Audio Production

Audio Production

Use Agent Audio as the local generation tool, and finish the requested work rather than leaving the user a separate audio-production task. The current installation uses Stable Audio 3 Medium. Small-SFX is a separate benchmark option, not a model selectable through this Skill's MCP workflow.

Decide and inspect

Do not wait for an explicit audio request when an intended audible event is missing from the requested deliverable. Appropriate cases include impacts, collisions, movement, UI feedback, transitions and environmental ambience. Music can be appropriate when the task calls for it; do not add a soundtrack merely because a generation tool is available.

Before generating, inspect the project for suitable audio assets, existing audio systems, timing, style and output conventions. Reuse suitable assets. Do not generate for documentation, analysis, cleanup or refactoring; where silence is intended; when sound would change rather than complete the user's design; or when the user has asked for no audio.

Autonomy and boundaries

Make routine decisions about sound, prompts, candidates and integration within the requested task. Do not introduce a mandatory sound-plan approval or ask the user to choose candidates before completing the work. Honor an explicit request for review or alternatives, and never override the user's scope.

This does not authorize bypassing host permissions, accepting model terms, accessing unrelated files, or paying for external services. If installation, credentials or new permissions are required, report the blocker and obtain the required authorization; do not silently install a different backend or use a cloud provider. No routine creative checkpoint is needed after authorized setup.

Prompt design

Use concrete acoustic descriptions: source/material, action, weight, texture, space and timing. Match the existing sonic style rather than requesting generic "epic" sound. For example, specify a dry, close metallic impact with a sharp transient and a short decay. Include "no voice" or "no music" in the positive prompt when appropriate.

The current tool rejects nonempty negative_prompt; put desired characteristics and exclusions in prompt. See prompt examples when useful. A request for a seamless loop does not establish that the generated file actually loops cleanly.

Generate within a bounded task

1. Call audio_status when runtime readiness is unknown. File readiness is not a successful generation test. Inspect the selected backend and reported limits.
2. Start with the shortest practical asset for the event. A roughly 3-second source is a useful starting point for a short effect, not a speed guarantee. The meaningful event may occupy only part of the generated file.
3. Call generate_audio with prompt, seconds and, when useful, a new .wav output_path in the intended project asset directory. Omitting the path uses Agent Audio's output folder; move/copy and integrate the asset when required. Never overwrite existing outputs or assume a failed call produced a new file.
4. Generate only the distinct assets the task needs. Reuse assets where appropriate; do not request large candidate batches by default. When another candidate is justified, change the acoustic direction deliberately. Stop repeated attempts on a persistent error and report the limitation.

The current interface is text-to-WAV. It does not expose audio editing, inpainting, continuation, loop construction or mixing as MCP operations. Use the project's existing tools for trimming, timing, fades or mixing when available and authorized. Do not invent tool parameters or install new editing software without the appropriate authorization. Long CPU requests can time out.

Integrate and verify the deliverable

When the larger request requires integration, generating a file is not the end of the task. Use the existing project tools and audio system:

• In games, connect the asset to the intended event and check triggering, repetition, restart behavior and the project's existing volume/mute controls.
• In videos, place the useful sound event at the intended timeline position, balance it against existing dialogue/music, and check the rendered result.

These are host-project tasks, not built-in Agent Audio engine/editor adapters. Preserve unrelated behavior and assets. Match the useful transient or audible event, not just the beginning or nominal duration of the generated file. Verify loop boundaries when looping is required; do not infer a clean loop from a prompt.

Check file existence, actual duration, sample rate and channels. With available tools, check silence, non-finite samples, peaks/clipping and the final mix. Such checks do not prove prompt adherence or artistic quality. Use actual listening or audio-analysis capabilities when available; never claim to have listened when only metadata was inspected. Without that capability, state the quality limit in the final report rather than inventing a listening assessment or forcing an intermediate user selection.

If generation or integration is blocked, do not silently substitute a paid service or procedural beep and present it as model-generated audio. Preserve completed work and report the specific missing part. Do not claim the whole sound workflow passed when only a WAV or a code reference was produced.

Final report

Report the finished result, where the audio was integrated, important asset paths, checks actually performed and remaining blockers or unverified quality. Do not expose hardware details unless relevant to performance or troubleshooting. Never claim CUDA, XPU or Apple acceleration unless the reported backend supports that claim. Model-license acceptance always remains a user action.