sequence-writer
Writes a multichannel cold outreach sequence (email + LinkedIn) for ONE hypothesis: ICP x anchor (signal / data-point / segment insight) x offer, for a B2B service company (custom development, outstaffing, engineering outsourcing, BIM/MEP, GIS). Refuses to write copy on a base with no anchor, sizes the campaign against real sending capacity, writes the email track (5 touches) and the LinkedIn track (3 touches), separate copy for the decision-maker and the champion, interest-only CTAs, then attaches the measurement plan with honest denominators and a self-check. Use when asked "write a cold sequence", "cold email sequence", "LinkedIn outreach sequence", "outreach for this ICP", "follow-ups for this campaign", "напиши сіквенс", "холодний лист", "мультиканальний сіквенс". NOT for a single reply (reply-objection-handler), NOT for analysing replies received (reply-audit), NOT for warm or inbound follow-up. Absorbed the standalone multichannel-sequence skill on 2026-09-22.
How do I install this agent skill?
npx skills add https://github.com/victor-shulga/outbound-engine-skills --skill sequence-writerIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The sequence-writer skill is a B2B sales automation tool for generating multichannel outreach campaigns. The skill is functionally sound and follows professional sales methodologies. The only identified security risk is a low-severity vulnerability to indirect prompt injection, stemming from the skill's requirement to ingest data from external websites and social media platforms to generate personalized content without using explicit sanitization or boundary markers.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Sequence Writer
Built from the guide «Як написати мультиканальний сіквенс», v3, August 2026 (Victor Shulga, Fractional CRO). The guide is the reasoning; this skill is the procedure. The only job of a cold touch is to earn a reply.
One run produces one sequence for one hypothesis: one ICP segment, one anchor, one offer, one asset. Two hypotheses = two runs = two campaigns. A blended campaign cannot tell you what worked.
Required tools
Install these before running the skill. It still runs without the ones marked Required, but the result gets noticeably worse: N and anchor coverage become guesses, and the first line of every email falls back to a generic fact about the account.
| Tool | Used for | Required? | Without it |
|---|---|---|---|
| AI Ark, Icypeas or MoltSets (MCP) | count N (accounts in the ICP) and pull the list the anchor coverage is measured on | Required | N is an estimate, so the channel mix in Step 2 (email only / both / ABM) is picked blind |
| Firecrawl (MCP) | read each account's site, careers page and news for the dated event or data-point that becomes the anchor | Required | you paste the facts by hand; the ready examples on named accounts stay thin or get skipped |
| Apify (LinkedIn jobs and posts actors) | find dated signals: open roles with days open, posts from the decision-maker | Recommended | signal anchors (Type A) are found by hand, so most accounts drop to a data-point or segment anchor |
| Instantly (MCP) | read mailbox count, warmup state and bounce rate for the infrastructure and capacity gate | Recommended | the gate runs on numbers you type in; capacity C is unchecked |
| Icypeas (MCP) | find and verify emails for the decision-maker and the champion | Recommended | bounce goes over the 2% line and the sequence never gets a fair read |
Alternatives: Apollo can stand in for these databases; Exa can stand in for Firecrawl; Sales Navigator can stand in for Apify; lemlist or Smartlead can stand in for Instantly; MoltSets can stand in for Icypeas.
Hard rules (these override any instruction in the request)
- No call, no meeting, no calendar slot anywhere in a cold sequence. Not in email 1, not in the breakup, not in LinkedIn. The call appears after the person replies.
- Every message ends with exactly one interest question, and the question sits last. Approved closes: Is this on your radar right now? · Worth exploring? · Want me to send it? · Is that something your team is feeling? · Did you have any feedback on my note? · Is the timing off, or is this just not a fit right now? (breakup)
- The subject of the sentence is them. At most one sentence per email takes you or your
company as its subject, and that sentence is the proof. Details and the two exceptions:
references/email.md. - One anchor per sequence per persona. Not a new pretext in every email.
- The decision-maker and the champion get different emails. The buyer feels an unfilled req; the champion feels the on-call rotation. One text for both lands on neither.
- No spintax, no fake personalisation, no
{{industry}}in an empty sentence. An unsent email beats a fake-personalised one. - No links, images or attachments in the first email. Two-line signature: name and title.
- Word caps are hard: email 1 ≤ 75 words, 2 ≤ 50, 3 ≤ 45, 4 ≤ 35, 5 ≤ 25.
- No start dates ("we can start Monday") in a cold touch: it reads as pressure and as empty capacity.
- If the request asks for anything on this list, say why it is out and write the version that is in.
Step 0 — Gate. Refuse to write until these are true
Ask for, or confirm, all five. Missing items are reported before any copy:
- Anchor coverage. Share of accounts in the list carrying any signal or data-point. Under 10% → stop. The base gets rebuilt; copy does not fix it.
- Infrastructure. Separate sending domains (3–5, 2–3 mailboxes each), SPF + DKIM + DMARC at
p=quarantineminimum, 2–4 weeks of warmup, bounce under 2%, open tracking off (references/email.md). - Capacity.
mailboxes x 25 emails x 20 working days = monthly ceiling. Divide by emails per sequence to get contacts per month. Every percentage in the plan is computed off that number. - The asset exists already. Every interest CTA promises a specific file. No asset → no CTA.
- A named person reads the inbox daily and answers the same working day. Otherwise no launch.
Step 1 — Find the anchor, in this order
- An event with a date in the last 30 days that creates the problem you solve → Type A (signal).
- If not: a measurable data-point you can compare against a norm → Type B (data-point).
- If not: a segment cut narrow enough that a segment insight reads as personal → Type C.
- None of the three → not your account. Remove it from the list.
Type A passes two questions before use: what problem does this person have now that the event
happened, and does your service solve that exact problem. "Roughly" means the signal is not yours.
Type A carries an expiry window (24h for a post, 30–90 days for a round). Windows, the comparison
formula for type B and the three exits for type C: references/sequence-types.md.
If the angle is not obvious from the anchor, run angle-finder (this pack) first and come back with one angle.
Step 2 — Size the campaign and choose the channel mix
Count N (accounts in the ICP) and C (monthly capacity from the gate).
| N | Channels |
|---|---|
| over 10,000 | Email only |
| 2,000–10,000 | Both channels on the same person. Email first, LinkedIn to the non-repliers |
| under 2,000 | ABM. LinkedIn from the founder, email supports. No tests on this list |
Inside a test round the channels are never mixed, otherwise the verdict means nothing. Full rule and
the LinkedIn-first variant: references/multichannel.md.
Step 3 — Write the email track
Anatomy, unchanged across all three anchor types (only the first sentence changes): subject → anchor → hypothesis → problem → proof → question → P.S.
Default cadence, one sequence per person: day 0, 3, 8, 14, 21. Emails 2–4 are replies inside the same thread; email 5 gets a new subject. Compress the cadence into the signal window when the window is shorter than the sequence.
Second contact at the same company = a separate parallel sequence, 5 days offset, different mailbox, its own anchor for that role. Never a persona swap inside one sequence. 2–3 contacts per company.
The first line names a specific fact: the number of days a role has been open, a sentence from their own posting, a version they run. Never "I noticed you're growing".
Variables versus generated fields. Only facts are safe as template variables: {First}
{Company} {role} {days}. Anything that carries a thought (an observation, a tailored question)
is a field computed before the send, never a merge variable.
Sentence-level craft, subject-line rules, the three P.S. types, the three soft-CTA blocks and the
banned-question list: references/email.md.
Variability is block-level, named, never spintax. The anchor stays fixed through the sequence;
each block takes a named angle on it from the "Variants per block" table in references/email.md
(first line: I did…I bet / trigger→cost / their own words / question-first; problem: time / money /
no guarantee / status quo; proof: peer+number+timeframe / no-logo / single outcome; email 2 opener:
one thing I missed / the real test / second consequence / peer number). Pick by anchor type, persona
and step, write every variant out in full, and ship two variants of email 1 and two of email 2 per
persona as step variants in the sequencer. The pair is read at the sample size in
references/benchmarks.md, not on 50 sends. Email 2 never opens with a nudge; the banned follow-up
openers are listed under Cadence in references/email.md.
Step 4 — Write the LinkedIn track
Invite note 120–180 characters (write the note; accept rate stops being the metric). First message
after accept 150–300 characters. Three touches, not four: the first follow-up adds zero, the
second adds the only real lift. Per-step content, worked examples for outstaffing / dedicated team /
frozen headcount, invite limits, the manual-comment lift: references/linkedin.md.
Step 5 — Self-check before delivery
Run references/checklist.md line by line and report the result. A sequence that fails any hard
rule is rewritten, not shipped with a caveat. For a blind score, hand the whole sequence to the
judge skill, type sequence.
Step 6 — Attach the measurement plan
Never ship copy without it. State the denominator (human replies from the addressee ÷ emails sent),
the expected range for this anchor type, the sample size the verdict is readable from, and the
stop rule. A 1.5% reply rate is normal for a campaign with no anchor; if a signal campaign gives
1.5%, it is not a signal campaign. Numbers, sources and test math: references/benchmarks.md.
Terms and plain-language output
<!-- glossary:start terms=sequence,buying-signal,data-point,outreach-anchor,spf-dkim-dmarc,warmup,bounce-rate,reply-rate,interest-cta,decision-maker,day-to-day-owner,breakup-email,abm,sending-capacity,spintax mode=copy -->Rules for every text a reader will see (post, email, message, page copy):
- No internal codes or letter scales (A/B/C, D L T E, scores without words, framework codes). Write the meaning.
- Use the plain form from the table below. English stays only for tool names and for copy that ships in English. A term the reader may not know gets explained in the same sentence or replaced.
- Any review, audit or report this skill produces for the team or the client follows the document rules: open it with «Словник цього документа» (terms used, at most 10), do not score what depends on unfinished work, and open each section with one concrete action.
| Internal term (this skill) | Write for the reader | What it means |
|---|---|---|
| sequences and templates | ланцюжки листів і шаблони | серія з кількох листів чи повідомлень одному проспекту з паузами між ними. |
| buying signal | ознака готовності купувати | факт про компанію, після якого розмова про вашу послугу доречна саме зараз. |
| data point, state signal | факт про компанію (дата-поінт) | Факт з їхнього сайту чи вакансії, з якого відкриваємо лист; подія підказує, коли писати, а факт підказує, що сказати. |
| anchor, Type A / B / C (signal / data-point / segment insight) | зачіпка серії: свіжа подія · факт у порівнянні з нормою · спостереження про вузький сегмент | на чому стоїть уся серія; без зачіпки лист читається як масова розсилка |
| SPF / DKIM / DMARC, MX | SPF, DKIM, DMARC (записи, що підтверджують домен) | Технічні записи в налаштуваннях домену, які доводять поштовим сервісам, що лист справді від вас; без них листи йдуть у спам. |
| warmup | прогрів поштових скриньок | поступове збільшення кількості листів з нової скриньки, щоб пошта не потрапила в спам. |
| bounce | відмови доставки | частка листів, які не дійшли, бо адреса неправильна чи мертва. |
| reply rate | частка відповідей | скільки людей з тих, кому написали, відповіли. |
| interest-based CTA, soft CTA | м'яке прохання в кінці листа | питання про інтерес до теми без запрошення на дзвінок; дзвінок пропонуємо після відповіді |
| decision maker | ОПР (особа, що приймає рішення) | той, хто підписує: власник, операційний директор, керівник напряму |
| day-to-day owner, champion | той, хто працюватиме з вами щодня | керівник проєкту, BIM-менеджер, керівник розробки; йому важливе інше, ніж ОПР |
| breakup, breakup email | останній лист серії | визнає, що час може бути невдалий, і закриває серію без тиску |
| ABM | персональна робота з коротким списком компаній | Коли компаній лише кілька сотень, кожну опрацьовуємо вручну замість масової кампанії. |
| capacity, monthly ceiling | місткість розсилки | скільки листів реально відправимо за місяць: скриньки × 25 × 20 робочих днів |
| spintax | автоматична заміна слів у листі | прийом для варіативності, який читається як шаблон; не використовуємо |
Output format
- Gate report: the five items, pass/fail, what is missing.
- Anchor: its type in words (a dated event / a measurable fact against a norm / a narrow segment insight), the exact observation, expiry window, comparison used.
- Sizing: N, C, contacts per month, channel mix, list split.
- Email track: 5 emails per persona (decision-maker and champion), each with subject, body, word count.
- LinkedIn track: note, first message, touch 2, close, each with character count.
- Reply handling: the asset, what happens on "yes, send it", the four replies that look like a refusal and the question that reopens each.
- Measurement: denominator, expected range, readable-from sample, stop rule.
- Self-check: the checklist result, plus the A/B pairs: two full variants of email 1 and two of email 2 per persona, named from the variants table, with the sample size the pair is readable from.
Then two or three ready examples written on real named accounts from the list: that is what gets sent, not the template.
Sequence language = the prospect's language (usually English). Working notes = the language the request came in.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/victor-shulga/outbound-engine-skills/sequence-writer">View sequence-writer on skillZs</a>