菜单
Publish durable spoken updates and answerable spoken questions through a standalone TTS29 daemon. Use when the user requests narration, a voice update, spoken progress, cross-device speech, or a bounded spoken question. The skill is a producer adapter only; it does not synthesize, pair devices, own playback, or run Nostr.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Capture iterative design explorations by naming the session, maintaining notes, and delaying implementation until clarity emerges. Use when the user is probing ambiguous architecture, system behavior, agent design, workflow design, product design, protocol design, meta-prompting/skill design, implementation strategy, or another complex design space across alternatives, objections, revisions, tradeoffs, ownership boundaries, invariants, failure modes, or integration risks. Complexity alone is not enough: use only when the user is exploring, probing, revising, comparing, or converging over time. Do not use for simple factual questions, clear implementation/edit requests, code review, CI fixes, releases, direct GitHub/PR work, simple prompt rewriting or one-shot prompt optimization, or when notes cannot be written.
Learn durable procedures from repeated delegation. Use when an agent should recognize a recurring request, recover how this user expects it done, execute from a lightweight runbook, or turn a successful new task into reusable procedural memory without making the user design the process.
Resolve and use a durable private home directory at ~/.agents/home/{identifier} for agent-only notes, scripts, caches, drafts, and other state that must persist across sessions.
Record evidence-based feedback after using another skill. Use when real use exposes unclear, missing, stale, inefficient, rigid, or ineffective guidance, a trigger misfire, or a skill-caused correction. Create or append canonical issues in the target skill's meta-feedback directory; exclude speculative reviews and general editing.
Explore a topic, codebase, system, document, tool, workflow, or unfamiliar concept and explain it succinctly at a high level for a user who does not want or need deep internals. Use when the user asks for a plain-language overview, "what is this?", "how does this work at a high level?", "help me understand", "explain without overwhelming me", "summarize the architecture", "give me the mental model", or similar broad orientation.
Use when turning an open-source repository into a project people understand, want to try, trust, star, share, or contribute to. Applies to README rewrites, launch preparation, repo audits, skill catalogs, CLI/library/app/framework/MCP/template repos, agent-facing repos, and root-structure recommendations. Separates human-facing positioning from agent-router metadata and produces adoption copy, proof, quick starts, trust signals, and maintainable repo shape.
| name | tts |
| description | Publish durable spoken updates and answerable spoken questions through a standalone TTS29 daemon. Use when the user requests narration, a voice update, spoken progress, cross-device speech, or a bounded spoken question. The skill is a producer adapter only; it does not synthesize, pair devices, own playback, or run Nostr. |
Resolve <skill-dir> to the directory containing this file and run
<skill-dir>/scripts/tts. Do not assume the current directory or PATH contains
the skill.
Every request needs a stable agent name, concise subject, factual preview, and the words to speak:
<skill-dir>/scripts/tts \
--agent-name "<stable-agent-name>" \
--subject "<2-to-5-word title>" \
--summary "<one factual preview sentence>" \
--message "<spoken update>"
Keep the same agent name during one workstream. Keep the subject under 10 words and 80 UTF-8 bytes. Keep the summary under 280 UTF-8 bytes and the primary message under 300 words; the adapter rejects more than 330 words. Write for listening: short plain-language paragraphs, no raw logs, and no code that would be painful to hear.
The command shapes one stable TTS29 producer request and invokes the standalone
tts29 CLI. It returns durable publication evidence as JSON. It never starts a
player or claims that any device played the item.
Before using --ask, read
asking-questions.md. Every ask requires an
explicit wait no longer than five minutes. A timeout does not undo publication.
Before adding supporting artifacts, read durable-artifacts.md. The adapter accepts only complete HTTPS artifact descriptors. It deliberately refuses local file paths because artifact upload belongs to the standalone daemon/product, not the installed skill.
The skill owns only request validation and shaping. tts29d owns Kokoro,
Blossom, durable jobs, membership repair, NMP publication, receipts, and bounded
answer observation. Apple and compatible NIP-29 clients independently project
and play the durable queue. Do not add pairing, local synthesis, queue state,
playback policy, Nostr code, or an MCP server to this skill.