ambient-tinystyler

Imported from ambient-xyz/AmbientDesktop (resources/ambient-cli-packages/ambient-tinystyler/SKILL.md). Install upstream with npx skills add ambient-xyz/AmbientDesktop --skill ambient-tinystyler. Copyright stays with the author.

Ambient TinyStyler

Use this package when the user wants local few-shot writing-style transfer from their own example texts. The v1 target is profile creation plus transfer with pretrained TinyStyler assets, not user-specific model fine-tuning.

Discovery And Setup

• Find this package with ambient_cli_search using terms like TinyStyler, style transfer, style profile, rewrite in this style, or writing style examples.
• Call ambient_cli_describe before first use so model assets, command contracts, current implementation status, and artifact behavior are visible.
• Run tinystyler_doctor --json before profile or transfer. It is non-mutating and does not download models.
• Profile creation uses local Style-Embedding assets when the approved Python dependencies and model cache are present. Transfer generation uses local TinyStyler/T5 assets when approved dependencies and caches are present. Use --fake only for deterministic local validation.

Commands

• tinystyler_doctor --json: reports Python/platform facts, declared Hugging Face revisions, required model assets, cache state, and safety boundaries.
• tinystyler_profile --examples-file <path> --output-profile <path> --profile-name <name> --json: creates a profile artifact using local Style-Embedding assets. Use --fake only for deterministic local validation, not for user-facing style profiles.
• tinystyler_transfer --input-file <path> or --text <text> plus --profile <path> or --examples-file <path> and --output-file <path> --json: writes transfer text plus bounded JSON metadata. Use --fake only for deterministic local validation, not user-facing transfer.

Prefer file inputs when text contains punctuation, whitespace, quotes, or user-provided examples. Do not pass long user samples as CLI arguments.

Artifact Contract

Profiles are JSON artifacts with schemaVersion ambient.tinystyler.profile.v1, provenance, aggregate source counts, model revisions, generation defaults, safety metadata, and embedding values. Raw example text and exact source verifiers are not persisted by default; raw text requires the user to explicitly request --include-source-text true.

Transfer always writes the exact generated text to required --output-file and returns concise JSON metadata on stdout. Large outputs should be files, not chat blobs.

Safety And Boundaries

• Treat this as style adaptation, not identity verification or impersonation.
• Never expose secret values. TinyStyler itself does not require API keys.
• Do not persist raw user examples by default.
• Do not silently fall back to prompt-only rewriting when model assets are missing.
• Optional reranking is deferred until LUAR/MIS remote-code, dependency, and runtime-cost risks are reviewed.
• If real model execution fails, report the failure class and safe next step rather than repairing output text after generation.