convez.ai
SkillMarketing & VendasIntermediário
luminik-io

Event Outbound Skill

luminik-io/event-outbound-skill

Sequências de outbound validadas para eventos B2B e trade shows

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que gera sequências de outbound validadas para eventos B2B, combinando emails e campanhas LinkedIn. Transforma dados de evento, ICP, identidade do sender e prova em múltiplos toques pré-evento, durante e pós-evento, validando cada mensagem contra regras rígidas antes do envio.

Instalação

$npx convez-ai luminik-io/event-outbound-skill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
37%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar campanhas de outreach para attendees de trade shows e conferências
  • Criar sequências multi-toque de email + LinkedIn com validação automática
  • Preparar campanhas de convite para eventos privados, jantares e reuniões com speakers

Veredito

Skill especializado e bem-estruturado para AEs, SDRs e founders que precisam escalar outbound para eventos B2B com qualidade. Diferencia-se por validação obrigatória de cada toque contra regras concretas, não apenas templates genéricos — ferramenta real para quem trabalha com sales acceleration.

README

A tradução para o português deste README ainda está na fila.

event-outbound

Buyer-first email and LinkedIn outreach for B2B trade shows and conferences. Claude Code + Claude Cowork. Free, MIT, open source.

License: MIT Claude plugin Tests Made by Luminik

Install · What it does · Worked examples · Validation rules · Why use this · Credits


Built on 20,000+ personalized touches across 50+ B2B events that sourced $6M+ in pipeline. Distilled from four years of fintech IDV and cybersecurity outbound run by hand.

The skill turns event, ICP, sender identity, proof, real campaign assets, and cadence constraints into a full multi-touch sequence per persona, pre-event, day-of, post-event. Every touch is validated before it lands: subject ≤ 4 words, buyer-first inbox preview, channel-specific body length, no buzzwords, illumination question on first touch, direct CTA, no permission-to-send gating, no invented assets, no unsourced proof, and no recycled pain angles across steps. Failed touches retry up to 3× with validator feedback. If they still fail, the skill stops and writes a QA report instead of shipping unchecked copy.

Install

Claude Code

From any Claude Code session:

/plugin marketplace add luminik-io/claude-plugins
/plugin install event-outbound@luminik-plugins

That's it. The skill registers itself, and any prompt mentioning a B2B event, attendee outreach, or pre-event sequencing routes to it.

Claude Cowork

Install event-outbound from Cowork's plugin directory once listed. In Cowork, open Customize, choose Browse plugins, then install event-outbound.

The same plugin package is used for Claude Code and Cowork. The local validator is referenced through ${CLAUDE_PLUGIN_ROOT}, so it resolves inside the installed plugin directory rather than depending on the current working directory.

Who this is for

AudienceWhat you get
AEs and SDRs working a trade-show attendee list four weeks outMulti-touch email + LinkedIn sequence per persona, ready for human review before loading into Apollo, Outreach, Salesloft, Instantly, or Smartlead
Event marketers running dinner invites, side events, speaker meet-and-greets, booth-visit campaignsChannel-appropriate copy that gets opened and RSVP'd, not blasted and ignored
Founders doing their own outboundThe same workflow and the same validated output, without hiring an SDR

Narrowed for fintech (identity verification, payments, regtech) and cybersecurity. The skill works outside those verticals, but the worked examples and the canonical pain-point library are tuned for those buyers.

What makes this different

Every cold-email generator claims "proven frameworks." This one validates every touch against hard rules before it lands.

LayerWhat it checks
SubjectAll lowercase, ≤ 4 words, no colons, no digits, no buzzwords
Inbox previewFirst-touch and post-connect DM openers must be buyer-first: no seller pronouns in the first 18 words, no event-first opener
Body lengthChannel-specific: cold email 50–100 words / 3–5 sentences, LinkedIn connect 18–35 words ≤ 200 chars, day-of nudge 30–60 words, post-event 40–90 words
Structure4T pattern: Trigger → Think (illumination question) → Third-party validation → Talk? (direct CTA)
Pronoun ratio"you/your" must outnumber "we/our"
No em-dashes, exclamation marks, or emojiHard-rejected
CTA clarityEvery touch must close with a clear lean-back question. Connection requests need an explicit connection ask like "Open to connecting?"
Event ask postureWhen sender event presence is real, one touch should include a natural event-route ask. When presence is unknown, meetup and coffee CTAs are blocked
CTA rankingmake_offer > ask_for_interest > ask_for_problem > ask_for_meeting (CTA-type reply-rate deltas from the Gong / 30MPC / Outbound Squad 85M-email report)
Cadence structureUser-configurable touch count, 4-day minimum gap by default, date-aware planning so steps do not land in the past
Angle diversityEvery touch carries pain_angle metadata; the sequence validator rejects repeated pain, repeated angle labels, repeated pain-anchor phrases, and high-overlap pain vocabulary across email and LinkedIn
Artifact gateFinal JSON must include summary, event dates, send_date, offset_days, snake_case pain_angle, checks, and empty validation_errors; final markdown is scanned for banned surface phrasing
Cliche blocklistTen categories, 195 phrases. See Validation rules below
No sequence mechanicsFollow-ups cannot open by announcing the outreach thread; each step has to stand on its own buyer-relevant angle
SpecificityEvery touch must reference a concrete persona priority, pain, or event signal, with no population-shape generalizations, forced event phrasing, or location-pasted CTAs
Strict truthIn strictTruth mode, asset promises require availableAssets, proof claims require proofPoints, and Apollo-ready {{first_name}} / {{company}} fields are required

What it does

You hand the skill five things:

  1. Event, name, dates, agenda, speakers, exhibitor list.
  2. ICP, industry, size range, website, and one or more buyer personas with concrete priorities and pain points (vague aspirations like "build the brand" or "scale the team" fail the specificity check).
  3. Buyer research, buyer job, current workaround, hidden risk, likely objections, and customer-language pain.
  4. Truth sources, proof points and assets the sender can truthfully attach or link, including lead magnets, tools, calculators, checklists, reports, templates, or event-specific prep materials.
  5. Sequence params, lead time in weeks (1–8, default 4), channels (email, LinkedIn, or both), optional touch count, minimum gap days, event dates, today's date, sending identity.

If proof or assets are missing, the skill asks for them before drafting. If the user explicitly proceeds without them, strict mode writes around the gap instead of inventing matrices, briefs, peer teams, or before/after numbers.

If the sender is confirmed to attend, sponsor, host, or have real availability at the event, the skill aims for one natural event-specific ask, for example a coffee or short conversation at the event. If attendance or availability is unknown, it avoids meetup CTAs and says why.

It returns an Outbound Research Brief plus a full sequence per persona. Six touches for a four-week email-only run by default, configurable when the user wants more or fewer steps. The cadence planner enforces at least four days between adjacent touches, infers the event start from structured or human-readable dates, and uses the runtime's local date when today is not supplied.

What the output looks like

A single validator-clean touch from the checked-in Claude showcase. This run was strict no-invention mode: no approved proof, no approved assets, and unknown sender attendance.

Touch 2 · T-4d · email cold

Subject: evidence pack

{{first_name}}, audit evidence at {{company}} can become a scavenger hunt
when the auth log, rail trace, and case notes live in different systems. How
quickly can your team reconstruct why a payment moved after authentication
changed? If that evidence pack still takes manual stitching across instant
rails and cards, is the gap worth looking into?

channel: email · offset: -4d · type: cold_email_followup_2 · cta: ask_for_interest · words: 55 · pain_angle: evidence pack latency

Apollo-ready merge-field syntax. The opening sentence names a specific evidence-chain pain, not a population-shape generalization. The CTA is a clean buyer-timing question. Because no approved assets or sender logistics were supplied, the touch avoids attachments, proof claims, and meetup language.

Worked examples

ExampleICPPersonasLead timeStatus
examples/black-hat-usa-2026/Cybersecurity (mid-market SaaS buyer)Director of Security Engineering + VP Security4 weeksPre-rendered, validator-clean
examples/money2020-europe-2026/Fintech (payments + neobank buyer)VP Risk and Fraud + Head of Compliance / KYC Operations4 weeksPre-rendered, validator-clean
examples/singapore-fintech-festival-2026/Fintech IDV(input fixtures)(n/a)Regenerate inside Claude with no API key
examples/claude-showcase/Real Claude runsPositive + guardrail casesMixedChecked-in outputs with deterministic showcase checks

Every shipped sequence is hand-verified against the full validator stack: zero hits across the ten cliche categories, channel-length compliance, illumination-question coverage, pronoun ratio in favour of the reader.

Quickstart

From Claude Code or Cowork, after installing the plugin:

Create an outbound sequence for Black Hat USA 2026 targeting Directors of Security Engineering at mid-market SaaS.
4 week lead time, email plus LinkedIn.

The skill picks up the request, researches public company/event context when URLs or company names are available, asks for missing proof/assets, and returns a full SequencerOutput plus a rendered markdown preview ready to paste into your sequencer.

Local development

git clone https://github.com/luminik-io/event-outbound-skill.git
cd event-outbound-skill
npm install
claude --plugin-dir $(pwd)

That's it. The installed skill runs inside Claude Code and Cowork with no extra API keys. Claude reads the rules from ${CLAUDE_PLUGIN_ROOT}/data/, generates each touch, validates it via node "${CLAUDE_PLUGIN_ROOT}/scripts/validate-touch.mjs", revises on failure, checks the full sequence, then runs the final artifact gate before writing deliverables.

To validate a single hand-written touch against the rule set:

echo '{"subject":"...","body":"...","channel":"email","touch_type":"cold_email_first_touch"}' \
  | node scripts/validate-touch.mjs --stdin

To run the full validator scan against every shipped artefact:

npx tsx scripts/scan-deliverables.ts

Run the tests

npm test -- --run

139 tests across 7 files (cliche-validator unit tests, strict context checks, date-aware timeline computations, installed-skill timeline CLI, sequence-level pain-angle validation, source-grounded craft evals, persona analyser, event scraper, end-to-end evals). Vitest, ~6 seconds cold.

The repo also includes real Claude-generated showcase outputs:

npm run check:showcase

This deterministic check reads examples/claude-showcase/, verifies cadence dates, validator-clean touches, strict no-invention output, and guardrail behavior for thin input, impossible cadence, and wrong-persona prompts. It does not call Claude, so it is safe for CI.

To validate pain-angle diversity for a full generated sequence:

node scripts/validate-sequence.mjs --sequence examples/claude-showcase/rich-positive-availability-unknown/sequencer-output.json

To validate a final generated JSON + markdown pair:

node scripts/validate-artifact.mjs \
  --sequence examples/claude-showcase/rich-positive-availability-unknown/sequencer-output.json \
  --final examples/claude-showcase/rich-positive-availability-unknown/final_sequence.md

To rerun the live Claude showcase locally:

npm run e2e:claude -- --case rich-positive-availability-unknown

Add --update-fixtures only when you intentionally want to refresh the checked-in showcase outputs after human review.

For broader local Claude coverage, run the matrix runner:

npm run e2e:claude:matrix -- --mode lite --max-cases 24 --command claude
npm run e2e:claude:matrix -- --mode validated --max-cases 5 --command claude

lite disables tools and verifies the skill blocks drafting when the validator cannot run. validated asks Claude to generate one touch per case, write validated-touch.json, and pass it back through scripts/validate-touch.mjs. Use --command for whatever local Claude command should be tested on your machine. Run artifacts and rollups are written under .tmp/claude-matrix-live/.

For full installed-skill runs that generate both final files and then run the artifact validator:

npm run e2e:claude:full -- --command claude --max-cases 2

Use --command for the authenticated Claude Code command on your machine. Run artifacts and rollups are written under .tmp/claude-full-sequence-live/.

Headless / batch generation (optional)

If you want to generate sequences outside Claude (CI, scheduled cron, batch backfill), src/agents/sequencer.ts exposes generateSequence() with a required injectable TouchGenerator. Bring your own LLM adapter. There is no required cloud API for using the skill inside Claude Code or Cowork, and the repo ships no default external-model dependency.

Parameters

Full TypeScript types in src/types/index.ts. The short version:

InputRequired fieldsNotes
EventContextname, dates, optional startDate, optional endDate, location, agendaTitlesSpeaker + exhibitor lists improve specificity but are optional
CompanyICPindustry, sizeRange, personas[]Optional website, productSummary, proofPoints, and availableAssets improve strict-mode output
AttendeePersonapersonaId, role, seniority, priorities[], painPoints[]Optional buyerJob, currentWorkaround, hiddenRisk, objections, proofPoints, and availableAssets prevent generic copy
SequenceParamsleadTimeWeeks (1–8), channels, optional touchCount, optional minGapDays, optional today, sendingIdentity4 weeks and 4-day gaps are the default

Output shape

type SequencerOutput = {
  summary: {
    totalTouches: number;
    requestedTouchCount?: number;
    minGapDays?: number;
    validatorStatus: string;
    sequenceValidatorStatus: string;
  };
  event: {
    name: string;
    startDate: string;
    endDate?: string;
    location?: string;
  };
  sequencesByPersona: {
    [personaId: string]: {
      personaId: string;
      touches: OutreachTouch[];
      leadTimeWeeks: number;
      channels: ('email' | 'linkedin')[];
    };
  };
};

Each OutreachTouch carries send_date, offset_days, snake_case pain_angle, checks, and validation_errors: [] so you can see exactly why it passed. Final sequence files are written only after validator-clean touches, sequence validation, and artifact validation. If a touch burns all three validation retries, the installed skill stops and reports the failed rule names for human review.

Strict mode adds missingMergeFields, assetPromiseHits, and proofClaimHits to the checks block so a reviewer can see whether Claude tried to invent a useful-sounding but unsourced claim.

Angle diversity adds pain_angle to every touch, plus reusedPainAngleHits and painAngleBodyOverlap in the checks block. A sequence only passes when validate-sequence.mjs sees one distinct pain angle per touch across all channels.

Validation rules

The cliche blocklist in data/cold-outbound-rules.json defines ten categories totalling 195 phrases. Nine are hard-banned in skill output; one is soft-warned for human review.

CategoryPhrasesRepresentative examples
performative_empathy11"stuck with me", "really resonated", "got me thinking"
generic_compliments7"amazing work", "love what you're doing", "curious to hear your thoughts"
sales_speak_openers21"hope this finds you well", "circling back", "touching base", "just checking in"
manufactured_intimacy8"in case it's helpful", "no pressure at all", "if it's of interest"
marketing_buzzwords32"leverage", "unlock", "transform", "seamless", "synergy", "drive growth"
cold_email_overused8"teardown" (as 1-pager), "playbook" (as marketing word), "blueprint", "north star", "table stakes", "low-hanging fruit", "double-click on", "do you have bandwidth"
lazy_generalization_openers51"Most teams...", "Most fintechs...", "Most VPs I talk to...", "Almost nobody...", "Nobody is...", "Everyone is...", "In our experience..."
llm_transition_tics20"Moreover,", "Furthermore,", "Additionally,", "It is worth noting that"
gpt_vocabulary29"delve", "tapestry", "navigate the landscape", "in today's fast-paced world"
hedge_softener_warnings8"I think", "perhaps", "it seems" (soft-warning, not auto-rejected)

Sources cited inline in data/llm-cliche-blocklist.md: Stanford CRFM evaluations of GPT-detected vocabulary, GPTZero linguistic markers, the Lavender Live transcript, the Gong / 30MPC / Outbound Squad 85M-email "Ultimate Cold Email Data Report", and the workspace voice canon.

Why use this over alternatives

You're considering...Where it falls short for event outbound
Apollo's built-in AI email writerGeneric templates per persona; no event-specific context (agenda, speakers, exhibitor neighbours). No validator.
A custom GPT or Claude prompt you wrote yourselfWorks, but every output passes through whatever LLM cliches the base model is trained on. No reproducibility, no per-touch quality flag.
Hand-writing each sequenceThe hand-written sequence is the bar this skill is calibrated against. The skill exists so you can run that bar across 10 personas in one afternoon instead of one persona per week.
A general-purpose copywriter or AE training courseBuys you craft, not throughput. This skill encodes the craft into validation rules and applies them to every touch automatically.

Project layout

.
├── src/                    Sequencer agent, validators, timeline, types
├── data/                   Validator inputs: cliche blocklist, channel rules,
│                           cold-email benchmarks, craft canon
├── examples/               Worked examples (Black Hat, Money20/20, Singapore Fintech)
├── tests/                  Vitest unit + integration tests
├── evals/                  End-to-end output evaluations
├── scripts/                Run-example, scan-deliverables, install verification
├── marketplace/            Cover image (1200x630)
└── .claude-plugin/         Claude plugin descriptor

License

MIT. See LICENSE.

Credits

Teachers and sources we learned from:

  • Josh Braun, whose public writing on buyer-first cold outbound has been a compass. The validator's tone rules (no pitch-speak, "you" > "we", concrete offers over meeting-asks) echo principles he teaches openly.
  • Gong's "Ultimate Cold Email Data Report", 85M emails analysed, co-authored with 30MPC and Outbound Squad. The benchmark numbers we validate against (subject length impact, CTA-type reply-rate deltas, word-count sweet spots) come from this publicly-published research.
  • Stanford CRFM and GPTZero for public research on GPT-detected vocabulary and linguistic markers, which seeded the LLM-cliche blocklist.
  • The Lavender Live transcript, for one of the few public corpora of in-the-moment cold-email rewrites by working SDR managers.

This plugin does not redistribute any proprietary content. It encodes general craft principles from publicly-taught material and published research into validation rules that run at generation time.


Visit the skill page on luminik.io →

Luminik is a product of DataRavel Inc. (Newark, DE).

Ver o original no GitHub

Bom para

Relacionados