Video summary

Cборка сложного b2b cold outreach в Claude Code

Main summary

Key takeaways

Technology

Overview (what the video demonstrates)

  • A B2B cold outreach “assembly” workflow in Claude Code, targeting a complex market with multiple stakeholders.
  • The creator claims near end-to-end campaign generation with roughly ~2 minutes of active input (pressing buttons / approving steps), while automation handles the rest (the system “watches” and executes).
  • The demo uses a case study designed for a mid-market illustration art vendor: hand-made character art for large studios across PC / PlayStation / Xbox.

Key technological/product concepts & workflow stages

The outreach is generated via a structured pipeline/methodology built from four main “skills” (autopilot stages):

  1. Assessment / qualification
  2. Research
  3. Copywriting
  4. Company assembly (creating outbound items and routing them into the outreach sequence)

Integration and automation notes

  • Skill loading / skills archive

    • Claude Code skills are loaded from a local “skills archive”, unpacked inside a Claude Code project folder.
    • The creator argues that downloading a ready-made skill from a repository won’t achieve meaningful personalization in complex markets. Effective results require preparing/testing prompts and skills for the target use case.

  • Data enrichment and source connections

    • Mentions integration via MCP to connect to external tools and data sources.
    • Talks about enrichment through linked systems (with an optional AI-extraction-style step).
    • Also mentions the use of multiple datapoints (e.g., parsing a digital footprint such as Steam titles and interviews).

Demo setup: target selection & scaling

  • The workflow starts with a dataset of ~200 potential target companies.
  • It selects 10 companies to build a campaign for.
  • A disqualification/qualification gate with “strict criteria” avoids mismatched niches (e.g., excludes “grey niches” and mobile/session-based games).
  • The demo narrows to 8 qualifying companies, with disqualification reasons noted.

Automation timing & active/passive effort

  • The creator sets an expectation of fast execution:
    • Assessment: ~15 minutes (measured on the clock).
    • The overall run misses a “20 minutes” goal, later reported as closer to ~30–32 minutes total.
  • Emphasis: active human time is mostly limited to starting/stopping stages and handling the system’s output.

Personalization logic (multi-stakeholder outreach)

  • The output is multi-stakeholder: different contacts/personas inside each company.
  • The system generates individualized emails rather than a single template with substitutions.
  • Constraint/workaround:
    • Claude Code can connect to tools via MCP, but (as described) can’t create custom fields in Instantly directly.
    • Workaround: generate content in a structured format, then pass it through MCP to Apollo (and/or other sequencing tools), producing unique letters per person.

Stakeholder handling approach

  • Uses a contact/title matching mask (a position-title “cloud”) to map contacts to relevant stakeholder roles.
  • Mentions behavior around multilingual history, where elements like studio names/entries appear in different languages (e.g., French/Polish examples).
  • Demonstrates stakeholder replacement when needed—e.g., swapping in a related role (such as a board games director) while maintaining relevance to computer games.

Tooling / stack mentions

  • Claude Code: orchestration layer.
  • Sequencer: mentions Apollo as the final sequencing stage (and notes it can connect other sequencers later).
  • Whisper: referenced as being used within the project context (from the Claude Code folder).
  • Instantly / email outreach providers:
    • The workflow is described as avoiding writing to these directly, instead generating via intermediates.

External enrichment tools/providers

  • Notes using Lemodatu as a backup.
  • Mentions it may have been unavailable during recording due to SSL/certificate or other issues.

Token / cost & performance considerations

  • The creator discusses token usage/cost and argues that spending tokens for accuracy is appropriate for enterprise/midmarket use cases—not “optimizing away” at the expense of quality.
  • Reports an example of token burn, with outputs like ~200,000 tokens gone during the run.

Review / guide-style claims: “what actually matters”

Core takeaway (how it works)

Claude Code automation works well when you:

  1. Pre-build skills (prompts + pipelines) for the target market and the realities of available data.
  2. Test prompts against what data exists vs. what’s missing.
  3. Package the tested logic into skills, then deploy into Claude Code for rapid runs.

Criticisms of common misinformation

  • The creator rejects the idea: “Install skills from a repository and it will personalize perfectly.”
    • Claimed false for complex B2B markets.
  • Ready-made skills may work for simpler tasks (e.g., website-based checks), but not deep multi-stakeholder context.

Course / onboarding promotion (program details)

  • Mentions a course stream starting Aug 17 (the 14th stream).
  • Promises:
    • Conceptual understanding plus a solved outreach workflow (“not just theory”).
    • ~30 minutes individual onboarding per participant.
    • Ongoing access to an updated materials library and free updates for a year.
    • Weekly group calls for troubleshooting and prompt/script fixes.
    • Access to an alumni community.

Main speakers / sources

  • Danil Gridin (speaker), associated with the Telegram channel: Den on ABM.

Original video