voice 🗣️
Four skills that decide how output is written, with an explicit routing rule for picking between them.
- human-voice — prose a human reads. Verdict first, scannable layers,
claims tagged
✅verified /⚠️inferred /❓unverified. - machine-voice — machine-read artifacts: agent traces, tool-call logs, status lines, state dumps, schemas. Compressed lexically, structured for scanning, emojis as typed status markers.
- ai-writing-mistakes — the wording pass. Removes the tells that mark prose as machine-written (negation flips, rule-of-three padding, hedge stacks, unearned intensifiers, empty closers) by restoring the claim each one was standing in for. Not a fourth voice: it never claims an element.
- second-opinion — an offer-only validation pipeline. Batches fact-checks, adds scoped advisor personas, re-emits the verdict grouped into verified / flagged / conflict with a mandatory delta.
Install
/plugin marketplace add JRichlen/agent-plugins
/plugin install voice@jrichlen
What's in the box
| Path | What it is |
|---|---|
skills/human-voice/ | Prose discipline: intent triage, layered layout, confidence tags |
skills/machine-voice/ | Compression layers, plus references/lexical-patterns.md loaded on demand |
skills/ai-writing-mistakes/ | The wording pass, plus references/tells.md loaded on demand |
skills/second-opinion/ | Validation pipeline with a hard subagent budget |
hooks/ | Optional session-start context injection (Claude Code only) |
commands/voice.md | Ask which voice governs a given element |
evals/cheap/checks.sh | Deterministic checks that the invariant is intact |
evals/promptfoo/ | Behavioral tier — an LLM judge on routing and refusal |
How routing works
Per output element, not per response. One reply commonly contains both: the
explanation is human-voice, an embedded trace is machine-voice.
ai-writing-mistakes is outside this decision. It claims no element; it revises
wording inside prose one of the voices already placed, and inside prose you were
asked to author into a file.
Code and file contents, commit messages, creative writing, and turns that are only a clarifying question or only tool calls fall outside both — they ship unstyled. That exemption beats the machine-voice list: a config or schema you were asked to author is file contents, not a machine-read artifact.
The always-on part
On Claude Code, a session-start hook injects the routing rule once per session (and again after a context compaction) so the right skill is consulted without being asked. It costs one short injection per session, not a per-turn tax.
The hook is a convenience, not a dependency. On any other harness the skills
still work — they are selected the ordinary way, by their descriptions, and no
skill's instructions assume the hook ran. To turn the always-on behaviour off
while keeping the skills, uninstall the plugin and copy the four skills/
directories into your own skills path.
The one thing this plugin will not do
Neither skill with teeth will report work it did not do. ai-writing-mistakes
will not label a draft "cleaned up" without having changed it, and will not
rewrite prose you asked it only to review — a review request gets the line
quoted, the tell named, and the rewrite offered.
second-opinion refuses to produce validation output it did not earn. Where no
subagent-spawning tool exists, it is required to say so and offer search-based
verification instead — emitting a grouped Verified/Flagged/Conflict block and a
delta without having dispatched anything would counterfeit exactly what the user
asked for when they said "are you sure".
License
MIT