translation-creative

name: translation-creative description: Translate ad copy / marketing brief / headline / catchphrase via faithful or transcreation mode. 5D reflection in transcreation. S1 back-translation = MUST in transcreation, SHOULD in faithful. Variants opt-in via --variants=N. version: 0.1.0

Layer 2-5 owner for ad copy / marketing brief / headline / tagline / catchphrase in translation-toolkit. Reads an intake-spec produced by translation-intake (Layer 1), runs an optional brand-brief intake, then runs the shared DRAFT / REFLECT / IMPROVE core loop with mode-conditional behavior: faithful mode = 4D reflection + S1 SHOULD; transcreation mode = 5D reflection (adds Effectiveness axis) + S1 MUST. Default emits a single output; --variants=N opts into N genuinely different alternatives.

Inputs

source — ad copy / headline / tagline / marketing-brief string, or a path to a short text file.
intake-spec — from translation-intake. If invoked directly without one, run intake first (--intake flag) before continuing. Intake-spec carries mode (faithful | transcreation) — the load-bearing field for this skill.
brand brief (optional but strongly recommended for transcreation) — captured per protocols/brand-brief-intake.md (archetype, tone-of-voice, do-not-say list, signature phrases, target persona, CTA style, brand-name handling). Skill-local; not part of the four shared service-interface fields.
glossary_path (optional) — repo-local project glossary (L1 tier of the 4-tier resolver). Defaults to <caller_repo>/docs/i18n/glossary-{target_locale}.md; missing → fall through to L2 bundled glossary in glossary/.

Service-interface contract (the four shared fields plus web_search) is defined in the design spec §Service Interface; no per-skill duplication here.

When to use

Routed in by using-translation-toolkit when the input is ad-shaped (short, persuasive, headline-like) or when the user explicitly names ad copy / marketing / catchphrase / tagline / brand voice.
Invoked directly by name when the user already knows creative is the right specialist.
Not for i18n format files (use translation-i18n), prose / Markdown docs (use translation-doc), or reviewing an existing translation pair (use translation-audit).

Modes

Mode is declared in the intake-spec and controls REFLECT axis count + S1 gate tier:

faithful — preserves source structure + sentence boundaries with target-natural phrasing. 4D REFLECT (per references/prompt-reflect-4d.md). S1 = SHOULD (WARN on similarity < 0.85, output proceeds).
transcreation — departs from literal surface to land equivalent persuasion in the target culture. Pun / rhyme / cultural-anchor substitution allowed and often required. 5D REFLECT (4D + Effectiveness axis, per references/prompt-reflect-5d.md and references/5d-effectiveness.md). S1 = MUST (FAIL on similarity < 0.70, output blocked). See protocols/transcreation-mode.md for the full mode-entry contract.

If the intake-spec does not specify mode explicitly, the router defaults ad-shaped genres to transcreation and prose-leaning marketing-brief content to faithful; the upstream translation-intake records the resolved mode + reason in the spec.

Web search policy

ON by default per using-translation-toolkit §"Web search trade-off note" (spec Decision #15) — creative work benefits from current cultural / campaign context. Override to OFF when working with an established brand voice that risks contamination from competitor copy:

--web-search=off

When OFF, glossary resolution stops at L2 (bundled) — L3 (web) is skipped, L4 (LLM-fallback) still runs. Effectiveness-axis critique still operates from training-time cultural knowledge; transcreation does not require web search but loses freshness on recent slogans / memes / campaign references.

Pipeline

This skill executes Layers 2-5 of the toolkit's 5-layer pipeline. (Layer 1 — intake — is the upstream translation-intake skill, augmented here by the optional brand-brief intake in Layer 2 step 1.)

Layer 2: Preparation

Brand brief intake. Per protocols/brand-brief-intake.md. Captures brand archetype (Hero / Sage / Outlaw / etc.), tone-of-voice axes (authoritative ↔ playful, formal ↔ casual, warm ↔ cool), do-not-say list, signature phrases (verbatim-preserve vs locale-transcreate decision), target persona, CTA style, and per-locale brand-name handling. Recommended for transcreation; optional for faithful (faithful mode falls back to intake-spec register + intent fields if no brief is provided). Brief outputs land in the audit-trail brand_brief block.
Protect-pass. Mask any structural spans with ⟦P:NN⟧ sentinels per references/protect-pass-spec.md (8 base classes — ICU plural, curly braces, printf, fenced code, inline code, HTML, URL, email). Ad copy is rarely placeholder-heavy, but URLs (campaign-link UTMs), HTML (rich-text email creative), and brand-name tokens still occur and MUST be masked when present.
Source analysis. Per chunk, extract: cultural references (idioms, allusions, anchor figures), wordplay (puns, rhymes, alliteration), CTA verbs + intensity, and untranslatability candidates (concepts with no clean target equivalent — flagged for I1). Feeds into glossary resolve, the WRITER prompt, and the Effectiveness axis in transcreation mode.
Glossary resolve. 4-tier fallthrough per references/orthogonal-axes.md and the plugin-level docs/glossary-format-spec.md:
1. L1 project — <repo>/docs/i18n/glossary-{target_locale}.md (caller-supplied, highest priority)
2. L2 bundled — glossary/glossary-{source}--{target}.md (Pontoon / GNOME / JLT / e-Stat / Tokyo / Cabinet / NAER ingested at build time)
3. L3 web search — only if web search is ON; cite URL in audit-trail
4. L4 LLM fallback — model proposes a translation; flagged in audit-trail with tier L4 and lowest trust
Only matched terms whose source forms appear in the active chunk flow into the WRITER / CRITIC / REVISER prompts — no full-glossary dump (see references/core-loop.md §8).

Layer 3: Core loop

Three roles, one LLM call each per chunk per pass, per references/core-loop.md:

DRAFT (WRITER) — references/prompt-draft.md. Receives brand-brief context (when present) alongside intake-spec axes; instructed to honor signature phrases + do-not-say list.
REFLECT (CRITIC) — axis count is mode-conditional:
- faithful mode → 4D per references/prompt-reflect-4d.md (Accuracy / Fluency / Style / Terminology)
- transcreation mode → 5D per references/prompt-reflect-5d.md (4D + Effectiveness, defined in references/5d-effectiveness.md). Effectiveness judges whether the target preserves the persuasion intent in target culture (CTA strength, cultural-reference landing, emotional pull, taboo avoidance, brand-voice consistency, anchor-figure substitution, genre-convention fit, phonetic / rhythm preservation, claim-framing convention).
IMPROVE (REVISER) — references/prompt-improve.md. Consumes the 4D or 5D critique and emits v2; preserves all ⟦P:NN⟧ tokens and any signature phrases marked verbatim-preserve in the brand brief.

WRITER and REVISER outputs MUST preserve every ⟦P:NN⟧ token exactly (same count, same IDs). CRITIC produces structured JSON critique only — never a rewrite.

Layer 4: Verification

Per references/verification-gates.md §"Per-skill gate application". translation-creative's defining trait is the mode-conditional S1 tier flip — back-translation is SHOULD in faithful but promoted to MUST in transcreation, the only place in the toolkit where a tier flip drives a HARD gate.

Gate	Tier	Action
M1	HARD	Placeholder integrity — count + ID set parity for any URL / HTML / brand-token sentinels between source and target. Deterministic regex check.
M2	HARD	Project glossary compliance — every L1-mandated source term renders as its mapped target form (PASS_ADVISORY for `notes: context-dependent`). In transcreation mode see `protocols/transcreation-mode.md` §"Glossary leeway" — culturally-driven violations are auditable but allowed. Deterministic lookup.
S1	MUST in transcreation, SHOULD in faithful	Back-translation — subagent dispatches a blind v2 → source retranslation; embedding-cosine similarity vs original source. Threshold = 0.85 in faithful, 0.70 in transcreation (relaxed, per `references/verification-gates.md` §S1). MUST in transcreation = similarity below threshold blocks output as FAIL (not WARN). SHOULD in faithful = similarity below threshold surfaces a WARN to the caller, output proceeds. Skipped with audit-trail flag `S1: SKIPPED (no isolation capability)` if the runtime provides no subagent dispatch.
S2	SHOULD	Register preservation — JUDGE classifies source vs target register on a discourse / formality axis. Reliable for short ad-copy chunks because brand brief + intake-spec pin the register tightly; mismatch surfaces a WARN.
I1	INFO	Untranslatability flagging — runs only when Layer 2 source-analysis flags a phrase. Non-interactive: records borrow / explain / approximate decisions in the audit-trail; never blocks, never prompts. Particularly active in transcreation where cultural references frequently force an explicit borrow / substitute decision.

S1's mode-conditional tier is the only case in the toolkit where a SHOULD becomes MUST. The reasoning, per spec Decision #4: in transcreation the WRITER is licensed to deviate substantially from the source surface, so M1 (placeholder count) and S2 (register) are insufficient guards against outright topic drift — S1 is the only gate that catches "the v2 is well-written copy about a different product". Promoting it to MUST in transcreation closes that hole; keeping it SHOULD in faithful avoids over-blocking on legitimate stylistic restructuring.

Layer 5: Output

Restore placeholder tokens. Replace every ⟦P:NN⟧ in the translated text with the original substring from the protect-pass token map. Tokens absent from the translation surface as M1 failures (already caught in Layer 4) — restore is a clean swap at this point.
Default emit: 1 translation. A single best-effort target-locale rendering, paired with the audit-trail block.
--variants=N opt-in: emit N genuinely different alternatives. Each variant is a full, independent DRAFT → REFLECT → IMPROVE run with the WRITER prompt instructed to vary on a tactical axis (per protocols/transcreation-mode.md §"How variants differ"): variant 1 might preserve source structure; variant 2 might restructure for natural target rhythm; variant 3 might substitute a culturally equivalent metaphor. Variants are NOT paraphrases of a single draft — that pattern is explicitly forbidden because it produces synonym-swap noise instead of strategically different options.
Emit audit-trail JSON. Schema per references/audit-trail-spec.md. Records: brand brief (when present), every chunk's (source, draft, reflect, improve) quadruple per variant, every gate verdict with diff, every glossary resolution with tier + audit_path, every untranslatability decision, every transcreation-mode glossary-leeway override with rationale, and a variant_index per chunk when --variants=N was used.

Variants and consistency

Distinguishing trait of translation-creative vs the other format specialists: variants are a first-class output mode, not paraphrase noise. Reasoning:

Ad copy benefits from A/B candidates; downstream creative review is human and prefers strategically different starting points to picking between near-identical wordings.
Each variant runs its own full core loop (3 LLM calls + gates) because shared draft → forked critique would produce critique that does not actually fit the variant being judged.
The variant_index field in the audit trail lets downstream review tools (or translation-audit if pointed at a variant set) attribute issues to the specific variant.
If a single variant fails S1 in transcreation mode (FAIL), that variant is dropped from the emitted set. If all N variants fail S1, the run halts and surfaces failure (rather than silently emitting fewer than N).

Roles

Same vocabulary as the rest of the toolkit (per using-translation-toolkit §Roles vocabulary):

WRITER — produces the draft, preserves ⟦P:NN⟧ tokens, honors brand-brief signature phrases + do-not-say list.
CRITIC — 4D (faithful) or 5D (transcreation) structured JSON critique; no rewrites.
REVISER — consumes critique, outputs v2; preserves placeholders + signature phrases.
BACK-TRANSLATOR — blind v2 → source retranslation, used by S1 (MUST in transcreation, SHOULD in faithful).
JUDGE — register classification, used by S2.

Roles are behavioral. Any LLM model can fill any role; this skill specifies behavior, not models.

Cross-plugin composition

copywriting-toolkit can be invoked after creative translation for additional tone polish (e.g., voice-quadrant repositioning, ethics gates, framework-fit checks). This is user-explicit composition only — translation-creative does NOT auto-invoke copywriting-toolkit. Users who want the chain run it as two skill invocations:

1. translation-creative ... → creative-translated copy + audit-trail
2. copywriting-toolkit (using-copywriting-toolkit) ← creative-translated copy as input

translation-creative does not duplicate copywriting-toolkit's voice / form / ethics gates; the two skills are complementary and run sequentially when the user asks.

What this skill does NOT do

Does not run intake. If invoked without an intake-spec, hand off to translation-intake first (or use --intake to run it inline).
Does not generate brand strategy. The brand brief intake captures existing brand decisions; it does not invent archetype / tone / persona from scratch. Missing brief in transcreation mode produces a WARN, not a generated strategy.
Does not replace human creative review. Creative output ships to a human reviewer downstream — the audit trail + variant index are designed to make that review fast.
Does not translate brand names without explicit intake instruction. Brand names default to verbatim-preserve unless the brand brief explicitly specifies translate / transliterate (katakana / 音譯 / hangul) / locale-substitute. Audit trail records the decision.
Does not auto-invoke copywriting-toolkit. Composition is explicit (see above).
Does not paraphrase a single draft into variants. --variants=N runs N independent core loops with axis-differentiated prompts (see Layer 5 step 3).
Does not bypass M1 / M2 in faithful mode. No --bypass-gates flag exists (anti-pattern per spec Decision #15). Fix the underlying issue and re-run.
Does not bypass S1 in transcreation mode. When S1 is MUST, a sub-threshold v2 blocks output; the resolution is to revise (re-run the loop with critique focus on the drift) or escalate to a human translator, NOT to flip the gate off.
Does not prompt the user during I1. Untranslatability decisions are recorded, not asked.

translation-creative

name: translation-creative description: Translate ad copy / marketing brief / headline / catchphrase via faithful or transcreation mode. 5D reflection in transcreation. S1 back-translation = MUST in transcreation, SHOULD in faithful. Variants opt-in via --variants=N. version: 0.1.0