRevision using nativeProjectRef.sha256. The domain plan carries no document; the adapter fills in expectedProjectSha256 and calls the public --source. Save as a new delivery and verify that the old delivery is unchanged; see the plan contract for the complete example rules. Development version 4 has fixed the installation lock contention; under four concurrent executors, 16 EffectCraft technical samples and the full ArtCraft parallel regression passed. This evidence does not represent production concurrency capacity.
  • During a run, cancel the registration intent and wait for the real child process and process group to stop; unclear results retain the write occupancy and do not replay the task. Development version 6 supervises native subtasks in an independent worker; after the scheduler exits, rerun the same public workflow command, verify the stop evidence of the same execution, and accept the artifacts. If the worker itself crashes or the commit window is unknown, continue waiting and retain the write occupancy; do not retry based on the PID disappearing. SeeRecovery Contract。
  • Deliver the SQLite ledger, frozen plan, installation receipt, result manifest, and each child node's native project, collected assets, previews, and exports. Native project references remain verifiable; moving the package may require the corresponding skill to relink assets.
  • review_ready indicates technical verification only. It is also necessary to check the actual visuals, text, audio, and user requirements; the shared budget upper bound has been controlled before native side effects; actual settlement of paid services and final creative review are not yet complete, so production release cannot be declared passed based on this.

The four independent skills hand off by name and do not use relative links to sibling skills. Independent installation:npx skills add full-aigc-skills/filmcraft-skills --skill filmcraft-use; the other names are effectcraft-use、photocraft-use、vectorcraft-use, belonging respectively to the same-named -skills package. The plugin also pins the release tag and digest; independent installation and plugin-hosted installation are different evidence scopes.

Current verification scope: handoff of the four native public workflows, downstream rebuild after a Logo semantic change, retention of the source project and existing audio, CLI reopen, and status verification. Complete clean first use and online release status are subject to the latest tests and repository evidence.

Development version 5 provides project packaging and moved-package verification: use the artcraft-harness skill's scripts/package.py create/verify, seeDelivery Package Contract. Collect native projects, registered inputs, and workflow records from the trusted ledger, retaining the technical pending-review status; after moving, verify and relink using the independent skill source project entry point.

ArtCraft dev.7 consumes the four-domain dev.2exchange loss report. The report reference is written into every public asset and packaged with the project; verification of native/export/reopen checks identity and rejects unknown fidelity masquerading as verified or lossy native substitution.

Select independent skills by task

Skill Trigger task
artcraft-cli Query actual command parameters and capabilities, invoke public CLI, MCP, and diagnostics
artcraft-cli-setup First installation, digest verification, version checks, and troubleshooting missing runtimes
artcraft-cli-plan Decompose the tools, DAG, and delivery constraints for brand graphics, posters, title sequences, and promotional videos
artcraft-cli-execute Execute an existing ArtCraft plan, verifying input versions and concurrent writes
artcraft-cli-assets Register asset digests, sources, and subproject dependencies, and verify affected artifacts
artcraft-cli-revise Replace a Logo or specified asset, modify the native subproject, and update downstream
artcraft-cli-recover Check task status, cancel the original task, and resume after the scheduler exits
artcraft-cli-deliver Collect native subprojects, assets, plans, and evidence, and verify after moving the package
artcraft-cli-review Check technical evidence for hybrid delivery and consistency of visuals, copy, and audio

Missing skill:npx skills add full-aigc-skills/artcraft-skills --skill <skill-name>. Implementation convergence is confined to the artcraft-harness skill; directly executing that skill's scripts/cli.py can also query the current CLI without relying on sibling paths.

When a finished video needs to be verified through an existing Video Factory, read this skill's public verification handoff. This adaptation preserves NOT_RUN and does not replace creative review or the native project.

When review evidence needs to be saved, use the artcraft-harness skill's scripts/review.py, contract see review record. It records named observations, responsible plugins, and binding to the current version; it does not execute models or automatic revisions, and package verification success cannot replace creative and human acceptance.

When reworking repeatedly according to review issues, use the artcraft-harness skill's scripts/revision.py, according to controlled revision guide freeze the strategy and execute explicit patches, saving rounds, stagnation, and the best package anchor. Unknown results are not automatically replayed; complete model planning/evaluation and creative acceptance are verified independently.

Brand palette mixed rework

This skill ships with a five-node example for brand graphics/posters/title sequences/finished videos and standalone icons. Follow this skill's brand palette guide to execute and verify affected artifacts; the domain required for first installation is determined by the fixed dependency lock.

Versioned requirement records can use the artcraft-harness skill's scripts/brief.py; seeBrief contract and current boundaries. Python workflows support --brief and --brief-sha, rejecting blocking requirements before installation; Node and Python executors verify requirement constraints, installation uses the artcraft-harness skill distribution lock to pin versions; technical pass and creative acceptance are recorded separately.

Local modifications to a PhotoCraft source project can be declared in the domain plan as protectedRegions(named rectangle id/rect=[x,y,width,height]); the domain workflow checks protected regions before publishing, and comparison PNGs and reports are retained with the subproject and project package. Do not treat protected-region reports as creative acceptance.

Dynamic transparent sequence handoff

When a hybrid task requires transparent animation, complete frame handoff, or downstream updates after a Logo replacement, use this skill'sdynamic sequence guideand examples/dynamic-brand-campaign.json. Verify all frames, time base, native references, and local rework according to actual receipts; do not substitute the first frame or ordinary JSON for the sequence.

When the result is unknown after saving, follow this skill's references/recovery.md verify failure.json, the original staging project, and stop evidence; do not replay the same task, and a failed staging is not a successful delivery.

Complete domain command handoff

Read this skill when native operations beyond the template are neededComplete domain command components, use the built-in domain_commands.py to query 2,646 commands, actual MCP parameters, and hand off the selected domain. Each skill includes four-domain creation/rework examples; unknown is not replayed, and component receipts do not replace native DAG delivery and acceptance.

Native gradients, multiple fills, and source-project color-change rework: use this skill's Vector appearance guide。

Native parent, transparency expressions, and source-project rework: use this skill's Effect expression guide。

For Photo native adjustment layers, selection masks, and trusted-source rework, see Local adjustment guide。

When you need to launch domain GUI commands independently, see Own desktop handoff。

For mixed tasks, refer to Business scenario manual, first confirm dependencies and input versions, then execute sub-projects, local rework, and delivery verification.

When installation fails, read First-use diagnostics, and use the receipt to locate the current skill's own setup entry point; handle installation failures and native invocation failures separately.

For real dubbing recognition and mixed subtitle tasks, please read this skill's Mixed ASR guide / Mixed ASR guide, and execute according to the fixed domain dispatch and persistent model directory; planning, rework, and acceptance each retain dependencies, native projects, and execution evidence.

For task identity and rejection details, read according to Task receipt description ; node summaries do not replace actual execution receipts.

Before reworking mixed posters, covers, and source projects, read Photo complete delivery, distinguish between standalone host plugins and Art's internal fixed dependencies.

For cross-artifact observation of fixed brand and subject references, see this skill's Consistency review contract; acceptance of source candidates and published fixed versions is reported separately.

Runtime upgrades and old ledgers

When encountering runtime_upgrade_busy , keep the old runtime, original project, and ledger; do not delete leases or forcibly modify state. First complete or reliably stop the task with the old version, then migrate according to the new version; read-only status 's old budget markers indicate historically untracked. For details, see this skill'sUpgrade guide。

For validation and mode boundaries of native command execution, see Command contract boundaries。

For classification of mode catalog candidates, parameter differences, and rejection boundaries, seeMode catalog usage; Vector desktop/bridge duplicate commands have not yet completed execution qualification and cannot be rolled back according to the headless catalog.