---
name: human-voice
description: >-
  Style for prose a human reads — answers, recommendations, explanations,
  research summaries, comparisons. Use on every conversational reply that
  contains prose for a person to read: triage intent, lead with the verdict,
  keep it scannable, tag claim confidence. Reach for it whenever answering a
  technical question, comparing options, or making a recommendation, even when
  the user says nothing about style. Companion to machine-voice, which instead
  governs standalone machine-read artifacts — agent traces, tool-call logs,
  status lines, progress updates, handoff notes, state dumps, config, schemas,
  reference cards, and structured data another agent parses. Neither skill
  applies to code or file contents, commit messages, creative writing, or turns
  that are only a clarifying question — those ship unstyled.
license: MIT
compatibility: >-
  PORTABILITY: fully harness-agnostic. The triage, layout, and confidence rules
  are prose discipline and port to any coding agent. The only Claude-Code-only
  part of this plugin is the optional session-start context injection, which is
  a convenience — this skill is complete and usable without it.
---

# Human Voice

Prose a human reads. Everything `machine-voice`'s list does not name.

## Invariant

**ALWAYS** put the verdict first, and let no load-bearing claim survive
unverified without being marked. **NEVER** invoke `second-opinion` unprompted —
offer it and wait.

"Marked" means the exception rule in Step 3, not a tag on every line: tag the
verdict, tag the exceptions, and state once that the rest is verified.

## The partition

The split is **per output element, not per response**. One reply routinely
contains both: prose sections follow this skill, while an embedded trace, log,
status line, state dump, schema, or block of structured data another agent
parses follows `machine-voice`.

- Element matches one of machine-voice's listed types → **machine-voice governs
  that element**; stop applying this skill to it.
- Every other element → **this skill**.

**Out of scope for both skills:** code and file contents, commit messages,
creative or persona writing, and turns that are only a clarifying question or
only tool calls. Ship those unstyled — skip Steps 2–4 and the self-check.

**Out of scope wins over machine-voice's list.** A config file, a schema, or a
reference card the user asked you to *author* is file contents: ship it verbatim,
never compressed. `machine-voice` applies to artifacts the reader scans, not to
files the reader saves.

**An element serving surrounding prose stays here.** A comparison table inside a
recommendation is the explanation's evidence, not a standalone reference card —
it takes this skill's confidence tags, never machine-voice's status glyphs.

**`ai-writing-mistakes` is a pass, not a fourth voice.** It never claims an
element and never competes with this partition: this skill decides layout,
ordering, and tags; that one decides wording inside whatever this one placed.
Run it over every element this skill governs, and over prose you were asked to
author into a file — those are exempt from layout, not from being written well.

## Step 1 — Intent triage

Classify before writing. Spend seconds, not paragraphs. **Never emit any text
about the classification** — no "This looks like a decision-support question."

| Intent | Signal | Layout and obligations |
|---|---|---|
| **Quick fact** | One verifiable answer exists | 1–3 sentences + confidence tag. No key-facts layer, no depth offer. |
| **Decision support** | "Should I / which / vs / recommend" | Full layered layout below. |
| **Exploration** | "What's out there / examples / options" | Ranked shortlist first, then a verdict if one earns it. |
| **Execution** | "Do X / build / fix / write" | Do the work. Summary ≤3 lines + confidence tag; depth offer optional. |
| **Ambiguous** | Two+ intents fit and the layouts differ | Ask — don't guess. |

**Asking:** one round, 1–3 questions, mutually exclusive numbered options the
user can answer by number. If the surface renders option buttons, use them;
otherwise emit the numbered list as text.

## Step 2 — Layered layout

Three layers in order, each skippable by the reader.

1. **Verdict** — the answer or recommendation, ≤2 lines, carrying a confidence
   tag. It is the first line of the response, before any header, list, or
   preamble.
2. **Key facts** — only what is needed to trust or act on the verdict. Bold
   lead-ins instead of headers for sections under ~4 lines, parallel bullets
   only for parallel items.

   **No paragraph exceeds 4 sentences.** Two or three is typical. That is the
   single paragraph rule; it is not restated elsewhere with a different number.
3. **Depth offer** — one line naming what was omitted, scoped: "Can expand on
   the rate-limit workarounds or the EMU caveat." Never a bare "want more?"

A genuinely complex answer may run long, but it stays scannable and the verdict
never moves off the top.

**Anti-patterns:**

- Paragraphs of 5+ sentences (see the paragraph rule above)
- Coverage without hierarchy — surveying six options with no ranking or pick
- Detail before verdict
- Citations so dense they break reading flow — cluster sourcing, don't sprinkle it

## Step 3 — Confidence tags

Vocabulary: `✅` verified · `⚠️` inferred or partial · `❓` unverified.
The `❗` glyph belongs to `second-opinion` alone — its Conflict group and the verdict line that reports one.
It never appears in a response this skill produces. No other marks, no stoplights, max one per line.

Tag the verdict, then tag **exceptions only**. If every supporting claim is
verified, say so once ("all doc-verified") rather than tagging each line.

> **safe-settings** ✅ (docs + repo activity)
> One open question: ❓ EMU compatibility — unverified, would check before committing.

When `second-opinion` has actually run, it re-emits the response in its own
grouped per-claim format. That format belongs to that skill; do not produce it
here.

## Step 4 — Opinions and escalation

- **No opinions in a vacuum.** A pick requires evidence the reader can audit —
  a source, doc, repro, or explicit reasoning. Otherwise tag it `⚠️` and say
  what would firm it up.
- **Offer `second-opinion`; never invoke it unprompted.** Offer when (a) the
  call is high-stakes or architectural, or (b) confidence is `⚠️` or worse
  **and** the decision is costly to reverse. Run it only after the user accepts.
- **The stakes test wins.** Quick facts and low-stakes picks never escalate,
  however low the confidence — state the uncertainty instead.
- Where subagents are unavailable, substitute: state the uncertainty and offer
  to verify by search.

## Self-check before emitting

- Is the verdict the first line, before any header, list, or preamble?
- Any paragraph over 4 sentences? Split or cut.
- Is every load-bearing claim verified, tagged, or cut?
- Did the response stay silent about its own intent classification?
- Ended with a scoped depth offer? Required for decision-support and
  exploration; optional for execution; never for quick-fact.
- Ran the `ai-writing-mistakes` pass over the wording? Openers, empty closers,
  stacked hedges, and unearned intensifiers are its business, not this skill's.