Cold Email Copy

Cold Email Copy

Write cold email copy the ColdIQ way. Carry the rules held across campaigns and the patterns learned from the ones that worked and the ones that did not. Use this skill to write a sequence from scratch or rewrite a draft that is not landing.

Treat approved facts as a closed claim set. Do not turn a capability into an unsupplied benefit, reduction, improvement, or causal outcome. A possibility label does not make an unsupported product outcome safe. Treat an offered artifact as closed context too. When the brief supplies only its name, offer it by that exact name. Do not invent its contents, coverage, usefulness, or effects.

Strategy first

Copy fails on strategy more often than on craft. Before writing a line, produce a strategy brief and put it in front of the operator. Run this on every job, including one with a full campaign spec, because part of its job is to test that spec.

STRATEGY BRIEF / <client> / <campaign>

ICP           Who receives this. Seniority, function, budget authority, what
              their week contains, and how often they get pitched.
TONE          The register, and one sentence that explains the choice.
EDUCATION     Expert / Aware-but-unpriced / Uninformed, plus the evidence.
ANGLE         One sentence. The single reason this person should care now.
REPLY DATA    What has earned replies on this account, with approved evidence.
DEPTH         Per-person openers or segment-level, and the segmenting fact.
SENDER        Which sending identity, and why this reader believes them.
SIGN-OFF      Claims and proof points that need written client approval first.
PUSHBACK      Where research disagrees with the operator, the evidence, and
              what would change in the copy.

Hand the brief to the operator and keep working in the same turn. It is a disclosure, not an approval gate. Stop and ask only when a PUSHBACK item would change the ANGLE itself. Rewrite the angle as though the pushback were correct. If it still names the same person and the same problem, proceed and mark the alternative reversible. If it names a different person or problem, stop. That would be a different campaign.

Four things decide the brief:

1. Load the client context and research the market. Read only the materials placed in scope. Use research tools when they are available and allowed. If required research is unavailable, state the gap instead of inventing a plausible answer.
2. Decide the register. Use seniority, profession, pitch frequency, budget authority, and language needs. Simple language is not the same as casual language. See resources/audience-register.md.
3. Decide how much the reader already knows. An expert inside the problem needs proof that the sender understands the work, not an explanation of the problem. See resources/strategy-brief.md.
4. Know the angle and check it against approved reply evidence. Write one sentence that names a person and a problem. If it needs a product clause to make sense, it is a feature, not an angle.

Research even when the operator supplied an ICP and angle. Where evidence contradicts the spec, state the conflict, cite the evidence, and explain what would change. Write on the operator's read unless the conflict changes the angle itself.

Once the brief is settled, find the per-person anchor: the real research each opening line will use. If the answer is nothing, say so before copy is generated. See resources/personalized-openers.md.

The rules

• Write two or three emails.
• Use 75 to 90 words per email by default.
• Keep every sentence under 20 words.
• Keep every subject between three and five words.
• Use one ask per email. Do not stack CTAs.
• Lead with the pain or wedge, never the product or a generic pleasantry.
• Do not use the U+2014 em dash character.
• Never add, suggest, or assume open or click tracking.
• Use simple language for audiences that are not native English speakers or not technical. This controls vocabulary and sentence length, not the register.
• Make clarity outrank cleverness. Use no irony, callbacks, wordplay, or observations about the reader's career. Name who does what instead of using an abstract noun.
• Show that you know the reader's business. Do not set out to educate an expert about their work.
• Use spintax only when the user, client rules, or selected sending tool requires it. Never nest a merge field inside a spintax block. Keep every approved claim sentence byte-stable.
• Remove machine-writing tells before review. Check lists of three, mirrored contrasts, self-answered questions, fragments as punchlines, filler transitions, speculation, and banned vocabulary. See resources/ai-tropes.md.

Longer proof-heavy copy can run to 150 words only when the user requests it and every proof point has a source. Nothing else buys extra words. Use the full reasons and exceptions in resources/copy-rules.md.

The per-person opening line

Default to one generated opener per recipient followed by a fixed, client-approved body when the available evidence and client review capacity support it. The client approves the body once. The opener is the only moving part, and it never carries a product claim.

Use two sentences and no more than 35 words. State what the person specifically does, then the problem next to it. Build it from real research about the person, not from a job title alone. With an expert audience, the second sentence often works better as a diagnostic question about their own operation.

Judge the anchor before using it. Not every fact earns a personal line, and forcing one is worse than not having one. Return usable, one short reason, and the opener. When usable is false, use an approved segment-level opener and mark the fallback.

Give the model the domain facts for each segment. A model asked to be specific without the facts will be specific and wrong. Match personalization depth to the client's review capacity. When the client cannot approve hundreds of unique lines, use one approvable opener per meaningful segment.

See resources/personalized-openers.md.

The sequence structure

• Email 1, core context. Give the strategy and the reason to reach out now. Open on the prospect's pain or a shift in their world, then the wedge.
• Email 2, the platform. Explain what the product does beyond the first hook and how it changes the reader's work. Keep it additive.
• Email 3, the close or route. Make it a truthful last note, offer a useful next step, or ask for the correct owner. Do not create false urgency.

Use resources/sequence-structures.md for the two-step variant and worked skeletons. Use resources/campaign-patterns.md for an approved campaign approach and resources/exemplars.md for fictional composites.

When the recipient is not the decision maker, write to the problem they own. Make the one CTA a routing ask. When a fact needed for segmentation is unavailable, ask the reader to identify it instead of making a guess look like personalization.

Choose subjects with resources/subject-lines.md.

Answering replies

Treat replies as a different genre. Answer the exact question in the first line, add one paragraph of substance, and offer one next action that matches the prospect's intent. Reply copy has no sequence word limit and never uses spintax. Never say that material is attached, linked, sent, or scheduled unless the artifact or approved delivery fact is present in the working context. When it is missing, return insufficient_context and name the blocker outside the prospect-facing draft. See resources/reply-copy.md.

Draft and review workflow

Keep one canonical draft throughout the workflow.

1. Create the strategy brief and canonical draft from approved context.
2. Run the mechanical checks in resources/ai-tropes.md. For generated per-row copy, regenerate with the failure reasons. Do not hide a failing prompt with hand edits.
3. Run Prospect clarity with resources/prospect-reviewer.md.
4. Run Client voice and requirements with resources/client-voice-reviewer.md.
5. Run Research and factual support with resources/research-reviewer.md.
6. Run ColdIQ copy standards with resources/coldiq-copy-reviewer.md.
7. Revise in this order: prospect, client, research, ColdIQ standards.
8. Run QC last with resources/qc-reviewer.md.
9. If QC fails, make one repair and run one final QC check.
10. If the second QC check fails, return needs_human_review with the blockers.

When the host can start subagents, run the four specialist reviewers as independent tasks. Start QC only after their findings are coordinated and the canonical draft is revised. Include real task IDs when the host provides them. Retry a failed native reviewer once. If it fails again, return incomplete_review and name the missing review. Report Review mode: native-subagent-review.

When the host cannot start subagents, run the same four rubrics as compact labeled passes. Then revise and run QC. Never claim that separate agents ran. Report Review mode: single-agent-review.

Limit each reviewer to five findings. Make each finding one short sentence with a location, issue, and proposed change. Quote no more than twelve draft words. Do not repeat the draft in reviewer output. Show the final copy once.

Deliverable

Return:

1. The strategy brief.
2. The real review mode.
3. Status: complete, insufficient_context, incomplete_review, or needs_human_review.
4. A compact review summary for the four reviewers and QC.
5. The final subjects and copy once.
6. A short fact ledger when the copy contains factual claims.
7. Personalization coverage and fallbacks when openers vary by person or segment.

Return Markdown by default. Create a DOCX only when the host supports files and the user asks for it. For a client-ready document, open with campaign metadata, show the approved copy and variants, list only questions the client must answer, and remove internal process commentary. Preview every opener type with its source evidence. Never ask the client to resolve an internal mistake.

Do not expose private campaign evidence. Public output can contain process rules, normative limits, and newly written fictional examples. It cannot contain measured campaign totals, response rates, conversion lifts, client identities, client results, launch findings, raw client copy, or facts derived from one client.