copywriting（文案写作）

Copywriting — Orchestrator

*Communication — Horizontal. Coordinates specialized sub-agents to produce craft-quality copy with annotations, alternatives, and quantitative evaluation.*

Core Question: "Is every key line visual, falsifiable, and uniquely ours?"

Philosophy

Copy frameworks (PAS, 3-Question Test, CTA formula) are proven tools — not mandatory templates. A skilled agent may combine frameworks or break a guideline when the extra words earn their place. The test is: does this stop the scroll and move the reader? This orchestrator dispatches specialist agents for each concern, then a critic agent ensures every key line meets the bar.

Inputs Required

• Brief: what to write (headline, hook, CTA, tagline, subject line, or full-page copy)
• Context: audience, product, awareness stage (from research/icp-research.md, research/product-context.md, or user-provided)

Output

• Inline annotations + alternatives on key lines (when called by another skill)
• .agents/mkt/content/[slug].copy.md (for standalone full-page copy tasks)

Quality Gate

Before delivering, the critic agent verifies:

• Every key line passes the Three-Question Test: visual? falsifiable? uniquely ours?
• Rubric score averages ≥3.5 across V/F/U for all key lines
• Every key line passes the Competitor Swap Test (swap in competitor name — if it still works, rewrite)
• 3-5 variations generated per key line, best selected with top 2-3 presented as alternatives
• Every key line annotated: rule that drove the choice, cut alternative, rubric score
• CTA follows formula: [action verb] + [what they get] (not "Learn More" or "Click Here")
• Every headline/hook contains concrete nouns or specific numbers (no abstract "better," "innovative," "leading")

Chain Position

Horizontal — called by content-create, lp-optimization, imc-plan. Can run standalone. Re-run triggers: When brand voice changes, when A/B test results suggest a different angle, or when key lines need fresh variations.

Skill Deference

• Need a full content asset (carousel, thread, email, video)? → Use content-create
• Content reads as AI-generated? → Run humanize after
• Diagnosing a live page's conversion problem? → Use lp-optimization
• Optimizing for search/AI citations? → Coordinate with seo

Agent Manifest

Shared References (read by multiple agents)

• references/headline-formulas.md — Headline formula catalog (used by hook-agent)
• references/page-sections.md — Page section specs and testimonial criteria (used by body-agent, social-proof-agent)

Routing Logic

Classify the task, then follow the matching route.

Route A: Single Key Line

When: Brief asks for a headline, hook, CTA, tagline, or subject line — not a full page.

1. Pre-dispatch: Gather context (Step 0 below)
2. Dispatch ONE agent:
   - Hook/headline/tagline/subject line → hook-agent
   - CTA → cta-agent
3. Dispatch: critic-agent (on the single agent's output)
4. If critic returns FAIL → re-dispatch the original agent with feedback (max 2 cycles)
5. Deliver annotated key lines

Note: Route A skips variant-agent. Hook-agent and cta-agent already generate 3-5 variations internally with rubric scoring. Variant-agent is designed for full-page copy where cross-section A/B alternatives are needed.

Route B: Full-Page Copy

When: Brief asks for landing page copy, full-page copy, or multiple sections.

1. Pre-dispatch: Gather context (Step 0 below)
2. LAYER 1 — Dispatch IN PARALLEL:
   - hook-agent
   - body-agent
   - cta-agent
   - social-proof-agent
3. MERGE: Assemble Layer 1 outputs into page structure
4. Dispatch: variant-agent (on merged output)
5. LAYER 2 — Dispatch SEQUENTIALLY:
   - voice-agent (receives merged + varianted output)
   - psychology-agent (receives voice-agent output)
   - zero-risk-agent (receives psychology-agent output)
6. Dispatch: critic-agent (receives zero-risk-agent output)
7. If critic returns FAIL → re-dispatch named agent(s) with feedback (max 2 cycles)
8. Deliver final artifact

Route C: Called by Another Skill

When: Invoked by content-create, lp-optimization, or imc-plan for inline copy work.

1. Pre-dispatch: Read context from calling skill's artifacts
2. Dispatch the relevant Layer 1 agent(s) based on what the caller needs:
   - Headline/hook → hook-agent
   - Body sections → body-agent
   - CTAs → cta-agent
   - Social proof → social-proof-agent
3. Dispatch: critic-agent
4. Return annotated output to the calling skill

Step 0: Pre-Dispatch Context Gathering

Before dispatching any agent, the orchestrator gathers context that ALL agents will need.

Product Context Check

Check for research/product-context.md. If available, read for product details, voice adjectives, and accuracy constraints. If research/product-context.md or research/icp-research.md date fields are older than 30 days, warn the user and recommend re-running icp-research before proceeding. This is a soft gate — proceed if the user confirms, but note "stale ICP data" in the artifact header. Tip: /navigate status (from meta-skills) gives a single-pass freshness report across all upstream artifacts.

Language

Default: English. If the user specifies another language, note it in pre-writing and pass to all agents. All agent instructions are currently optimized for English copy — other languages may need adapted idioms and cultural references.

Optional Artifacts

Pre-Writing Framework

Answer these 4 questions before dispatching. Pass the answers to every agent as the pre-writing input:

1. Who am I talking to? What do they currently believe? What language do they use?
2. What should they believe after reading this? What's the one shift?
3. What can I say that nobody else can? What's our unique proof?
4. Where is the traffic coming from? (Ad, search, email, social, direct) — determines what they already know and expect.

If research/icp-research.md exists, pull VoC quotes and pain language. Write how the buyer talks, not how the brand talks.

If .agents/mkt/content-research.md exists, pull the Audience Language Map (brand language → audience language pairs) and Winning Patterns (hook types that are performing in this niche). Feed these into pre-writing as research_language and research_patterns for hook-agent and voice-agent.

Dispatch Protocol

How to spawn a sub-agent

For each agent dispatched below, use the Agent tool with a prompt constructed as follows:

1. Read the agent instruction file (e.g., agents/hook-agent.md) — include its FULL content in the Agent prompt
2. Append the brief and pre-writing context after the instructions
3. Resolve file paths to absolute: replace relative paths with absolute paths rooted at this skill's directory. Example: if this skill is at /Users/you/skills/copywriting/, then references/headline-formulas.md becomes /Users/you/skills/copywriting/references/headline-formulas.md. Tell the agent: "Read the reference file at [absolute path] for domain knowledge."
4. Pass upstream artifacts by content, not path: the orchestrator reads research/product-context.md, research/icp-research.md, and .agents/mkt/content-research.md (if it exists) FIRST, then includes relevant excerpts (VoC quotes, voice adjectives, pain language, research-backed audience language and winning patterns) in the pre-writing object. Sub-agents should NOT read artifact files directly — the orchestrator curates what they need.
5. If feedback exists (from a critic FAIL cycle), append it at the end of the prompt with the header "## Critic Feedback — Address Every Point"

Single-agent fallback

If multi-agent dispatch is unavailable (no Agent tool, single-agent runtime, or context constraints), execute each agent's instructions sequentially in-context:

• Layer 1: run each agent's domain instructions one at a time, writing each section
• Layer 2: apply each refinement pass to the full document in order
• Critic: self-evaluate using the critic-agent's rubric and quality gate

The output quality should be equivalent — the multi-agent pattern optimizes for parallelism and focus, not capability.

Layer 1: Parallel Section Writers

Spawn the following agents IN PARALLEL (multiple Agent tool calls in a single message). For each agent, follow the Dispatch Protocol above.

For single key line tasks (Route A): Dispatch only the relevant agent, not all four.

Merge Step

After all Layer 1 agents return, assemble their outputs into the page structure.

Section Order (full-page copy)

1. Hero — Hook agent's recommended headline + subheadline + CTA agent's hero CTA
2. Social Proof Bar — Social proof agent's bar element
3. Problem — Body agent's Problem section
4. Solution — Body agent's Solution section
5. How It Works — Body agent's How It Works section
6. Testimonials — Social proof agent's testimonials
7. Key Stats — Social proof agent's stats (if applicable)
8. Final CTA — CTA agent's final CTA + risk reversal

Assembly Rules

The merge is deterministic assembly, not creative synthesis. Slot each agent's output into the template by ownership:

Conflict Resolution

• Each agent owns specific sections (table above). If two agents mention the same fact (e.g., both hook and social-proof reference "12,000 users"), keep the version from the section owner.
• If hook-agent's headline contradicts body-agent's problem framing, the orchestrator adjusts the body to align with the hook — the hook is the anchor because it's what the reader sees first.

After Merge: Variant Agent

Dispatch the variant-agent with the assembled document as upstream. It returns A/B alternatives for key sections. Variant alternatives are appended to the artifact (not replacing originals) — they're testable options, not replacements.

Layer 2: Sequential Craft Refiners

Dispatch these agents ONE AT A TIME, IN ORDER using the Dispatch Protocol above. Each receives the previous agent's full output as the upstream field.

voice-agent → psychology-agent → zero-risk-agent → critic-agent

Each agent returns the full document with their edits applied + a change log.

Critic Gate

The critic agent returns one of two verdicts:

PASS

The copy meets all quality standards. Deliver the critic's annotated output as the final artifact.

FAIL

The critic returns specific failures with:

• Which lines failed and on which dimension
• Specific fix instructions
• Which agent to re-dispatch

Rewrite loop:

1. Read the critic's failure report
2. Re-dispatch ONLY the named agent(s) with the critic's feedback attached as the feedback input
3. Run the modified output back through the critic
4. Maximum 2 rewrite cycles. After 2 failures, deliver the copy with the critic's annotations and flag to the user: "Copy scored [X] — manual review recommended on [specific lines]."

Artifact Template

---
skill: copywriting
version: 1
date: [today's date]
status: draft
---

# Copy: [Title / Brief Description]

**Date:** [today]
**Skill:** copywriting
**Audience:** [who]
**Awareness Stage:** [stage]
**Traffic Source:** [where they're coming from]

## Pre-Writing

1. **Talking to:** [audience + current belief]
2. **Shift to:** [desired belief after reading]
3. **Only we can say:** [unique proof/angle]
4. **Traffic context:** [what they already know]

## Key Lines

### [Line Type: Headline / Hook / CTA / Tagline / Subject Line]

**Selected:** "[winning line]"
  Rule: [which principle]. Score: V:[n] F:[n] U:[n].
  Cut alternative: "[runner-up]" — [why cut].

**Alternative A:** "[second option]"
  Rule: [principle]. Score: V:[n] F:[n] U:[n].

**Alternative B:** "[third option]"
  Rule: [principle]. Score: V:[n] F:[n] U:[n].

## [Additional sections for full-page copy — Hero, Problem, Solution, etc.]

## A/B Variants
[Variant agent's alternatives with hypotheses and test priority]

> On re-run: rename existing artifact to `[slug].copy.v[N].md` and create new with incremented version.

Next Step

Run humanize to refine voice and compress. Run lp-optimization if this copy is for a landing page. Run attribution to track performance.

Worked Example — Full-Page Copy (Route B)

Brief: Landing page for StatusZero, a tool that replaces standups with async status updates. Audience: Engineering managers at 50-200 person companies, problem aware. Traffic: LinkedIn ads (cold).

Step 0: Pre-Dispatch

1. Talking to: EMs who feel buried in status overhead but think it's part of managing.
2. Shift to: Status updates are a solvable problem, not a cost of doing business.
3. Only we can say: Internal data shows 12 hrs/week lost to status theater.
4. Traffic: LinkedIn ad click — cold, haven't searched, hook must interrupt.

Layer 1: Parallel Dispatch

→ Hook agent returns 5 headline variations, recommends: "Your team loses 12 hours a week to status updates nobody reads." (V:5 F:5 U:4) → Body agent returns Problem (status theater pain), Solution (async updates), How It Works (3 steps) → CTA agent returns hero CTA ("See how teams ship without status theater →"), mid-page, and final CTAs with risk reversal → Social proof agent returns logo bar (4 tech companies), 2 testimonials with metrics, key stat

Merge (orchestrator assembles)

Slots Layer 1 outputs into template: Hero → Social Proof Bar → Problem → Solution → How It Works → Testimonials → Final CTA

Dispatch: Variant Agent

Receives merged document. Returns A/B alternatives: → Hero variant B: "I cut meetings 80%. Output doubled." (tests story hook vs. data hook) → Final CTA variant B: "Try StatusZero free — no credit card, cancel anytime" (tests specificity vs. risk-first)

Layer 2: Sequential Dispatch

→ Voice agent receives merged + varianted doc. Fixes 3 AI patterns, adjusts tone to "direct, technical." Returns modified doc + change log. → Psychology agent receives voice output. Adds "which means..." bridges to 2 features, strengthens emotion in problem section. Returns modified doc + change log. → Zero-risk agent receives psychology output. Adds "No credit card required" near hero CTA, "Cancel anytime" near final CTA. Returns modified doc + change log.

Dispatch: Critic Agent → PASS

Receives zero-risk output. Scores all key lines. All average ≥3.5. Annotated and delivered.

Agent Files

Sub-Agent Instructions (agents/)

• agents/hook-agent.md — Headlines, hooks, 3Q scoring, formulas
• agents/body-agent.md — Problem, Solution, How It Works sections
• agents/cta-agent.md — CTA formula, risk reversal, placement
• agents/social-proof-agent.md — Testimonials, stats, credibility
• agents/variant-agent.md — A/B alternatives, rubric scoring
• agents/voice-agent.md — Clarity, voice consistency, AI slop removal
• agents/psychology-agent.md — So What, Prove It, Specificity, Emotion
• agents/zero-risk-agent.md — Barrier removal, guarantees, exit grace
• agents/critic-agent.md — Final evaluation, scoring, PASS/FAIL
• agents/_template.md — Reusable template for creating new agent files

Shared References (references/)

• references/headline-formulas.md — Headline formula catalog
• references/page-sections.md — Page section types, templates, testimonial selection

Page-Specific Guidance

Anti-Patterns

Skipping the pre-writing — Dispatching agents without answering the 4 pre-writing questions. Every agent needs this context to write relevant copy.

Dispatching all agents for a single headline — Route A exists for a reason. A single key line only needs one Layer 1 agent + critic.

Ignoring the critic's FAIL — If the critic fails the copy, the orchestrator MUST re-dispatch. Delivering failed copy to the user breaks the quality contract.

Re-dispatching the wrong agent — The critic specifies which agent to re-dispatch. Read the failure report carefully.

More than 2 rewrite cycles — Diminishing returns. After 2 cycles, deliver with annotations and let the user decide.