Orchestrates the full copy generation pipeline. Contains NO domain knowledge. Only knows the sequence of steps and how to route between skills.
Overview
This skill is the central orchestrator for copy generation. It chains together the briefing, domain routing, drafting, quality enforcement, and revision steps in a fixed sequence. It does not write copy itself, does not know copy frameworks, and does not make domain-specific decisions. Those are delegated to the appropriate skills at each step.
Every copy generation request starts here and ends here.
The 8-Step Workflow
Step 1: Brief
Invoke the copy-brief skill.
Pass any input the user or calling agent provided. copy-brief handles validation and gap-filling through its three modes (interview, agent handshake, fast path).
This step is complete when copy-brief returns a complete structured brief in YAML format with type, goal, and audience.who all present.
Do not proceed to Step 2 until the brief is complete.
Step 2: Domain Routing
Match the brief's type field to a domain skill using keyword matching.
Ad rule (checked before the table): if the type contains "ad", "ads", or "advertisement" as a standalone word, route to marketing-copy immediately. Paid ads are marketing copy even when they name a platform: a LinkedIn ad routes to marketing-copy; a LinkedIn post routes to social-copy.
Routing table (checked in row order, first match wins):
Keywords in type field |
Route to |
|---|---|
| "email", "subject line", "newsletter", "drip", "sequence", "outreach", "cold email", "cold outreach", "nurture", "campaign", "re-engagement", "transactional" | email-copy |
| "microcopy", "button", "button label", "error message", "error state", "onboarding", "empty state", "tooltip", "UX", "UI copy", "notification", "dialog", "confirmation", "placeholder", "helper text" | ux-copy |
| "LinkedIn", "Twitter", "X post", "tweet", "Instagram", "TikTok", "social", "social post", "thread", "carousel", "caption" | social-copy |
| "blog", "blog post", "article", "SEO", "seo content", "thought leadership", "whitepaper", "opinion piece", "long-form content" | editorial-copy |
| "proposal", "case study", "pitch deck", "one-pager", "battle card" | sales-copy |
| "brand voice", "voice profile", "tone guide", "messaging hierarchy", "style guide", "tagline", "elevator pitch", "brand guidelines" | brand-copy |
| "pricing", "pricing page", "signup", "sign-up", "checkout", "A/B", "variant", "trial", "conversion", "funnel" | conversion-copy |
| "landing page", "lander", "ad", "advertisement", "CTA", "value prop", "value proposition", "hero", "banner", "homepage", "sales page" | marketing-copy |
Match is case-insensitive. Row order encodes the tie-breaks: a named social platform beats topic keywords (a LinkedIn thought leadership post is social copy, not editorial), and marketing is checked last because its keywords are the most generic. This table mirrors routeToDomain() in @copydoc/core; keep the two in sync when editing either.
If no match: Ask the user which domain skill to use, or default to marketing-copy if in agent mode.
Note that "sales email" matches the email row and routes to email-copy, which handles it with PAS and the cold-outreach reference (the same treatment sales-copy would give it).
Step 3: Framework Selection
Defer entirely to the domain skill.
Pass the complete brief to the routed domain skill. The domain skill selects the appropriate framework (or principles, in the case of ux-copy) based on its own internal mapping. Do not override or second-guess the domain skill's framework choice.
Step 4: Draft
The domain skill generates the copy using its selected framework and domain patterns.
Receive the draft output from the domain skill. Note the framework used and any domain-specific flags. Pass draft to Step 5.
Step 5: De-Slop Pass
MANDATORY. This step cannot be skipped under any circumstances.
Load ../quality-frameworks/references/anti-slop.md.
Apply the full de-slop pass to the draft:
- Scan for every banned word and phrase
- Flag em dash usage
- Flag tricolon patterns that read like jingles
- Flag AI tell-tale phrasing (hedging language, hollow transitions, vague superlatives)
- Apply humanization techniques from the anti-slop reference
- Rewrite every flagged section before proceeding
Do not move to Step 6 with any unfixed anti-slop violations. The de-slop pass is non-negotiable.
Step 6: Critique
Score the de-slopped draft against the 7-dimension quality rubric in quality-frameworks SKILL.md.
| Dimension | Scale |
|---|---|
| Clarity | 1-10 |
| Specificity | 1-10 |
| Voice Match | 1-10 |
| AI-Tell Score | 0-10 (lower is better) |
| Persuasion | 1-10 |
| Action | 1-10 |
| Overall | 1-10 |
Record all 7 scores. Pass scores and draft to Step 7.
Step 7: Revise
Revision threshold: If AI-Tell Score >= 3 OR Overall score < 7, revision is required.
If revision is required:
- Return the draft to the domain skill with the critique scores and specific flags.
- Domain skill revises based on the flagged dimensions.
- Re-run Steps 5 and 6 on the revised draft.
- Evaluate against the threshold again.
Hard cap: maximum 2 revision cycles. If the draft still fails the threshold after 2 cycles, proceed to Step 8 with the best available draft. Set flags in the output to note the unresolved quality issues and which threshold was not met.
If no revision is required (AI-Tell < 3 AND Overall >= 7): Proceed directly to Step 8.
Step 8: Return
Return structured output matching the output contract below.
Include the framework used, domain routed to, all 7 quality scores, and any flags from the revision process.
Human Checkpoints
Checkpoints are on by default for human users (a user can ask to skip them). They are skipped entirely in agent mode.
Checkpoint A (after Step 1):
"Here's what I understand: [brief summary of type, goal, audience]. Correct?"
Wait for confirmation before proceeding to Step 2.
Checkpoint B (after Step 3):
"I'm going with [framework name] for this. Sound right?"
Wait for confirmation before proceeding to Step 4.
Checkpoint C (after Step 4):
"Here's the draft. Want me to revise it, or should I run it through the quality check?"
Wait for direction before proceeding to Step 5. If the user asks for manual revisions, apply them and re-present. Once the user confirms, proceed.
Agent Mode
When invoked by an orchestrator, pipeline, or another agent:
- All three checkpoints are skipped.
- Execution is end-to-end without interruption.
- Domain routing defaults to
marketing-copyif no keyword match is found. - Revision cycles run automatically up to the 2-cycle cap.
- Return the structured output contract when complete.
Output Contract
copy:
primary: "..."
variants: ["...", "..."]
metadata:
framework_used: "PAS"
domain: "email-copy/cold-outreach"
quality_scores:
clarity: 8
specificity: 7
voice_match: 9
ai_tell_score: 2
overall: 8
flags: []
Fields:
copy.primary: The final copy, ready to use.copy.variants: Optional alternate versions. Empty array if none were produced.metadata.framework_used: The framework or approach the domain skill used (e.g., "PAS", "AIDA", "4Us").metadata.domain: The domain skill routed to, with sub-type if applicable.metadata.quality_scores: All 7 scores from the final critique in Step 6.metadata.flags: Empty array on clean pass. Populated if revision cap was hit or recommended brief fields were absent.
Example flags:
flags:
- "AI-tell threshold not met after 2 revision cycles: ai_tell_score 3"
- "brand_voice not provided: voice_match scored at 5"
Hard Rules
No domain knowledge in this skill. Framework selection, copy patterns, audience targeting, and format specifics all belong to the domain skill. This skill sequences steps and routes -- nothing else.
Brief must be complete before routing. Never skip copy-brief. Never guess at missing fields.
De-slop is mandatory. Step 5 runs on every draft, every revision cycle, without exception.
Revision cap is 2 cycles. Never exceed two revision cycles regardless of quality scores. Return with flags.
One domain skill per request. Route to exactly one domain skill. Do not blend outputs from multiple domain skills.