Speak With Profile

Use this skill as the required front door for speech tasks. Treat built-in $speech as the synthesis engine, and keep this skill as the policy and profile adapter.

Purpose

• Normalize speech requests with deterministic profile precedence.
• Enforce disclosure and manifest policy consistently.
• Delegate generation to $speech by default in Codex App/CLI conversations.
• Keep a script fallback for deterministic local/automation execution.

Required UX pattern

1. Route speech requests through $speak-with-profile first.
2. Resolve profile/default/override values.
3. Require disclosure text in end-user output context.
4. Delegate generation to built-in $speech using resolved fields.
5. Record effective configuration in a run manifest.

Execution modes

• delegate (default for conversational usage): use built-in $speech for synthesis.
• local-cli (fallback for scripted/automation runs): invoke scripts/text_to_speech.py via scripts/speak_with_profile.py.

Use delegate unless there is a concrete reason to run the local script path.

Inputs needed

• Input source: either the exact text to speak or a path to a text file.
• Profile choice: a --profile ID if the caller wants a named profile, or no profile if baseline defaults are acceptable.
• Profile data source: a --profiles-file path when profile-based resolution is required.
• Delivery guidance: any explicit voice, instructions, speed, or format overrides the caller wants to force.
• Output intent: target output path or filename expectations for the generated audio.
• Execution preference: whether built-in $speech is acceptable or deterministic local-cli behavior is required.
• Playback intent: whether the generated file should also be played locally with open or afplay.
• Optional skill-local workflow customization from config/customization.yaml or config/customization.template.yaml.

Hard constraints

• Do not bypass this skill for speech tasks in this repository workflow.
• Do not modify built-in $speech behavior assumptions; adapt inputs/outputs around it.
• Never modify the bundled generation script scripts/text_to_speech.py.
• Use built-in voices only.
• Require OPENAI_API_KEY for live API calls.

Disclosure policy

Always include a clear disclosure in user-visible output when speech is produced, for example:

• "This audio was generated by AI text-to-speech."

If a profile provides disclosure, use it. Otherwise, use the default disclosure above.

Profile resolution order

1. Explicit wrapper flags (--voice, --speed, --instructions, --format, output path flags).
2. Selected profile (--profile).
3. Default profile from profiles file (default_profile).
4. Baseline defaults (voice=cedar, speed=1.0, format mp3, model gpt-4o-mini-tts-2025-12-15).

Workflow configuration precedence

1. Explicit user input and wrapper flags.
2. Skill-local workflow customization in config/customization.yaml, when present.
3. Skill-local defaults in config/customization.template.yaml.
4. Workflow defaults described in this skill and references/wrapper-contract.md.

The local wrapper loads supported customization values from config/customization.yaml and config/customization.template.yaml at runtime. Use skill-local customization to guide wrapper defaults and agent decisions. preferredExecutionMode remains documentation-only because the wrapper script implements only local-cli.

Workflow

1. Resolve input text source (--text or --text-file).
2. Resolve effective workflow settings using explicit input, skill-local customization, then workflow defaults.
3. Resolve profile and defaults using references/wrapper-contract.md.
4. Validate configuration against references/profile-schema.md.
5. Choose execution mode:
   • delegate: call built-in $speech with resolved fields.
   • local-cli: run scripts/speak_with_profile.py.
6. Emit/validate run manifest and include disclosure.

Output Contract

• delegate: return a user-visible result that confirms speech generation intent, includes the disclosure text, and surfaces the resolved speech settings that matter for the request.
• local-cli: produce an audio file at the requested or resolved output path and write an adjacent <audio>.manifest.json file.
• local-cli manifest/result reporting must include:
  • output audio path
  • manifest path
  • disclosure text
  • execution mode (local-cli)
  • playback result when open or afplay is used
• Playback failures must remain explicit rather than being treated as silent success.

Failure modes

• Missing OPENAI_API_KEY: stop before generation and tell the caller that the local environment must provide the API key.
• Missing text source or unreadable text file: fail fast rather than guessing input content.
• Missing profiles file when --profile is requested: fail fast and require the caller to provide a valid JSON or YAML profiles file.
• Invalid profiles file: stop on schema/parse errors and surface the validation failure clearly.
• Unknown profile ID: fail with the known profile list from the supplied profiles file.
• Input over 4096 characters: fail fast and require the caller to split or chunk the input before running.
• Missing bundled scripts/text_to_speech.py on the local CLI path: stop and report that the expected bundled CLI is unavailable.
• Downstream speech CLI failure: propagate the subprocess failure rather than masking it.
• Playback failure with open or afplay: return the deterministic playback failure outcome, record it in the manifest, and respect --stop-on-error / --no-stop-on-error.

References

• Profile schema and validation rules: references/profile-schema.md
• Starter profile set and examples: references/starter-profiles.md
• Adapter contract and mode behavior: references/wrapper-contract.md
• Skill-local workflow customization: references/customization.md

Validation helper

Use scripts/validate_manifest.py to verify required manifest keys:

uv run python speak-with-profile/scripts/validate_manifest.py path/to/output/file.manifest.json