Copy brief

Copy Brief

Context-gathering skill. No copy generation. Receives inputs, validates fields, returns a complete structured brief.

Overview

This skill handles the gap between what a consuming agent or human knows and what a copy skill needs. It validates field completeness, identifies gaps, and resolves them through three distinct modes. Output is always a complete structured brief in the standard YAML format.

Input Contract

The brief uses this YAML structure:

type: "cold outreach email"
goal: "Book a demo call"
audience:
  who: "US food enthusiasts, 28-45"
  pain: "Bored of supermarket sake"
  sophistication: "casual"
product:
  name: "SakeBox"
  description: "Monthly sake subscription"
  differentiator: "Direct from 40+ Japanese breweries"
brand_voice:
  tone: "knowledgeable, casual"
  avoids: ["artisanal", "curated", "elevate"]
  examples: ["Example copy that matches desired voice"]
constraints:
  length: "150 words max"
  format: "plain text email"
  cta: "single CTA, link to booking page"
  language: "en-US"

Field Classification

Field	Status	If missing
type	Required	Cannot proceed
goal	Required	Cannot proceed
audience.who	Required	Cannot proceed
product.name	Recommended	Proceed, flag reduced quality
audience.pain	Recommended	Infer from type + audience
brand_voice	Recommended	Use domain defaults
constraints	Optional	Use domain-standard defaults

Required fields (cannot proceed without all three): type, goal, audience.who

Recommended fields (proceed with a quality flag if missing): product.name, audience.pain, brand_voice

Optional fields (sensible defaults apply): constraints

Mode Selection Logic

Evaluate in this order:

1. If all required fields are present in the input, proceed to Mode 3 (fast path).
2. If the calling context is non-interactive (invoked by an agent or pipeline), use Mode 2 (agent handshake).
3. Otherwise, use Mode 1 (interview).

Mode 1: Interview (Human User)

Use when a human is providing context conversationally.

Rules:

• Ask one question at a time. Never present a list of questions.
• Start with required fields. Confirm each before moving on.
• After required fields are covered, move to recommended fields.
• Load references/brief-questions.md for the question bank. Select the question that fits the copy type.
• Build the brief progressively as answers arrive.
• When minimum requirements are met, confirm the brief and offer to proceed.

Sequence:

1. Ask for type if not provided: use the universal question for copy type.
2. Ask for goal if not provided.
3. Ask for audience.who if not provided.
4. Ask for recommended fields one at a time, selecting from the relevant domain section in brief-questions.md.
5. When all recommended fields are covered (or the user declines), confirm and output the brief.

Confirmation step: Before proceeding to copy generation, surface the brief summary:

Here's what I have:
- Writing: [type]
- Goal: [goal]
- Audience: [audience.who]
- [any other fields collected]

Ready to proceed, or anything to adjust?

Mode 2: Agent Handshake (Multi-Agent)

Use when invoked by an orchestrator, pipeline, or another agent.

Rules:

• No conversational interaction. Structured request and response only.
• Receive the partial or complete brief from the calling agent.
• Check required fields first. If any required field is missing, respond with the structured missing-fields format and stop.
• If required fields are present, check recommended fields. Identify gaps and request them.
• Proceed once minimum requirements are met.

Response format when fields are missing:

missing:
  required:
    - audience.who: "Who is this for? (age range, role, context)"
  recommended:
    - brand_voice: "Any tone or voice guidelines?"
    - audience.pain: "What problem does this audience have?"

Only list fields that are actually missing. If no required fields are missing, omit the required key. If no recommended fields are missing, omit the recommended key.

Handshake flow:

Step 1: Orchestrator sends partial brief
  { type: "drip email step 3", goal: "re-engage inactive users" }

Step 2: Copy-brief identifies gaps, responds with missing-fields YAML

Step 3: Orchestrator fills what it can, resends the updated brief

Step 4: Copy-brief confirms minimum requirements met, outputs the complete brief

When to proceed with gaps: If required fields are all present and the calling agent cannot provide recommended fields, proceed. Include a flags entry in the output brief noting which recommended fields were absent.

Mode 3: Complete Brief (Fast Path)

Use when all required fields are already present in the input.

Rules:

• Validate that type, goal, and audience.who are present and non-empty.
• Confirm minimum requirements are met.
• Proceed immediately without asking questions.
• If recommended fields are missing, add a flags entry to the output noting reduced context.

Validation:

• type must be a non-empty string describing the copy format.
• goal must describe a specific outcome, not a vague intent.
• audience.who must identify the audience concretely.

If any required field is empty, missing, or too vague to act on, fall back to the appropriate mode (Mode 1 for humans, Mode 2 for agents).

Output

The output of this skill is always a complete structured brief in the standard YAML format. Include all fields that were collected or inferred. Add a flags array for any recommended fields that were not provided.

type: "cold outreach email"
goal: "Book a demo call"
audience:
  who: "US food enthusiasts, 28-45"
  pain: "Bored of supermarket sake"
  sophistication: "casual"
product:
  name: "SakeBox"
  description: "Monthly sake subscription"
  differentiator: "Direct from 40+ Japanese breweries"
brand_voice:
  tone: "knowledgeable, casual"
  avoids: ["artisanal", "curated", "elevate"]
  examples: ["Example copy that matches desired voice"]
constraints:
  length: "150 words max"
  format: "plain text email"
  cta: "single CTA, link to booking page"
  language: "en-US"
flags: []

If recommended fields were absent, the flags array lists them:

flags:
  - "brand_voice not provided: using domain defaults"
  - "audience.pain not provided: inferred from type and audience"

Hard Rules

No copy generation. This skill gathers and validates context. It does not write copy.

Required fields block all modes. If type, goal, or audience.who cannot be determined, stop and ask. There is no fallback.

One question at a time (Mode 1). Never present a list of questions to a human user.

Structured responses only (Mode 2). No prose explanations to agents. Return YAML.

Infer, do not hallucinate. If a field can be reasonably inferred from other provided fields, make the inference and note it in flags. Do not invent audience pain points or brand voice that contradict provided context.