How to use Draftship with Anthropic Claude

Claude is good at email copy when you ask it correctly. Bad prompts give you LinkedIn-flavored slop. Good prompts return drafts that need a light edit pass and ship. The trick is telling Claude what shape you want the output in (block by block, in JSON) and what voice it should match (with two or three real examples from your archive).

Three ways to wire Claude into a Draftship workflow

• Console (claude.ai): paste a long prompt, get drafts, iterate. No code.
• API: programmatic generation. Useful when you're populating templates from CRM data or running batch drafts.
• Workbench (console.anthropic.com): prompt engineering with side-by-side runs and version comparison.

Prompt template: full email draft

The pattern that consistently produces paste-ready output:

txt
You are writing a marketing email for {{ company }}, a {{ one-line description }}.

Brand voice anchors:
- Direct. No marketing fluff.
- Short sentences over long.
- Avoid these words: synergy, leverage (as a verb), unlock, journey.
- Tone examples (real emails from our archive):
  [paste 2-3 short past emails here]

Email goal: {{ describe what the recipient should do }}
Audience: {{ who is reading this }}
Constraints:
- Subject line under 50 characters.
- Preheader under 90 characters.
- One H1 heading.
- 2-4 body paragraphs.
- One primary CTA button.

Return JSON in this shape:
{
  "subject": "...",
  "preheader": "...",
  "blocks": [
    { "type": "heading", "text": "..." },
    { "type": "text", "text": "..." },
    { "type": "button", "label": "...", "href": "..." }
  ]
}

The JSON output paste-imports into Draftship's block library cleanly. Each block becomes one Draftship block. The structure forces Claude to think in blocks, not in continuous prose.

Prompt for subject lines only

txt
Generate 10 subject lines for this email. Constraints:
- Under 50 characters
- No exclamation points
- No emoji
- Match the tone in this past sample: "[paste a real subject from your archive]"

Email body for context:
[paste your draft body]

Run the output through the subject line checker to filter for spam-trigger words and length.

Prompt for variant generation (A/B)

txt
Take this email and produce 3 variants:
- Variant A: matches the original tone exactly.
- Variant B: more direct, removes one paragraph.
- Variant C: more conversational, adds a personal note.

Original:
[paste your draft]

Return as JSON with keys A, B, C, each containing { subject, preheader, body }.

Using the API for batch generation

For populating CRM-bound campaigns, the API call:

bash
curl -X POST 'https://api.anthropic.com/v1/messages' \
  -H 'x-api-key: sk-ant-...' \
  -H 'anthropic-version: 2023-06-01' \
  -H 'content-type: application/json' \
  -d '{
    "model": "claude-opus-4-7",
    "max_tokens": 2000,
    "messages": [
      { "role": "user", "content": "Your full prompt here" }
    ]
  }'

Use the Sonnet or Haiku tiers for higher-volume drafting; Opus for the highest-stakes one-off campaigns where voice precision matters most.

Pasting Claude's output into Draftship

Three patterns for moving from Claude's response into Draftship:

1. Block by block: copy each section, create the corresponding Draftship block, paste content. Slowest but gives you the most editorial control. 2. JSON paste: if you've prompted for JSON, write a small script that parses the JSON and creates Draftship blocks via the import format. Fastest for batch. 3. HTML fragment: prompt Claude for inline HTML fragments per block. Paste each into Draftship's HTML block. Loses some block-level features but is fast.

For one-off launches, pattern 1. For campaigns running at volume, pattern 2.

Test send checklist

• Run the final email through the size checker.
• Send a test through your ESP, not just Draftship's preview.
• Check the rendered output in Gmail and Outlook.
• Verify any AI-generated link URLs are real, not hallucinated.

When Claude isn't the right tool

For pure visual edits (image cropping, color swaps), Claude is overkill. Use Draftship's editor directly. For structured personalization (cart abandon items, dynamic offers), use your ESP's templating engine, not AI generation.

For prompt engineering broadly, the brand voice principle generalizes: anchor with examples, constrain output shape, validate output with explicit tests.