content-create（内容创建）

Content Creation — Orchestrator

*Communication — Step 3 of 4. Coordinates specialized agents to turn IMC angles into production-ready content assets with A/B variants.*

Core Question: "Would the target persona actually stop scrolling for this?"

Critical Gates — Read First

• Do NOT write content without format specs. Platform constraints (character limits, dimensions, native patterns) must be resolved FIRST. Hook/body/CTA agents cannot write without them.
• Do NOT dispatch all agents for a single social post. Route A exists for quick assets. Only Route B uses the full agent stack.
• Do NOT write outlines. Every agent must produce complete copy — every word, slide, timestamp. "Discuss benefits" is not a deliverable.
• Stale IMC data (>30 days) produces misaligned content. Recommend re-running imc-plan before proceeding if artifact dates are old.
• Platform compliance is non-negotiable for regulated channels. SMS requires TCPA/GDPR opt-in/opt-out. Store listings must follow Apple/Google metadata policies. Forum posts must respect community rules. OOH materials need legal disclaimers. Email requires CAN-SPAM/GDPR compliance.

Philosophy

This skill focuses on content assets — choosing the right format, structuring for the platform, and producing deliverables. For copy craft (variation workflows, evaluation rubrics, annotation), use copywriting.

Inputs Required

• Angle + channel from .agents/mkt/imc-plan.md (or user-provided)

Output

• .agents/mkt/content/[slug].md

Quality Gate

Before delivering, the critic agent verifies:

• Hook ≤8 words or within platform character limit
• Key lines pass 3Q test (Visual, Falsifiable, Uniquely ours)
• CTA follows formula: [action verb] + [what they get]
• A/B variant changes exactly ONE element with testable hypothesis
• Complete copy — every word written, no outlines or placeholders
• Format compliance — all platform specs met
• VoC language present — buyer words, not brand words

Chain Position

Previous: imc-plan | Next: attribution Re-run triggers: When IMC plan angles are updated, when targeting a new platform, or when A/B test results suggest a new direction.

Skill Deference

• Need craft-quality headlines or page copy? → Run copywriting for variation workflow and rubric.
• Page exists, problem is conversion? → Run lp-optimization first.
• Content reads as AI-generated? → Run humanize after.
• Optimizing for search/AI citations? → Coordinate with seo.

Agent Manifest

Shared References (read by multiple agents)

• references/platform-specs.md — Platform dimensions, character limits, native patterns
• references/examples.md — Complete worked examples with full copy
• references/repurposing-cascade.md — Hero → derivative content workflow

Routing Logic

Route A: Single Asset (Quick)

When: One social post, one email, or one ad — not a multi-asset campaign.

1. Pre-dispatch: Gather context (Step 0)
2. LAYER 1 — Dispatch IN PARALLEL:
   - format-agent (determines platform constraints)
   - voc-extraction-agent (pulls buyer language)
3. LAYER 1.5 — Dispatch IN PARALLEL (after format-agent returns):
   - hook-agent (receives format specs)
   - cta-agent (receives format specs)
4. Orchestrator writes body inline (single asset doesn't need full body-agent)
5. Dispatch: platform-compliance-agent
6. Dispatch: critic-agent
7. If FAIL → re-dispatch named agent(s) with feedback (max 2 cycles)
8. Deliver artifact

Route B: Full Asset (Campaign or Complex)

When: Multi-slide carousel, video script, email sequence, or multi-asset campaign.

1. Pre-dispatch: Gather context (Step 0)
2. LAYER 1 — Dispatch IN PARALLEL:
   - format-agent
   - voc-extraction-agent
3. LAYER 1.5 — Dispatch IN PARALLEL (after format-agent returns):
   - hook-agent (receives format specs + VoC)
   - body-agent (receives format specs + VoC)
   - cta-agent (receives format specs + VoC)
4. MERGE: Assemble hook + body + CTA into unified content asset
5. LAYER 2 — Dispatch SEQUENTIALLY:
   - platform-compliance-agent (receives merged asset)
   - ab-variant-agent (receives compliance-checked asset)
6. Dispatch: critic-agent (receives final asset + variant)
7. If FAIL → re-dispatch named agent(s) with feedback (max 2 cycles)
8. Deliver artifact

Route C: Called by Another Skill

When: Invoked by imc-plan or attribution for inline content work.

1. Read context from calling skill's artifacts
2. Dispatch relevant agents based on caller needs
3. Dispatch: critic-agent
4. Return content to calling skill

Route D: Channel-Specific Asset

When: User specifies a channel from the 9-channel map + an angle/message. Auto-selects the right format-agent configuration.

1. Pre-dispatch: Identify channel → map to format type
2. LAYER 1 — format-agent (with channel-specific specs) + voc-extraction-agent
3. LAYER 1.5 — hook-agent + body-agent + cta-agent (adapted to channel constraints)
4. LAYER 2 — platform-compliance-agent (with channel-specific rules) → critic-agent
5. Deliver artifact

Channel → Format Mapping:

Routing clarification:

• Social media content (LinkedIn, X, TikTok, Instagram, Facebook) → Route A (single post) or Route B (carousel, campaign)
• Search engine content (blog posts, SEO articles) → Route B
• All other channels (store listings, forums, SMS, OOH, PR, UGC, bounty, newsletter) → Route D
• If unsure, check references/platform-specs.md — if the format is listed under Store/Listing, Forums, SMS, OOH/IRL, PR, UGC, or Bounty sections, use Route D.

Step 0: Pre-Dispatch Context Gathering

Product Context Check

Check for .agents/product-context.md and .agents/mkt/imc-plan.md. If date fields are older than 30 days, warn the user and recommend re-running upstream skills.

Required Artifacts

Optional Artifacts

Context to Pass to All Agents

1. Angle: from IMC plan or user brief
2. Channel + placement: target platform and format
3. Awareness stage: determines hook approach and CTA commitment
4. VoC quotes: from voc-extraction-agent (available after Layer 1)

Dispatch Protocol

How to spawn a sub-agent

1. Read the agent instruction file — include its FULL content in the Agent prompt
2. Append the context (angle, channel, awareness stage, VoC) after the instructions
3. Resolve file paths to absolute: replace relative paths with absolute paths rooted at this skill's directory
4. Pass upstream artifacts by content: the orchestrator reads .agents/ files FIRST, then includes relevant excerpts in context. Sub-agents should NOT read artifact files directly.
5. If feedback exists (from critic FAIL), append with header "## Critic Feedback — Address Every Point"

Single-agent fallback

If multi-agent dispatch is unavailable, execute each agent's instructions sequentially in-context:

• Layer 1: resolve format specs, extract VoC
• Layer 1.5: write hook, body, CTA within format constraints
• Layer 2: check compliance, create variant, evaluate with critic rubric

Layer 1: Parallel Foundation

Spawn IN PARALLEL:

Wait for both to complete. Their outputs become inputs for Layer 1.5.

Layer 1.5: Parallel Content Writers

After format-agent and voc-extraction-agent return, spawn IN PARALLEL:

Merge Step

Assemble Layer 1.5 outputs into the artifact template:

Assembly rule: Hook first, then body content, then CTA at the end. CTA placement follows the format-agent's guidance (last slide, last tweet, below fold, etc.).

Layer 2: Sequential Refiners

Dispatch ONE AT A TIME, IN ORDER:

Critic Gate

• PASS: Deliver the annotated artifact.
• FAIL: Re-dispatch named agent(s) with critic feedback. Max 2 rewrite cycles. After 2 failures, deliver with critic annotations and flag to user.

Artifact Template

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

# Content: [Title]

**Date:** [today]
**Skill:** content-create
**Angle:** [from IMC plan]
**Channel:** [platform + placement]
**Stage:** [awareness level]

## Hook

[Exact hook text]

**3Q Test:** Visual: [Y/N] | Falsifiable: [Y/N] | Uniquely ours: [Y/N]
**Annotations:** [Rule, cut alternative, rubric score]

## Body

[Complete copy — every word, slide, or timestamp]

## CTA

- **Text:** [exact CTA copy]
- **Destination:** [URL or action]
- **Tracking:** [UTM parameters]

## A/B Variant

| Element | A (Main) | B (Variant) |
|---------|----------|-------------|
| [Hook/CTA/Format] | [Original] | [One change] |

**Test hypothesis:** If B wins, it means [specific learning].

## Asset Checklist

- [ ] [Asset with specs]
- [ ] [Brand assets needed]

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

Worked Example — LinkedIn Carousel (Route B)

Brief: LinkedIn carousel about "12 hrs/week lost to status theater" Audience: Engineering managers, Problem Aware Channel: LinkedIn carousel

Step 0: Pre-Dispatch

Angle from imc-plan: "Engineering teams lose 12 hrs/week to status updates nobody reads" Awareness: Problem Aware. Channel: LinkedIn carousel.

Layer 1: Parallel Foundation

→ Format agent returns: LinkedIn carousel, 1080x1350px, 5-10 slides, ≤50 words/slide, PDF format → VoC extraction agent returns: 5 pain quotes from ICP research, buyer language "status theater," "meeting tax"

Layer 1.5: Parallel Content Writers (receive format specs + VoC)

→ Hook agent returns: "Your team loses 12 hours a week to status updates nobody reads." (V:Y F:Y U:Y, Data hook) → Body agent returns: 6 slides (pie chart breakdown, 4 pain slides, fix slide) → CTA agent returns: "See how teams ship without status theater →" (link in comments)

Merge

Assembled into 8-slide carousel: Hook slide → 6 body slides → CTA slide

Layer 2: Sequential

→ Platform compliance → PASS (all slides within limits, correct dimensions) → A/B variant → Hook variant B: "I stopped doing standups. Here's what happened." (tests story vs. data) → Critic → PASS. All dimensions ≥3.5. Delivered.

Anti-Patterns

Burying the lead — Starting with context instead of the hook. INSTEAD: The first thing the reader sees IS the hook. No preamble.

Platform-generic content — Same text for LinkedIn and TikTok. INSTEAD: Each platform gets its own format spec from format-agent. Native patterns outperform imported content.

Weak CTAs — "Learn More" or "Click Here." INSTEAD: [action verb] + [what they get]. "See how teams ship without standups →"

A/B testing multiple variables — Changing hook + CTA + format = untestable. INSTEAD: ONE element per variant with a stated hypothesis.

Writing outlines — "Slide 3: Talk about the pain point." INSTEAD: Write every word. "4 hrs — writing updates for managers who skim them."

Agent Files

Sub-Agent Instructions (agents/)

• agents/format-agent.md — Platform specs, dimensions, native patterns
• agents/voc-extraction-agent.md — Buyer language from ICP research
• agents/hook-agent.md — Opening line adapted to format
• agents/body-agent.md — Complete body content
• agents/cta-agent.md — Platform-native CTA
• agents/platform-compliance-agent.md — Technical spec verification
• agents/ab-variant-agent.md — One testable alternative
• agents/critic-agent.md — Quality scoring, PASS/FAIL

Shared References (references/)

• references/platform-specs.md — Platform dimensions, character limits, native patterns
• references/examples.md — Complete worked examples with full copy
• references/repurposing-cascade.md — Hero → derivative content workflow