# Golden Voice Packets > Use when writing, revising, speaking, or thinking in a named human or agent voice; using golden voice packets, voice packs, voice guides, tone guides, or prompt packs; creating or updating voice packets; or learning voice preferences from user-edited model drafts. - **Type:** Skill - **Install:** `agentstack add skill-dbmcco-claude-agent-toolkit-golden-voice-packets` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [dbmcco](https://agentstack.voostack.com/s/dbmcco) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [dbmcco](https://github.com/dbmcco) - **Source:** https://github.com/dbmcco/claude-agent-toolkit/tree/main/skills/golden-voice-packets ## Install ```sh agentstack add skill-dbmcco-claude-agent-toolkit-golden-voice-packets ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Golden Voice Packets ## Purpose Use the canonical voice packet system in `/Users/braydon/assistant-kb/people/`. Do not store packet content in this skill. Load only the files needed for the current task, and treat private examples as sensitive source material. ## Canonical Layout ```text /Users/braydon/assistant-kb/people//voice.md /Users/braydon/assistant-kb/people//voice-pack/ /Users/braydon/assistant-kb/private/gold// ``` Known human packet starts: - Braydon: `/Users/braydon/assistant-kb/people/braydon/voice.md` - Logan: `/Users/braydon/assistant-kb/people/logan/voice.md` - Amrish: `/Users/braydon/assistant-kb/people/amrish/voice.md` Agents use the same `people//` pattern. If an agent has `style.md` or `persona.md` but no `voice.md`, use those as source material and create `voice.md` only when the user asks to develop that voice pack. ## Load Order 1. Resolve the subject slug under `/Users/braydon/assistant-kb/people/`. 2. Load `voice.md` first if it exists. 3. Load `voice-pack/README.md` when you need the packet index or provenance. 4. Load the gold master when the task needs high fidelity or packet updates. 5. Load one matching mode pack when the task clearly maps to a mode. 6. Load private gold examples only when the user explicitly asks for few-shot fidelity or packet maintenance that requires them. Useful discovery commands: ```bash find /Users/braydon/assistant-kb/people -maxdepth 2 -name voice.md -print find /Users/braydon/assistant-kb/people//voice-pack -maxdepth 1 -type f -name '*.md' -print ``` If the subject is ambiguous, ask one concise clarification rather than guessing. ## Drafting Rules - Use voice packets as guidance, not as a license to impersonate without review. - Never label generated text as a real quote. - Never send, publish, or externally represent generated text as someone else's words without explicit human approval. - Preserve provenance boundaries: observed source, inferred pattern, and speculative adaptation should be distinguishable. - For outreach or external messaging, combine this skill with the relevant drafting/review skill and keep the output human-reviewable. ## Creating Or Updating A Pack Use this structure for humans and agents: ```text people// |-- voice.md `-- voice-pack/ |-- README.md |-- _Voice_Gold_Master.md `-- __Prompt_Pack.md ``` Update the smallest file that carries the change: - `voice.md`: compact runtime rule used by agents first. - `voice-pack/README.md`: inventory, provenance, privacy notes, update rules. - Gold master: general voice principle, source model, or prompt sandwich. - Mode pack: email, posting, scientific writing, meeting talktrack, TTS, or another recurring mode. - `/Users/braydon/assistant-kb/private/gold//`: private examples. When adding examples, redact or summarize unless the user explicitly wants a private gold-bank entry. Do not put raw private examples in public repos, campaign folders, or skill files. ## Learning From User Edits When Braydon edits model-written output, treat the edit as supervised signal for the relevant voice pack if both versions or a clear description of the change are available. Process: 1. Identify the subject voice and mode. 2. Compare model draft vs user-edited version. 3. Extract the reusable preference: cadence, diction, structure, claim discipline, warmth, hedging, examples to use, or examples to avoid. 4. Update the smallest canonical file. 5. Mark confidence in the update: - `Observed from explicit edit` - `Inferred from repeated preference` - `Speculative; confirm before relying on it` Do not update a packet from typo fixes, factual corrections, one-off reversals, or edits that change the underlying argument rather than the voice. Good update shape: ```markdown - Observed from explicit edit, 2026-06-08: In short partner follow-ups, prefer the concrete next decision before the relationship context. Avoid opening with broad appreciation when the ask is operational. ``` ## Consolidation Rules - Assistant KB is canonical for packet content. - Repo-local copies are mirrors, frozen campaign exports, or working material unless their README explicitly states otherwise. - Logan's Synthyra HANTA packet copies have historically matched the assistant-kb packet by hash; verify before replacing or deleting. - Braydon's `personal/dbm_dossier` material is legacy/tooling source unless the assistant-kb packet points there for scripts. - Amrish's assistant-kb packet is canonical; add LFW-specific modes only when a recurring use case justifies them. Prefer README pointers or symlinks over duplicate copies only when the consuming tool handles symlinks safely. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [dbmcco](https://github.com/dbmcco) - **Source:** [dbmcco/claude-agent-toolkit](https://github.com/dbmcco/claude-agent-toolkit) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-dbmcco-claude-agent-toolkit-golden-voice-packets - Seller: https://agentstack.voostack.com/s/dbmcco - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.