Copywriting Voice Tone Stage

copywriting-voice-tone-stage

Phase 6 — tactical voice + tone fine-tuning. Phase 5 fixed the macro quadrant (Authority↔Affinity × Reason↔Emotion); this phase works inside that quadrant: 4-axis micro voice calibration, Mailchimp-style tone context-switching, and voice-anchor register application (80+ individual- creator anchors across JP / ZH / EN / zh-HK with explicit LLM reproduction gap awareness).

Output is a polished draft + structured tone_notes, handed to Phase 7 ethics check.

When to Use

Invoke this skill after copywriting-voice-quadrant-stage (Phase 5) has stamped a voice_quadrant onto the envelope AND a workflow skill (Phase 4: long-form-pasona / long-form-extended / mid-form / short-form / light-action) has produced a draft.

Skip only when:

• The brief is single-sentence short-form (≤15 chars) AND no voice maestro

is referenced — Phase 5 output already embodies tone; no micro-tuning adds value.

• The run is an audit — audit re-runs Phase 5-8 against existing copy;

this skill still runs, but treats the provided copy as the draft input rather than regenerating.

When Pass 3 (Voice Anchor) applies

Pass 3 activates iff either condition holds:

1. brief.voice_reference names a specific creator (any language — 糸井重里 / 岩崎俊一 / 眞木準 / 谷山雅計 / 仲畑貴志 / 許舜英 / 李欣頻 / 葉明桂 / or any named author with a matching anchor-{slug}.md in the library).

1. voice_quadrant.primary + voice_quadrant.position are set AND the brief requires register-signal-level voice lock (beyond 4-axis tuning). Signals: tone_cue contains register keywords, message_thesis is stance-loaded, or upstream skill annotates register_lock_required: true.

Otherwise Pass 3 is skipped; tone_notes.register_signal_applied = null.

Cross-tradition transplant forbidden — never force JP signatures onto ZH output (or vice versa), never force either onto Anglo output. voice-and-tone.md §Anti-Patterns + voice-anchor-meta.md §Cross-Master Context document the ban. Cross-language BORROWING allowed only when anchor frontmatter's cross-reference-valid-for[target_lang] == STRONG AND brief explicitly permits it.

If lineage-language conflict detected (e.g., voice_reference: 糸井重里 but output_language: zh-TW) → emit violation envelope per ../../CLAUDE.md §Envelope Violation; router routes to copywriting-intake.

Violation payload:

{
  "violation": {
    "detected_by": "copywriting-voice-tone-stage",
    "missing": [],
    "malformed": [{
      "field": "brief.voice_reference + brief.output_language",
      "expected": "consistent lineage (JP maestro + ja OR ZH maestro + zh-*)",
      "got": "cross-lineage conflict (e.g., 糸井重里 + zh-TW)"
    }],
    "bounce_to": "copywriting-intake",
    "user_message": "Voice reference and output language disagree on lineage — intake needs to resolve which tradition applies."
  }
}

What This Skill Owns

• Standards:
• standards/voice-and-tone.md — Tier 2 cross-framework voice SSOT. Ogilvy 1963/1983 Anglo long-copy canon + JP emotional-resonance tradition + 18F / Mailchimp 4-axis voice model + Mailchimp "one voice, multiple tones" + tone context-switching table.
• standards/voice-anchor-meta.md — consolidated voice anchor metadata (schema + selection rubric + over-mimic fallback registry + verified lineage map + cross-reference registry + cross-cultural label rubric + drift corrections + cross-master context + axis extreme candidates). Consolidated in v1.13.0 from former meta-core + meta-detail + axis-extreme files.
• standards/{jp,zh,en}-q{1,2,3,4}-anchors.md (12 files) — quadrant router index, landmark-organized (center / extreme / toward-Q{N}).
• standards/anchor-{slug}.md (80+ files, all schema_version: 2.0) — per-creator v2 anchor bodies (Voice direction / Native critical read / Prose mechanics / Examples / Don't-over-mimic / Metadata).
• Rubric (SHOULD gate):
• rubrics/voice-consistency-gate.md — cross-stage / cross-candidate voice stability + tone contextual appropriateness + anchor fidelity. Runs AFTER tone tuning as the last voice-layer check before handing to Phase 7 ethics gates.

This skill does not own:

• Voice quadrant selection — copywriting-voice-quadrant-stage (Phase 5).
• Framework structure (PASONA stage ordering etc.) — Phase 4 workflow skills.
• Ethics / legal / ステマ — copywriting-ethics-check-stage (Phase 7).
• Form-appropriate character discipline — copywriting-form-check-stage (Phase 8).
• Library curation (4-condition rubric enforcement) — scripts/lint-anchor-library.py at CI time (v1.13.0+).

Phase 6 Operating Model

Three sequential passes on the draft. Each pass returns either an edited draft or an annotation the next pass must honor. No pass may silently rewrite Phase 4's structural skeleton — frameworks are replaceable, voice is an irreplaceable asset, but structural edits belong to Phase 4 re-runs, not here.

Pass 1 — 4-axis micro calibration (always runs)

Per standards/voice-and-tone.md §Voice definition — 4 axes:

| Axis | Left end | Right end | |----|------|------| | Formality | formal | casual | | Seriousness | serious | funny | | Respectfulness | respectful | irreverent | | Enthusiasm | matter-of-fact | enthusiastic |

Schwartz alignment passthrough: voice_quadrant (including schwartz_alignment) passes through Phase 6 unchanged per ../../CLAUDE.md §Immutable fields. If schwartz_alignment == "conflict_flagged", Pass 1 MAY lightly soften axis defaults (e.g., keep enthusiasm matter-of-fact when L2 forced into Q2), but does NOT elaborate rationale — Phase 8 8b owns the rigorous consumer check.

Action:

1. Extract brand voice axis positions from brief.voice_reference + voice_quadrant. If absent, derive default from Phase 5 quadrant (Q1 ≈ formal / serious / respectful / matter-of-fact; Q3 ≈ casual / warm / respectful / warm-enthusiastic; etc.) and disclose the derivation in tone_notes.axis_default_derived.
2. Self-check for axis drift: lightly scan the draft for sentences whose axis reading obviously conflicts with the target vector by ≥2 steps. Rewrite the flagged sentences. The Voice Consistency SHOULD gate handles rigorous per-sentence enumeration; Pass 1 catches obvious drift.
3. Preserve framework stage boundaries (PASONA A / S / O labels, BEAF stages, etc.) verbatim.
4. Apply Ogilvy commandments explicitly: no empty-hype vocabulary (amazing / revolutionary / game-changing), headline voice = body voice (no click-bait exemption), long-copy must earn every body sentence.

Pass 2 — tone context-switching (runs if the draft contains context-sensitive segments)

Per standards/voice-and-tone.md §Tone context-switching table:

| Context | Tone direction | Caution | |------|------|------| | Onboarding / Welcome | warm, encouraging | no bait-and-switch to hard sell | | Error / Failure | calm, specific, non-deflecting | no humor masking | | Crisis / Incident | serious, timely, transparent | no jokes, no hype, label uncertainty | | Celebration / Launch | enthusiastic, user-grateful | user is subject, not brand |

Voice stays constant; tone varies. If brand voice is "playful," error-state remains playful in voice, but tone becomes calm and joke-free. Mailchimp-explicit rule (§voice-and-tone §Tone context-switching table).

For non-SaaS copy (product ads, long-form sales, キャッチコピー), the four contexts still map — e.g., PASONA's Affinity stage ≈ Onboarding tone; Offer stage ≈ Celebration tone (see voice-and-tone.md §Integration with PASONA / BEAF / キャッチコピー frameworks).

Pass 3 — Voice Anchor Selection (runs only per §When Pass 3 applies)

One linear 8-step flow. v1.13.0 collapsed the former 4-branch dispatch (Pass 3a/3b/3c/3d) into a single flow — named creator, axis-extreme landmark, and register-signal routing all traverse the same steps; differences are candidate-pool filters or Step 2.3 priority, not separate control paths.

Step 1 — Load

load standards/voice-anchor-meta.md
load standards/{output_language}-q{voice_quadrant.primary}-anchors.md §Landmark: {position}

{position} ∈ {center, extreme, toward-Q1..toward-Q4, axis-authority, axis-affinity, axis-reason, axis-emotion}. Axis-extreme landmarks live in voice-anchor-meta.md §Axis Extreme candidates (cross-language section); per-quadrant router holds center / extreme / toward-Q{N} sections.

Cross-lang opt-in: additionally load standards/{cross_ref_lang}-q{primary}-anchors.md §Landmark: {position} ONLY when:

• Some candidate anchor's frontmatter has cross-reference-valid-for[{output_language}] == STRONG
• brief.cross_lang_borrowing_allowed == true (default false)

Step 2 — Candidate pool

Read all anchor entries in the Landmark section. Apply three filters in order:

2.1 Automatic rejection — exclude anchors matching voice-anchor-meta.md §Automatic rejection:

• Voice inseparable from ideological / traumatic content
• Biographical-tragic weight overpowers style on bare name
• Corpus LLM-latent-space illegible
• Register non-transferable to commercial frame
• Corpus THIN in target language

2.2 Negation-of-axis — if brief.tone_cue contains explicit axis negation (e.g. "not corporate", "not preachy"):

1. Forbid the negated axis; exclude candidates whose primary axis is negated
2. Record forbidden axis in voice_quadrant.rationale (carry-forward) for Phase 8 verification
3. If post-filter candidate pool is empty → emit violation envelope to intake requesting tone_cue clarification

2.3 Named-creator forced rank 1 — if brief.voice_reference names a creator AND the corresponding anchor-{slug}.md is in the candidate pool:

• The named anchor is FORCED to rank 1 (primary). No override. User intent is authoritative.
• Agent still MUST evaluate fit relative to the brief. If fit_score (Phase 4 drafter)
• brief.output_language missing → bounce to copywriting-intake (Level 1 / 2 gap)

Input Envelope (consumed)

Shape inherited from copywriting-toolkit/CLAUDE.md §Handoff Envelope, post-Phase 5:

{
  "phase": "phase-5-voice-positioned",
  "form": "long-form-pasona | long-form-extended | mid-form | short-form | light-action | audit",
  "brief": { "voice_reference": "...", "output_language": "ja | en | zh-TW", "tone_cue": "...", "...": "..." },
  "message_thesis": "...",
  "voice_quadrant": {
    "primary": "Q1 | Q2 | Q3 | Q4",
    "edge": "Q2-Q3 | Q1-Q4 | null",
    "position": "center | extreme | toward-Q{N} | axis-{axis} | null",
    "rationale": "Phase 5 rationale",
    "schwartz_alignment": "ok | hard_rule_applied | conflict_flagged"
  },
  "draft": "the Phase 4 draft",
  "next_stage": "copywriting-voice-tone-stage"
}

Required keys: voice_quadrant, draft. Missing either → return to orchestrator with status: upstream-incomplete (Phase 5 must run first).

Output Envelope (produced)

{
  "phase": "phase-6-toned",
  "form": "long-form-pasona | ...",
  "brief": { "..." : "..." },
  "message_thesis": "...",
  "voice_quadrant": {
    "primary": "Q1 | Q2 | Q3 | Q4",
    "edge": "Q2-Q3 | Q1-Q4 | null",
    "position": "...",
    "rationale": "...",
    "schwartz_alignment": "ok | hard_rule_applied | conflict_flagged"
  },
  "draft": "the polished draft after 3 passes",
  "tone_notes": {
    "axis_target": { "formality": "casual", "seriousness": "serious", "respectfulness": "respectful", "enthusiasm": "warm" },
    "axis_default_derived": false,
    "tone_contexts_detected": ["onboarding", "celebration"],
    "tone_adjustments": [
      { "span": "sentence 3 of stage A", "context": "error", "before": "...", "after": "...", "rationale": "calm + non-deflecting" }
    ],
    "register_signal_applied": {
      "primary_anchor_slug": "...",
      "landmark_position": "center | extreme | toward-Q{N} | axis-{axis} | null",
      "named_master_forced_primary": true | false,
      "named_master_fit_warning": null,
      "mitigation_clauses_applied": ["..."],
      "anchor_candidates_ranked": [{"rank": 1, "slug": "...", "fit_score": "HIGH", "fit_reasoning": "..."}],
      "substitute_suggested": null,
      "secondary_anchors": [],
      "thesis_self_check": "clear | revised_once | escalate | not_applicable",
      "agent_selection_confidence": "HIGH | MEDIUM | LOW",
      "native_critical_vocab_cited": ["..."]
    },
    "ogilvy_flags": ["removed 'revolutionary' empty-hype token in stage S"]
  },
  "gate_verdict": "PASS | PASS_WITH_NOTES | NEEDS_REVISION",
  "gate_report": { "...see rubric §Output Format..." },
  "next_stage": "copywriting-ethics-check-stage"
}

Breaking changes from v1.12.x → v1.13.0:

Three tone_notes top-level fields are REMOVED and their semantics merged into register_signal_applied:

• tone_notes.lineage_applied → register_signal_applied.primary_anchor_slug
• tone_notes.lineage_gap → register_signal_applied.named_master_fit_warning (richer structured warning)
• tone_notes.axis_extreme_applied → register_signal_applied.landmark_position (axis landmarks flow through uniformly)

Downstream Phase 7/8 evaluators that read the old field paths must migrate. See rubrics/voice-consistency-gate.md for the updated Dimension 6 / Dimension 7 read paths.

Downstream (copywriting-ethics-check-stage) reads draft + tone_notes + brief. tone_notes.ogilvy_flags can surface 景品表示法 / FTC-relevant removals for Phase 7 cross-check.

Gate — SHOULD (Voice Consistency)

• Rubric: rubrics/voice-consistency-gate.md
• Agent: copywriting-toolkit/agents/copywriter-evaluator.md
• Trigger: multi-stage OR multi-candidate artifact (PASONA 6 stages, BEAF 4 stages, ideation N candidates, or a series). Per the rubric's §Scope Boundary: single-stage / single short-form only → gate is skipped with reason: cross-stage evaluation not possible recorded in the envelope.
• Position: runs AFTER all 3 passes, as the last voice-layer check before Phase 7 legal gates pick up the artifact.
• Dimensions (per rubric): brand voice stability, tone contextual appropriateness, anchor style fidelity (if Pass 3 ran), JP emotional-resonance vs Anglo benefit-clear fit, voice quadrant coherence, over-mimic adherence (scoped to Pass 3), and thesis alignment (scoped to Pass 3 — catches anchor-induced drift from envelope.message_thesis).

Verdict handling:

• PASS → emit envelope; next_stage: copywriting-ethics-check-stage.
• PASS_WITH_NOTES (≤2 WARNINGS, no FATAL) → apply named fixes in one auto-revise round, re-run gate; if still PASS_WITH_NOTES, emit with notes disclosed (handoff-format §Section 3 max-2-rounds rule).
• NEEDS_REVISION (any FATAL or ≥3 WARNINGS) → BLOCKED. Return to the specific pass the rubric's next_action names. Do not forward to Phase 7.

Anti-Patterns

• Rewriting framework stage structure during voice tuning. PASONA A / S / O boundaries are not Phase 6's territory.
• Swapping voice mid-artifact to match tone context ("error → serious voice"). voice-and-tone.md §Anti-Patterns: voice stays constant, only tone shifts.
• Transplanting signature mechanics across traditions (JP ↔ ZH ↔ EN). Cross-transplant ban is absolute; cross-language BORROWING (via STRONG cross-reference + brief opt-in) is the only permitted cross-tradition path.
• Re-running Step 2 voice-anchor-meta.md 4-condition validation at runtime. Library curation (scripts/lint-anchor-library.py) owns baseline eligibility; runtime trusts the library.
• Overriding Step 2.3 named-creator forced rank 1. User intent is authoritative; agent emits named_master_fit_warning but does NOT reassign primary.
• Filling the Voice Guide axes abstractly ("friendly, professional") with no side-by-side on-brand / off-brand examples (voice-and-tone.md §Brand Voice Guide checklist).
• Headline breaks voice for attention, then body "returns" — trust already lost at headline layer.
• Swallowing a NEEDS_REVISION gate verdict and forwarding to Phase 7.

**Drift

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kouko
• Source: kouko/monkey-skills
• License: MIT

Install and usage instructions live in the source repository linked above.