Gary Copy Production Orchestrator

Gary 文案生产总控

Use this skill when the user asks to process a Gary male-account topic batch, send a batch for expansion, redo a batch after route drift, or connect knowledge-base transplant, Gemini expansion, opening matching, and acceptance.

This is an orchestration layer. It does not replace the specialist skills:

• Read huanxin for topic/knowledge route design and mechanism deduplication.
• Read 男版扩写 and douyin-copy-production-workflow-custom for Gemini expansion and version contracts.
• Read baokuan-kaitou-sheding for learned opening matching. Never hand-write a production opening from memory.
• Read humanize-copy only when the user explicitly asks for AI-tone cleanup after content acceptance.

Mandatory sequence

1. Freeze the current request. Identify account, batch, titles, version, and whether the user wants topic work, source-route work, expansion, opening matching, or acceptance. Treat 2.8 as ordinary 2.8 Safe Draft unless the user explicitly names another route.
2. Preflight the batch. Preserve the user's exact titles. Check topic fingerprint, promised count, mechanism, result, relationship relevance, and within-batch/history similarity. Build one route per article with one primary mechanism and at most one supporting mechanism.
3. Build the short source. Use traceable knowledge cards for mechanism explanation and case material. Keep the source lean and public-facing. Do not write 知识库案例, source labels, prompt instructions, or backstage process language into the manuscript. Do not expose knowledge IDs to Gemini.
4. Run one calibration article first. For a new batch or after a route failure, expand only one article. Send the complete selected Gemini prompt plus the article-specific route lock and source. Do not batch-run the remaining articles until the user explicitly approves the test direction.
5. Keep body and opening separate. Gemini's first paragraph is not the formal opening. After the body route is accepted, use baokuan-kaitou-sheding to choose an existing approved template by topic fingerprint, count, action, and result. Keep fixed template text unchanged; replace only permitted variables. Do not let a generic 2.8 hook become the production opening.
6. Batch only after approval. Once the user says the calibration is acceptable, apply the approved route method to the remaining articles, with separate mechanism locks and anti-cross-contamination rules for each.
7. Accept narrowly. For ordinary 2.8, require at least 3000 Chinese characters, promised points, a complete natural case before the mid-body CTA, natural professional terminology, the fixed ending, and no backstage leakage. Repair only definite wording, grammar, TTS, or formal-label defects. Do not redesign an otherwise complete article without user instruction.
8. Learn every manual correction. When the user rewrites a title or opening, preserve the assistant original, the user's exact version, the differences, and the transferable reason in a dated knowledge-base record. Put only the generalized rule in the relevant rule page or skill reference. Approval without a manual rewrite is a positive candidate signal, not a universal rule.
9. Verify the learning loop. Before the next batch, consult the newest correction pairs and run a regression check for mechanism, result, count, title skeleton, forbidden wording, and within-batch similarity. Never claim a correction was learned unless both the exact pair and its generalized rule have a recorded location.

User-calibrated writing rules

• Preserve the exact public topic and promised number. 六个步骤 or 六个节点 is usable when the body contains executable actions; reject vague structural labels that do not create operational value.
• The user's accepted opening logic prioritizes: strong judgment or observable condition -> low-cost learnable action -> strong male result -> professional topic/mechanism later. Weak benefits such as 理解心理, 提升认知, or 愿意回复 cannot carry the hook.
• Do not use 礼貌回应 or similar weak intermediate outcomes as the main opening reward. Avoid 挑选你, 见面前, and 普通聊天 when they weaken the intended relationship stage.
• Do not reuse the generic body opening skeleton 女人拿走了什么 / 你留下了什么 / 全砸在你身上 / 卡在这个死局里 across articles. Choose one topic-native entry: a concrete scene, one hard judgment, one case action, or one immediate conflict.
• Separate near mechanisms. Projection is not Barnum self-disclosure; intermittent reinforcement is not expectation design; relationship-battle rules are not random rewards; self-iteration is not a generic self-improvement list.
• Cases are evidence, not point substitutes. They must naturally convey the man's initial problem, the adjustment, observable female feedback, and the result; never expose QA field labels in public copy.
• AI-tone cleanup is local. Remove repeated formulaic phrases and rigid transitions without weakening force, relationship specificity, or valid mechanism vocabulary.

State boundaries

Record and report these states separately: source-route-ready, calibration-generated, user-approved, batch-generated, 正文待确认, opening-matched, Word-exported, and published. A generated file is never silently treated as accepted.

Stop conditions

• Stop after the calibration article when user approval is missing.
• Stop and report a route failure when the article drifts from its mother topic, promised count, primary mechanism, or relationship relevance.
• Do not rerun a full batch merely for local wording defects; repair narrowly after acceptance review.

Read the relevant specialist skill and the user-calibrated reference before acting. The companion reference is references/user-calibrated-rules.md.