The Claude GTM Agency Playbook

A coding agent you run locally from your PC, installed in two lines of script.

Step 1: install

Download and run the install script. Two lines. Done.

Step 2: enable Plan Mode first

Before building anything, switch to plan mode and describe what you want. Claude Code gives you 2–3 approaches. You pick one, and it executes without you touching anything.

If you don't know what to build or how, start here. Plan mode removes the guesswork.

Step 3: build your CLAUDE.md

This is the master knowledge file. It stores everything Claude Code knows about your business, your processes, and your rules.

Update it constantly. Every new workflow you build, log it here so future builds reference it automatically.

Step 4: invest 1 hour per day

Minimum one hour per day per person building in Claude Code. The ROI compounds fast — within a few hours of starting you're already above 80% of users.

The core GTM use cases

Custom scrapers. Tell Claude Code: go to this website, click each exhibitor profile, extract email, website, phone, address. It builds the scraper from zero and you get a CSV while you do something else. This previously required hiring Python developers from Fiverr.

Scraping behind login walls. Some databases (e.g. Shoptalk attendee lists) are only visible when logged in. Claude Code told us exactly which tools to install — in this case Cookie Editor — to capture the authenticated session and pass it to the scraper. Result: 400 filtered attendees scraped automatically.

Full process automation. Onboarding, infrastructure setup, campaign building, and reporting can each be built as an automated process. The agent runs without you monitoring it and delivers the output when done.

From the moment a client joins your Slack channel to a completed onboarding form — including deep research, curated questions, and a golden ICP document — the entire process runs without manual work.

Step 1: trigger on Slack join

Client joins Slack. Automation fires. No manual trigger needed.

Step 2: create client folder on GitHub

Claude Code creates a dedicated client folder in your GitHub repo. All files, workflows, and outputs for this client live here.

Step 3: run 3 research agents in parallel

Three agents run simultaneously:

• Industry analysis
• Competitor landscape
• Market and future predictions

Also feed in transcripts from pre-sales calls. These contain details the client already gave you that should feed into the onboarding form.

Step 4: generate the onboarding form

Claude transforms the research into curated questions specific to this client — not generic ICP questions, but questions shaped by what the research already surfaced.

It covers: responsible contact, sending infrastructure, DNC list, offers, ICPs per decision maker, pain points per role, buying signals, and what triggers inbound for them.

Useful side effect: clients often say filling this in forced them to rethink their own business. That's strong positioning for your agency.

Step 5: push to Typeform and notify the client

The form goes automatically to Typeform. A Slack bot posts the link to the client with a message to go and fill it. Zero manual steps.

Step 6: after submission, generate outputs

Once the client fills the form:

• Golden ICP document — one source of truth, used by every subsequent agent
• Full GTM strategy generated
• Initial campaigns and offers drafted, 80–90% done

Manual infrastructure setup takes 40 minutes per client. With Claude Code plus the Salesforge API, it takes 5. The code runs, confirms each step, and you watch accounts appear in the platform without touching anything.

Inputs

Fed straight from the onboarding form:

• Sending account names
• Email signatures
• Profile pictures

What it builds automatically

• Creates all sending accounts inside the platform
• Pushes accounts automatically to the warm-up tool
• Inboxes start warming without any manual steps

The critical dependency

This only works if your sending tool has solid API documentation. If the documentation is missing steps or unclear, Claude Code hits a wall and stops.

When evaluating any tool for your stack, treat API quality as non-negotiable. Poor documentation is a bottleneck for everything you try to automate.

The offer bible

Build a database of your best-performing campaigns and offers. This is the reference file Claude Code pulls from when generating new campaigns.

The more campaigns you add over time, the better the output quality gets. Every client adds to the bible.

The campaign generation flow

After the onboarding form is filled, Claude Code reads the golden ICP document, the offer bible, and the research from onboarding.

Output: 80–90% complete campaign copy across all angles and decision makers. Offers are well-matched because the model has reference examples.

The target is to push campaigns live directly from Claude Code within days of onboarding, including:

• All omni-channel steps
• Leads with spintax and variables already applied
• Sequences loaded into the sending tool automatically

Your team's job becomes quality control and creative variation, not building from scratch.

The automatic feedback loop

Connect your sending infrastructure API to Claude Code and pull campaign data across all decision makers at message level:

• Reply rates per message
• How many replies converted to interest or booked meetings

This data feeds automatically back into future campaign generation. Every batch of campaigns makes the next batch better.

Automated reporting

Pull live stats via API and deliver automatic data reports to clients covering campaigns live, messages sent, interest rate, and meetings booked.

This adds predictability. "We sent 1,000 emails, got 10 positive replies, booked 4 calls" means the client can project what doubling volume produces. Transparency builds trust.

The LinkedIn vs email reply rate gap

Same message, same prospect, same value proposition — sent via email, then LinkedIn:

Channel	Reply rate
Email	2–3%
LinkedIn	20–30%

LinkedIn outperforms email by roughly 10x on reply rate. The implication: for any account with a limited TAM, LinkedIn touches aren't optional. They're the primary driver.

When to run email only

Use email-only when your TAM is large — tens of thousands of decision makers in one category — and you need fast volume and quick results. Email scales faster.

When to run multi-channel

For campaigns targeting 1,000–2,000 prospects where you want meaningful test data, run multi-channel. The results aren't close: multi-channel outperforms both email-only and LinkedIn-only by a significant margin across all tested clients.

How to structure multi-channel sequences

Treat every touchpoint as a first touch. Don't assume the prospect saw your email before they got the LinkedIn message. Always include the core value proposition in every single message on every channel.

Don't water down channel two because "they already saw channel one." They probably didn't.

Adding cold calling

Cold calling top prospects on top of email and LinkedIn lifts overall results further.

US market pick-up rate is roughly 3% — low. But when you do get someone on the line, meeting conversion is significantly higher because you handle objections live. Target cold calling at your top-tier prospect segments only, not the whole list.

The allbound model

For entering new markets seriously:

• Email to the full ICP list
• LinkedIn messages to the same list
• Cold calls to top-tier prospects

This isn't a future ambition. Build toward it now, because it's already what separates agencies getting results from those that aren't.

Event outreach play

For each client, identify the top events in their industry, then build a workflow that:

1. Finds the attendee list for the event
2. Filters by the right job titles
3. Sends a message suggesting a coffee or meeting during the event

This works independent of client because the intent signal is universal — people go to events to meet people and do business.

Add social listening on top. Find everyone who posts on LinkedIn that they're attending the event, and trigger a message off that mention: "Saw you're going to [Event] — me too. Worth a quick chat about [problem]?"

Personalised without any manual research. The signal does the work.

Competitor followers play

1. Identify the top 5 competitors for the client
2. Find their LinkedIn audience
3. Filter to the target job titles only
4. Scrape that filtered list
5. Run research to find why your client is a better option than the competitor they currently use
6. Generate displacement copy highlighting that comparison
7. Launch outreach

The message is automatically personalised to a competitor-specific pain point, not a generic pitch.

Both plays can be fully automated

Event outreach and competitor displacement can both be built as automated workflows in Claude Code. Once built, they run for any client with minimal configuration changes — the playbook scales across your whole book of business.

Host a webinar on a topic your ICP cares about. Instead of posting about it on LinkedIn and hoping people find it, send the registration link via outbound to 2,000 targeted prospects.

I'm hosting a webinar on [topic]. I think you'd find it valuable. Here's the link if you want to join.

No hard sell. Just an invitation.

Why it converts

If someone attends and watches 30–40% or more of the webinar, conversion probability to sale or sign-up is high. They self-selected as interested — you didn't have to push them. You let the content do the qualification work.

How to execute without burning deliverability

Do this on LinkedIn, not email. Use multiple team accounts: five people each sending to hundreds of prospects weekly gives you significant reach without touching email infrastructure.

No landing page required. Just the LinkedIn invite from a real person.

What changed

Old system: credits charged only for enrichments — finding emails, phone numbers, web research.

New system, from March 11th: two separate credit types.

• Data credits — emails, phone numbers, data lookups
• Action credits — every single action performed inside the platform, including third-party API calls you make with your own keys

If you bring your own API key for a third-party tool and use it inside Clay, that still costs action credits. You pay the third-party tool and you pay Clay.

Who gets hit hardest

Agencies running a single shared workspace for multiple clients. The old model — one Pro plan with 50k credits, running 10–20 clients from one workspace — is now unviable.

40k action credits in the new Pro plan is roughly one well-built Clay table. If you need 10 tables for 10 clients, you're looking at 10x the cost.

Data costs went down 50–90%. But enrichment, scoring, research, and third-party tool usage went up significantly.

What to do now

Keep legacy pricing if you have it. Existing users aren't forced to switch. Stay on legacy as long as possible and don't voluntarily migrate.

Have clients cover the Clay license. Pitch it on the sales call: the client pays for their own license as part of the engagement. Fair framing — if they ever leave, they keep the license and can use it internally. Most serious GTM clients accept this.

Diversify to Bitcale and Databar. Both have credit systems more favourable to agencies. Worth testing for use cases getting expensive in Clay.

Move some Clay functions into Claude Code. Google Maps scraping and website enrichment using your own OpenAI API key can move out of Clay entirely. One example: scraping 75,000 companies from Google Maps cost zero Clay credits when done inside Claude Code. The same enrichment in Clay under the new model would burn through action credits in days.

A free, open-source autonomous AI agent. Install from GitHub, runs locally. It gets access to CRM, calendar, Slack, and web, and executes repetitive tasks without you watching it.

The best GTM use case: automated prospecting

• Give the bot a company-level list
• Specify target job titles
• The bot searches web databases and tools like Prospeo to find the right decision makers and their contact details
• It delivers a complete prospect list with verified data

For a team of 70 cold callers needing fresh lists daily, this bot delivered new prospect lists in roughly 10 minutes, triggered directly from Slack.

How to run it safely

Use a separate device. Don't run OpenClaw on your main work laptop — use a dedicated machine such as a Mac mini. It gets access to everything you give it, so if that machine holds sensitive client data, contracts, or financial info, that data is at risk.

Limit what it can access. Install only the apps it needs on that device. Give it access only to data you're comfortable with it using. Don't connect it to anything you wouldn't want leaked or used externally.

GDPR and data legislation. Any data it processes must comply with applicable legislation. Treat this the same way you would any automated data processing tool.

Watch credit spend. OpenClaw can burn through API credits fast if left unchecked. Set limits and monitor usage.

The core advantages clients are buying

Faster execution. Onboarding to live campaigns in days, not weeks. Infrastructure setup in 5 minutes, not 40. First campaign drafts generated automatically from the onboarding form. Time to value is radically shorter.

Systematic process. For automation to work, every process must be mapped. AI-native agencies have been forced to think systematically about every stage of GTM, so the client gets a repeatable system rather than a person doing ad-hoc work.

Faster feedback loops. More campaigns live in less time means more data faster. The agency can test more angles, value propositions, and decision makers simultaneously, and use that data to improve. Manual agencies test one thing at a time.

The one risk to manage

If AI generates all the copy, every agency's output starts to look the same. The risk is a reinforcing loop where AI trains on AI-generated content and everything converges.

How to counter it: human review on every piece of copy. Add creative variation. Make offers specific to the client's actual results and language. AI gets you 80–90% of the way — humans add the 10% that makes it sound real.

What the GTM engineer role is becoming

Not manual task execution. Not tool operation. Instead: orchestrator of AI agents.

The skill is knowing what to automate, in what order, and how to connect the agents into one system. The output is a productised GTM fulfilment machine that scales without proportional headcount growth.