Event Outbound

Event Outbound Skill

Turn an event plus a target persona into a validated multi-channel outreach sequence. Use the active Claude or Codex model as the generator. No external model API key is required. Run the bundled local validator between drafts.

For every command below, replace <absolute-skill-directory> with the resolved absolute directory containing this SKILL.md. Do not type the placeholder literally and do not rely on a new environment variable.

• Claude Code or Cowork: expand ${CLAUDE_PLUGIN_ROOT}/skills/event-outbound to its absolute path.
• Codex: use the absolute path shown for event-outbound in the available-skills catalog.

Install the complete plugin. A copied SKILL.md does not include the canonical validator and data files.

Scope

Use this skill when the user wants:

• A pre-event outreach cadence (4-week lead time by default, configurable 1-8 weeks).
• A post-event follow-up sequence keyed to a booth scan, session attendance, or speaker drop-by.
• Side-event invites (dinners, founder breakfasts, after-parties) at a major conference.
• Channel-aware copy: cold email, LinkedIn connection request, LinkedIn DM, day-of nudge, post-event email.

Do not use this skill for: lifecycle/nurture email, transactional email, content marketing, generic always-on outbound. The generator is tuned for event-shaped triggers and will produce weak copy when forced into general use.

Non-Negotiable Quality Contract

Do not generate a sequence from thin inputs like event + persona + channel + sender name. That produces plausible vendor copy, not buyer-first cold outbound.

Before drafting, build a short Outbound Research Brief. If the user gives a company website or event URL, fetch it. If the user only gives a company name, use web search to find the official site before asking. If web tools are unavailable, say that and ask for the missing facts.

Validation gate

The local validator is mandatory, not advisory. If shell/node tools are unavailable, if the user asks to disable tools, or if node "<absolute-skill-directory>/scripts/run.mjs" validate-touch --stdin cannot run after substituting the resolved path, do not draft copy. Return a blocked status with the missing capability and the next input needed. Do not include sample subject lines, sample bodies, or rejected phrase examples in that blocked response.

The no-garbage rules apply to every skill-authored surface: outreach copy, research brief, blocked responses, notes, summaries, QA reports, and final JSON metadata. Do not use em dashes anywhere. Blocked responses must be ASCII-only JSON or plain text. Do not say the work is ready to send, cleared for deployment, or approved for outreach. The approved status phrase is "ready for human review."

Before the first draft, do this preflight:

• Confirm each touch uses a canonical touch_type: cold_email_first_touch, cold_email_followup_2, cold_email_followup_3plus, linkedin_connection_request, linkedin_dm_post_connect, linkedin_day_of_nudge, or post_event_followup. Documented aliases from timeline labels are okay, but improvised labels are not.
• Email subjects are static, lowercase, 1-4 words, no digits, no colons, and no merge fields.
• LinkedIn touches have no subject at all. Use "" or omit the field in prose, never n/a.
• Cold first touches and post-connect DMs need a neutral how/what/why illumination question that matches the validator.
• Bodies start directly with {{first_name}}, followed by the buyer-relevant trigger. Do not open with Hi {{first_name}}, Hey {{first_name}}, or Hello {{first_name}}.
• Every touch ends with one clear lean-back question. Connection requests end with a direct connection ask.
• Asset language is allowed only when availableAssets is non-empty. Proof language is allowed only when proofPoints is non-empty.
• Event-specific coffee, meetup, booth, or side-event asks are allowed only when sender logistics were supplied.
• No pain angle repeats across the sequence, including across email and LinkedIn.
• Do not put sender names or signatures inside touch.body. Signatures belong outside sequencer import fields.

If any preflight item is missing, ask targeted questions or continue in strict no-invention mode only when the user explicitly says to proceed. Even in strict no-invention mode, the validator gate remains mandatory.

The brief must contain:

1. Buyer job, the progress this persona is trying to make.
2. Current workaround, how they probably handle the job today.
3. Hidden risk / cost of doing nothing, the expensive, awkward, or audit-facing thing they may not be seeing.
4. Customer-language pain, concrete words from case studies, customer quotes, docs, or the user's notes. Avoid marketing nouns.
5. Trigger, either a real observed signal or a truthful situation trigger. Never fake a post, session, hire, or customer.
6. Proof points, named customers, public numbers, case-study facts, or a clear statement that no proof was supplied.
7. Available lead magnets / useful assets, actual links/files/attachments the sender can truthfully include: tools, calculators, benchmark reports, checklists, field guides, audit worksheets, teardown templates, relevant blog posts, or event-specific prep notes. If none are supplied, do not promise a matrix, brief, worksheet, one-pager, report, recap, audit, or doc.
8. Event-specific ask feasibility, whether the sender is attending, sponsoring, hosting, has a booth/table, has coffee availability, or wants one event-route ask. If unknown, do not use meetup or coffee CTAs.
9. Likely objection/anxiety, the thing the buyer would silently think before replying.
10. Cadence feasibility, event start date, event end date if known, today's date, requested touch count, minimum gap, and whether any requested touch would fall in the past.
11. Pain-angle ledger, one distinct angle per planned touch. Each angle needs a 2-5 word label, source pain, illumination mechanism, and cost of inaction. Do not reuse an angle across email and LinkedIn.

If any of buyer job, current workaround, hidden risk, proof points, or available assets are unknown, ask targeted follow-up questions before drafting. If the user explicitly says to proceed without proof or assets, write in strict no-invention mode and say the sequence will avoid asset promises and customer proof.

Inputs

The user (or upstream agent) provides three things:

1. Event context, name, startDate (YYYY-MM-DD), endDate (YYYY-MM-DD if multi-day), location, optionally agenda titles, speakers, exhibitors, venue. Either inline or as a JSON file. If the user provides an event URL only, fetch it first; if you can't fetch it, ask the user for the basics.
2. Company ICP + personas, for each target persona, supply: role, seniority, buyerJob, currentWorkaround, priorities (3-5 outcomes the persona owns this quarter), painPoints (3-5 specific operational scars in their own language), hiddenRisk, objections, proofPoints, and availableAssets. Sample fixtures live at examples/*/company-icp.json.
3. Sequence parameters, leadTimeWeeks (1-8, default 4), channels (email, linkedin, or both), optional touchCount, optional minGapDays (default 4), optional preEventOnly, and sendingIdentity (sender name, title, company).

If any of these are missing or vague, ask the user for them before generating. Vague inputs produce vague output, and strict validation must reject the touches rather than letting the model invent substance.

Minimum follow-up questions

Use these when the user gives a thin request:

1. What is the sender company website, and what does it sell in plain English?
2. Who buys it, and what job are they trying to get done this quarter?
3. What are they using today or manually stitching together?
4. What goes wrong if they do nothing for 30-90 days?
5. What proof can we truthfully use, named customers, public data, customer quotes, or before/after numbers?
6. What lead magnets, tools, reports, checklists, calculators, templates, useful posts, or event-specific prep materials can we truthfully attach or link?
7. Will the sender attend, sponsor, host, have a booth/table, or have real coffee availability at the event? If yes, what exact logistics can we use?
8. What is the event URL or agenda page, and which track/session makes this outreach timely?
9. How many steps do you want? If unsure, say "default" and use the cadence planner.

For event-led outbound, it is acceptable to ask for the sender's website first and research the ICP yourself. Prefer researching before asking the user to explain basics that the website can answer.

Execute the workflow

For each persona, build one outreach sequence with a user-configurable number of touches distributed across the lead-time window. The standard gap is at least 4 days between adjacent steps. Determine today's date yourself by omitting today from the planner input unless the user gives an explicit date. Determine event dates from the user, the fetched event page, JSON-LD, or the visible event date text. A 4-week email-only sequence defaults to 6 touches, but if today is already inside the lead window, the first touch starts no earlier than today. If the user explicitly asks for pre-event only, omit day-of/post-event touches and keep the pre-event touches. Each touch is a separate generation step.

Before drafting, run the deterministic cadence planner:

node "<absolute-skill-directory>/scripts/run.mjs" plan-timeline

Pass JSON on stdin:

{
  "leadTimeWeeks": 4,
  "channels": ["email"],
  "touchCount": 6,
  "minGapDays": 4,
  "eventDates": "June 2-4, 2026"
}

The planner uses the runtime's local date when today is omitted. Prefer passing eventStartDate when available; otherwise pass the human-readable eventDates string and let the planner infer the start date. When the planner returns send_date, copy that exact date into the final files. Do not calculate calendar dates in your head. If the planner returns "isValid": false, do not draft. Explain the feasibility issue and ask whether to reduce touch count, adjust channels, change minGapDays, switch to pre/post-event mix, or accept fewer steps. Preserve the 4-day gap as the recommended default unless the user explicitly overrides it.

Use this validator path. Do not assume the current working directory is the plugin root:

node "<absolute-skill-directory>/scripts/run.mjs" validate-touch --stdin

1. Pick the channel, offset, and send date from the planner output. Do not invent dates by hand.
2. Assign a fresh pain angle from the ledger. The touch must use that angle and no previous angle. Across the whole sequence, including mixed email + LinkedIn runs, no pain angle may repeat. If you do not have enough angles, split the buyer problem into different dimensions (current workaround, ownership, timing, audit evidence, stakeholder conflict, cost of inaction, objection) or ask the user for more customer-language pains.
3. Choose CTA posture. If the sender supplied real event presence or availability, include exactly one natural event-specific ask in the sequence, usually the last pre-event or day-of touch: "Worth coffee at {{event_name}} if this is already on your list?" or "Worth a quick conversation at {{event_name}} if this is already on the roadmap?". Do not use this if attendance, booth, side-event, or availability is unknown. If relevant assets exist, include at most one asset-backed CTA in the best-fit touch. If no assets exist, do not mention assets.
4. Draft the touch following the 4T framework, the channel-specific length rule, and the hard validator rules below.
5. Validate by running node "<absolute-skill-directory>/scripts/run.mjs" validate-touch --stdin with the resolved absolute path and the touch as input. Include strictAngleDiversity: true, the current painAngle, and all usedPainAngles from prior touches. Do not invent, summarize, or approximate validator results. A touch only "passes" when the actual CLI returns "isValid": true. If it returns errors, read the errors, revise the draft, and re-validate. Up to 3 attempts. If still failing on attempt 3, stop the sequence, report the failed rule names, and ask for the missing input or manual review. Do not include the failed copy in the final answer or final files.
6. Score and band the touch: 5/5 top-tier, 4/5 ship, 3/5 review, 1.5/5 rewrite (keyed to validator pass + specificity).
7. Append to the sequence; move to the next touch.

After all touches generate, run the sequence-level validator:

node "<absolute-skill-directory>/scripts/run.mjs" validate-sequence

Pass the full sequencer-output.json on stdin or with --sequence. This catches pain-angle recycling across all channels. The sequence is not validator-clean unless both every touch validator and the sequence validator return "isValid": true.

If the user supplied confirmed sender attendance, booth, table, side-event, or coffee availability, set eventSpecificAskRequired: true on the top-level sequencer-output.json or per sequence before running validate-sequence.mjs. The validator will then require at least one natural event-specific ask. If sender event presence is unknown, leave this false and include a note: no meetup CTA because event logistics were not supplied.

After all touch and sequence validators pass, write the final output as final_sequence.md (human-readable) and sequencer-output.json (machine-readable). Include a sequence summary header: total touches, requested touch count, min gap days, average quality, score-band counts, CTA mix, illumination-question coverage, distinct pain-angle coverage, touch validator status, and sequence validator status from the actual CLI output.

The final JSON must include top-level summary, event.startDate, event.endDate, and sequencesByPersona. Every touch must include touch_slot, offset_days, send_date, channel, touch_type, subject, body, cta_type, snake_case pain_angle, quality_band, validation_errors, and the actual checks object from validate-touch.mjs. Do not use camelCase painAngle in the final file. validation_errors must be an empty array on successful final touches.

Then run the final artifact validator:

node "<absolute-skill-directory>/scripts/run.mjs" validate-artifact --sequence sequencer-output.json --final final_sequence.md

The sequence is not finished unless this returns "isValid": true. If any validator cannot run or any touch still fails after retries, do not write final sequence files. Write a QA report with failed rule names instead. Never call the sequence "ready to send" or "cleared for deployment"; use "ready for human review."

Treat the cold-outbound craft canon as the canonical playbook. Re-read it any time you are unsure how to handle an edge case such as multilingual events, dinner-invite touches, or very short lead times.

Before the sequence, include the Outbound Research Brief in the response so the user can see the buyer job, hidden risk, proof, assets, and assumptions. The brief is part of quality control, not extra decoration.

Voice (read this before drafting anything)

The bar is authentic, honest, empathetic, no theatrics. Not slang. Not coolness-signalling. Not corporate-sales-speak either. Write as a real person who has worked the job and is genuinely curious whether the recipient is dealing with the same thing.

Concrete rules that follow from that:

• Don't fabricate a personalization signal. If you didn't actually see the prospect's panel, post, hire, customer quote, or session, don't write "saw your X" or "noticed you're Y" or "caught my eye". Forced personalization reads as fake the moment the recipient checks. When you have no real signal, use a situation trigger sized to the role ("end of Q3 and the CFO is locking the FY27 plan", "the Q4 chargeback target is about to land in the CFO review"). The situation is honest; the fake signal isn't.
• Poke the bear via pain illumination + cost of inaction, not theatrics. Name the specific tradeoff the persona is managing this quarter, then quietly draw the line to what happens if it goes unaddressed (Q4 board deck, regulator audit, CFO QBR). No drama. No "this is wild." No "honestly?". Just the operational reality the persona is already living.
• Empathy over swagger. When the persona is dealing with a hard tradeoff, name it honestly. "Tightening the rules drops approval rates 4-7 points and Sales escalates inside 48 hours" is honest empathy. "Most VPs walk in already knowing the pitch" is performative.
• No slang and no faux-casual swagger. Banned tells: caught my eye, caught my attention, wall-to-wall, no worries, no biggie, hot mess, needle in a haystack, fire drill, all hands on deck. The validator catches these. The deeper issue: copy that needs slang to sound human is masking weak substance.
• Bodies use proper grammar. Capitalize the first letter of every sentence. Use full punctuation. The "all-lowercase" convention applies only to subject lines (per the cold-email-subject canon). Body copy is not a Slack message.
• No generic greeting wrappers. Start with {{first_name}}, and the buyer's situation. Hi {{first_name}} wastes the preview line and makes every touch feel templated.
• Protect the first 18 words. The inbox preview is where the recipient decides whether to keep reading. For cold first touches and post-connect DMs, the opener must be buyer-first: no I/we/us/our in the first 18 words, and no event-first opener like "Black Hat is coming up...".
• Don't force the event into the body opener or CTA. The event is the occasion for outreach; it is not the buyer's reason to care. The strongest touches anchor on a persona responsibility first, then use the event only when it makes the ask more natural ("Worth a coffee at Black Hat if this is already on your audit list?"). Forcing "before RSA" / "week of Money20/20" / "into m2020" into every sentence reads like a sequence template. Forcing "before Amsterdam" or "for Amsterdam prep" into the CTA is worse: it makes the city pretend to be the buyer priority.
• One event-specific ask is good when it is true. If the sender is actually attending, sponsoring, hosting, or available for coffee, one touch should usually make the event route explicit: "Worth coffee at Money20/20 if this is already on your audit list?". Do not sprinkle event CTAs across the sequence. One is enough.
• Use human shorthand for proof references. Stiff: "Adyen and Marqeta wrote up a same-week false-positive cohort review that compresses the loop. There is a write-up if useful." Better, the way someone would actually say it: "Attached a one-pager from our work with Adyen and Marqeta on this. Is the false-positive review on your roadmap this quarter?" Shorthand reads as a peer talking, not a vendor reading from a script. Patterns that work:
  • "Attached a one-pager from our work with [A] and [B] on [topic]."
  • "I attached the writeup from [A]'s review last quarter."
  • "[A] and [B] hit [number]. I attached the short version."
• Do not ask permission to send the useful thing. If the asset is real and useful, attach it or link it. Banned because they add fake friction: "should I send", "can I send", "want me to send", "want the one-pager", "happy to send". Better: "I attached the worksheet. Worth a coffee at Black Hat if this is on your audit list?" If the asset does not exist, do not mention it.
• Proof is sacred. Never invent "three orgs", named customers, before/after numbers, or "peer teams" because the framework wants third-party validation. If proof is missing, ask for it. If the user says none exists, use a mechanism or buyer-risk sentence instead and mark the proof gap in the brief.
• Close with a real question, not a polite ritual. "Worth a look?" is fine but overused. Prefer natural buyer-timing asks: "Worth looking into?", "Open to taking a look?", "Worth a closer look?", "Is this on your roadmap, or parked for later?", "Does this belong in the roadmap conversation?". Do not close with "Open to a look?"; it sounds clipped and unnatural. Do not bolt a location onto the ask unless the ask is literally about meeting there.
• LinkedIn is not exempt from CTAs. Connection requests still need a direct connection ask: "Open to connecting?" or "Worth connecting?". DMs and nudges need a clear lean-back question at the end. Do not end with a clever observation or an "I attached..." statement and make the reader guess how to respond.
• Reread aloud. If you'd be slightly embarrassed sending this to a peer, the copy is wrong. Rewrite.

The 4T framework

Every cold email and post-event email follows this shape. LinkedIn DMs follow it loosely. LinkedIn connection requests compress it to 2 sentences (Trigger + Talk).

1. Trigger, why this person, why now? A specific observation or situation. Patterns: "noticed [observable thing] which suggests [deduction]", "saw you're [doing X] at [event]". If you have no signal, use a situation trigger sized to the role (e.g. "end of Q3 and the CFO is locking the FY27 plan"). Never open with a population claim ("Most teams…", "Most VPs…", "In our experience…", "Many fintechs…"). Auto-rejected.
2. Think, the illumination question. A neutral how/what/why-are-you question that shines a light on a problem the persona owns. Not a leading question. Auto-rejected: "if I could…", "would you be interested?", "wouldn't you agree?", "don't you think?". Required for cold emails and post-connect DMs.
3. Third-party validation, let other people toot your horn. One sentence. Pattern: "[Peer A] and [Peer B] [outcome] [SHARP NUMBER] compared to [old number] before." Use real published peer references, customer-approved proof, or supplied case-study facts. If no proof exists, stop and ask for proof before drafting; if the user explicitly says to proceed, omit third-party validation rather than fabricating it. Banned: "we're the best", "industry-leading", "world-class".
4. Talk?, interest-based CTA, with a question mark, lean-back energy. Approved shapes: "Worth looking into?", "Open to taking a look?", "Worth a closer look?", "Is this on your roadmap, or parked for later?", "Does this belong in the roadmap conversation?", "What do you think?". For connection requests, use "Open to connecting?" or "Worth connecting?". Banned: "Open to a look?", "Got 15 minutes?", "Book a call", "schedule a meeting", "calendar link", "worth pressure-testing before [city]", "useful for [city] prep", and the over-deployed "comparing notes" / "compare notes" / "open to comparing notes".

Channel-specific length rules (hard)

These are enforced by scripts/validate-touch.mjs. Read data/cold-outbound-rules.json for the canonical numbers.

Subject lines: ≤ 4 words, all lowercase, no digits-only buzzwords, no banned subject buzzwords (AI, platform, leverage, transform, unlock, …). Email subjects are static text, do not put merge fields inside them.

Timeline (default 4-week lead time, both channels)

Both-channel default

Email-only default

If the user says "pre-event only", use T-28, T-21, T-14, and T-7 only. Otherwise event-led outreach includes day-of and post-event touches by default.

For other lead times, touch counts, dates, or single-channel cadences, use scripts/plan-timeline.mjs; do not adjust proportionally in your head.

Hard validator rules (auto-rejected)

Every touch is run through scripts/validate-touch.mjs. On failure, the script returns JSON with each failed rule. Read the errors, fix the draft, re-validate.

• Subject ≤ 4 lowercase words, no digits in cold-email subjects, no banned subject buzzwords.
• Unknown or improvised touch_type values are rejected. Use only the canonical keys in the channel table or the documented timeline aliases.
• LinkedIn touches must not include a subject. Empty or omitted is correct; "n/a", event names, and label-style subjects are rejected.
• Body within the channel length rule (above).
• Every generated touch must close with a clear lean-back CTA question. Do not end with an asset statement, clever observation, or implied next step.
• No comma-spliced asset CTAs like "I attached the review, worth looking into?". Use a clean CTA sentence when sentence count allows it, or rewrite the final question around the buyer's decision.
• LinkedIn connection requests must close with a direct connection ask such as "Open to connecting?" or "Worth connecting?".
• LinkedIn DMs and nudges must close with a clear lean-back CTA question such as "Worth looking into?", "Open to taking a look?", or "Does this belong in the roadmap conversation?".
• Do not put signatures in body. Daniel, Maya, job titles, and sign-off blocks should not be inside the sequencer JSON body.
• No awkward comma-plus-gerund clauses such as a sentence that tacks on , and talking... or , and comparing.... Rewrite the idea as a direct sentence.
• Cold first touches + post-connect DMs must keep the first 18 words buyer-first: no seller pronouns (I/me/my/we/us/our) and no event-first opener.
• Cold emails + post-connect DMs must contain a how/what/why-are/do/is-you/your illumination question.
• No leading questions: if I could…, would you be interested, wouldn't you agree, don't you think.
• No em-dashes. Use commas, periods, colons, parens.
• No exclamation marks. No emoji.
• "you/your" must outnumber "we/our" in the body.
• No banned phrases, see additional_banned_phrases in data/cold-outbound-rules.json. Includes happy to send, should I send, can I send, want me to send, want the one-pager, 15 minutes, 30 minutes, calendar link, book a call, schedule a meeting, cutting-edge, industry-leading, world-class, etc.
• No forced event phrasing: "keeps coming up before RSA", "week of Money20/20", "today at RSA", "into m2020", a question that bolts "before [event]" onto the end, or a CTA that makes the location do the buyer-priority work ("worth pressure-testing before Amsterdam", "useful for Amsterdam prep"). Use buyer responsibility as the reason to write and the event as the route to a clear ask.
• No generic post-event pleasantries: "hope the event went well", "hope the event was productive", "hope the week in [city] went well". Start from the buyer's returned-to-desk work instead.
• No invented sender logistics: do not say "I'm around", "I am around the [track] side of the agenda", "I'll be at booth X", or name a session/track unless the user supplied that fact. If sender availability is missing, use a buyer-timing CTA instead of a meetup CTA.
• No pain-angle recycling. Every touch needs painAngle metadata and must use a different buyer problem, consequence, or illumination route than all previous touches. This applies across email and LinkedIn in the same sequence. Rewording the same pain or repeating a concrete pain-anchor phrase from an earlier touch is still a failure.
• No sequence-mechanics openers. Do not write "separate thread", "sent a note", "earlier note", "previous note", or "following up on my...". Follow-up steps should stand alone with a fresh buyer-relevant angle, not announce how the sequence is organized.
• No vague calendar-trigger openers. Do not write "this is usually the week...", "new rails get attention", or any filler that explains the send date instead of the buyer's job. Name the buyer-side consequence directly.
• No invented event logistics. Do not invent hosted roundtables, coffee times, speaker-lounge locations, Business Hall locations, or prior conversations. Only use those when the user supplied them.
• No invented buyer state. Do not write as if the buyer's auditor flagged something, the buyer is carrying a named gap, or the buyer discussed something with the sender unless that fact was supplied.
• In strict mode, no unsourced assets or proof. If the touch mentions an attached/linked/pulled-together asset, pass availableAssets. If it uses named customers, peer teams, or before/after numbers, pass proofPoints. Otherwise the validator must reject it.
• No LLM-cliché phrases, 200+ phrases across 10 categories. See llm_cliche_blocklist in the same file. Categories: performative_empathy, generic_compliments, sales_speak_openers, manufactured_intimacy, marketing_buzzwords, cold_email_overused, lazy_generalization_openers, llm_transition_tics, gpt_vocabulary. (hedge_softener_warnings is soft, does not fail.)
• No anti-flex selling tics: "no deck", "no demo", "no calendar invite", "no follow-up deck", "reply yes and i'll send". These read as soulless-selling-with-extra-steps. If you'd write one, just don't include the assurance, write copy that doesn't need it.
• No floor/booth-shopping framing in cold copy. Don't talk about "47 fraud-platform vendors", "five vendors with real numbers", "which booths can answer X". The recipient cares about the risk they manage, not your vendor census.
• One problem per email. Don't mash multiple value props.
• Address the recipient with {{first_name}} and reference their company as {{company}} at least once. Never hard-code real names or companies in the output, those are merge-field substitutions that happen at send time.
• Never say the sequence is ready to send, cleared for deployment, or approved for outreach. This skill produces drafts for human review.

Validating a touch from inside this skill

node "<absolute-skill-directory>/scripts/run.mjs" validate-touch --touch <(cat <<'JSON'
{
  "subject": "money20/20 chargeback prep",
  "body": "Neha, ...",
  "channel": "email",
  "touch_type": "cold_email_first_touch",
  "eventName": "Money20/20 USA 2026",
  "personaPriorities": ["ship a measurable reduction in chargeback rate to the CFO before Q4 close"],
  "personaPainPoints": ["rules-vs-models false-positive tradeoff..."],
  "strictTruth": true,
  "availableAssets": ["field-level worksheet approved for this campaign"],
  "proofPoints": ["Adyen and Marqeta public case-study comparison, 94% vs 12% inbox placement"],
  "strictAngleDiversity": true,
  "painAngle": {
    "label": "stale rule ownership",
    "sourcePain": "stale rules keep tier-1 closing the same detection",
    "mechanism": "rules ownership drifts between product and risk",
    "costOfInaction": "audit answer gets harder to defend"
  },
  "usedPainAngles": []
}
JSON
)

Returns { "isValid": true | false, "errors": [...], "checks": {...} } on stdout. Exit code is 0 either way (errors are part of the contract, not a process failure).

Output structure

Write two files per run, in a directory named for the event:

• final_sequence.md, human-readable. One section per persona; one subsection per touch with channel, offset, type, body, and quality band. Sequence summary at top.
• sequencer-output.json, machine-readable. Top-level shape: { "summary": {...}, "event": { "startDate": "YYYY-MM-DD", "endDate": "YYYY-MM-DD" }, "sequencesByPersona": { [personaId]: { personaId, leadTimeWeeks, channels, touches: [...] } } }. See src/types/index.ts for the full type. Every touch carries send_date, offset_days, snake_case pain_angle, its checks object, and validation_errors: [].

Working examples that exercise the full contract:

• examples/black-hat-usa-2026/, cybersecurity, Black Hat USA 2026, Director of Security Engineering + VP Security personas.
• examples/money2020-europe-2026/, fintech, Money20/20 Europe 2026, VP Risk and Fraud + Head of Compliance personas.
• examples/singapore-fintech-festival-2026/, fintech APAC, fixture inputs only (regen target).

Voice, the bar

Every touch must read like a text from a smart peer who noticed something specific and is genuinely curious whether the recipient is dealing with the same thing. Not a vendor. Not an AI. Not a polished sales rep. A peer.

Concrete tells of getting it wrong:

• Polished/symmetrical sentences. Real human copy is asymmetric and a little terse.
• Hedge stacks: "Just wanted to see if you might possibly have a moment to perhaps connect." Cut all of them.
• Abstract pain ("attribution challenges", "scaling outbound") instead of specific scars ("the $250K RSA budget line reads as sourcing $0 because of the 90-day last-touch window").
• Ending with a sales-coded reassurance ("no deck, no demo, no follow-up"). If you wouldn't say it on a coffee invite, don't write it.

When you're unsure if a draft is good, read it aloud. If you'd be embarrassed sending it to a peer, rewrite it.

Source files

• data/cold-outbound-rules.json, channel rules, banned phrases, cliché blocklist.
• data/cold-email-benchmarks.json, subject + body benchmarks.
• data/cold-outbound-craft.md, full canon, the playbook this skill encodes.
• data/cold-outbound-frameworks.md, 4T worked examples + voice anti-patterns.
• data/cold-outbound-canonical-examples.json, pass + fail reference touches.
• src/types/index.ts, full TypeScript types for EventContext, CompanyICP, OutreachTouch, OutreachSequence, SequencerOutput.
• scripts/validate-touch.mjs, the validator CLI.
• scripts/validate-sequence.mjs, sequence-level pain-angle and cadence validator.
• scripts/validate-artifact.mjs, final JSON + markdown artifact validator.

Optional: headless runs

For CI, batch generation, or scheduled runs outside an installed plugin session, use src/agents/sequencer.ts, which exposes generateSequence() with an injectable TouchGenerator. Bring your own LLM adapter. Interactive use in Claude or Codex requires no separate model API key.