GTM Engineering With Claude Code: Build an AI SDR Workflow

By Ammar Ahamed, Founder, Cocoon · 3 November 2026 · Facts checked against treg.to, its docs and its GitHub repository on 3 November 2026

What is GTM engineering?

Every sales and marketing team has a playbook: who we sell to, how we find them, what we say first, and when. GTM (go-to-market) engineering is the practice of turning that playbook into systems instead of weekly manual effort. A list that rebuilds itself. Enrichment that fills in company size and industry before a rep opens the lead. A signal feed that says which accounts are hiring for the role you sell to. A first line of research for every email.

The job title took off because the tools changed. A few years ago this meant stitching together spreadsheets, a data vendor and a sequencing tool. In 2026, an agent like Claude Code can run the whole chain from a written prompt, as long as it can reach the data. That last part is where most teams get stuck, and where a data layer like treg comes in.

What does a GTM engineer do?

A GTM engineer owns the automated parts of the revenue system. In a typical week that means:

1. Writing the ideal customer profile (ICP) as fields a data provider can filter on, based on the deals the team actually won and kept.
2. Building and enriching lists cheaply: checking the fields you already have first, and paying for people, emails and verification only on rows that pass.
3. Watching for timing: hiring, funding, job changes and posts, each with a link and a date.
4. Handing reps a reason to write, with its source, while a person stays on the send button.
5. Keeping the data honest: verifying before sending and re-checking people who just changed jobs.
6. Measuring: tagging every row so pipeline can be traced back to a list, a signal and a data provider.

The title overlaps with others, so here's a simple way to split the work:

Adapted from treg's GTM engineering playbook.

Roughly: RevOps decides how the system should work, and the GTM engineer builds the automated parts of it. In a small company, one person often does both, and the "GTM engineer" may just be a marketer who is comfortable with an agent, a spreadsheet and a command line.

The GTM engineer's stack in 2026

• An agent to run the work: Claude Code, Codex or Cursor.
• A data layer: company and people search, enrichment, email finding and verification, hiring, funding, news and social data. treg is one option, with one key, per-call pricing and your own vendor keys used first.
• Skills: written recipes the agent follows, such as treg's lead-signals.
• A sending tool for email sequences, with deliverability managed properly. treg doesn't do this part.
• A CRM where every row and outcome lands.

Without a shared data layer, each of those lookups is a separate seat. treg's homepage lists Apollo at $59 per seat, Hunter at $34 a month and Crunchbase at $99 a month for the jobs its catalog serves per call (for example, email find and verify from $0.0005 a call, funding data at $0.01).

What an "AI SDR" should and shouldn't do

"AI SDR" is a crowded search term (about 4,400 searches a month worldwide in our keyword pull) and an even more crowded product category. Many promise an agent that finds leads, writes emails and sends them on its own. The evidence for letting it send is thin. When treg scored the first lines its agent researched from news events, 4 of 19 were rated decent or better on a 0 to 3 scale. treg's own conclusion: let the agent research, and let a person write, or at least approve.

A safer split, and the one this guide builds:

• The agent does: list building, enrichment, qualification, signal checks, verification, research notes with sources, draft first lines.
• A person does: decides the ICP and the rules, reads every message, presses send, handles replies.

Build it: an 8-step AI SDR workflow in Claude Code

Set up treg once (paste set up treg - https://treg.to/llms.txt into Claude Code; details in our setup guide). Then work through the steps below. Each has a prompt to adapt and the rule worth keeping.

1. Write the ICP as filters, and count the market for free

Split your ICP into filters (country, headcount, industry, funding, technology) and checks (judgement calls per row). Then count before you buy. In treg's run for US B2B software companies with 51 to 200 staff, one provider's free count went from 10,403 companies to 1,357 when it added "funded 2024 or later", and to 25 with "Series A" instead. Each filter changes the market by an order of magnitude, and counting is free.

Rule: don't buy a list until two free counts roughly agree and ten sample rows look right.

2. Qualify on the fields you have before any paid lookup

In treg's recorded run, 50 companies became 48 with a usable domain; 27 passed the ICP check and 21 were dropped before any paid lookup. The run ended with 20 verified-deliverable leads for $2.33 in total, or $0.12 per deliverable lead. Enriching all 48 up front would have cost an estimated $4.12.

Rule: no expensive lookup (people, emails, news) runs on a row that hasn't passed the check.

3. Enrich a company from its domain

When a row is too thin to judge, one cheap enrichment comes first. We ran treg's routed companies.enrich tool on a single domain: it skipped eight providers that needed a LinkedIn URL, a name or an email we didn't have, and the first provider that could answer returned the company name, industry, headcount and LinkedIn page for $0.0005. treg's own test on 20 fresh domains took a median of 2.2 seconds and $0.053 in total.

treg call treg.companies.enrich --method POST --data '{"domain":"acme.com"}'

The same call is how you route inbound leads: enrich the email domain the moment a form is submitted, so the lead reaches the right rep before anyone opens it.

4. Find people by function, not "decision makers"

A "decision makers" endpoint sounds perfect and often isn't. Across 8 accounts (3 were rate-limited), such an endpoint returned 45 senior people in treg's run; none had "marketing" in their title, even though marketing was the buyer. A search by title ("marketing") returned 20 people across all 8 accounts.

5. Find emails with a waterfall, then verify them

A waterfall tries email providers one after another, cheapest first, until one finds an address. treg's routed finder works this way and doesn't bill misses on per-success providers. In treg's published benchmark on 292 people, the cost per correct work email was $0.0056 through treg against $0.0395 through Clay, with exact-match accuracy around 90% for both; that's treg's own test, and worth repeating on your own rows.

Found isn't the same as safe to send. In another treg study, the finder returned an address for all 60 people, and the verifier rated 32 of them deliverable (53%), 16 risky, 4 unknown and 8 invalid.

Rule: only verified-deliverable addresses go into the main sequence. Keep "unknown" apart, and treat a blank catch-all field as unknown, not safe.

6. Use signals you can open and date

Timing matters as much as fit. Good signals are ones a person can click and check: a job post with a date, a funding announcement, a public post about a new role. On 27 accounts, treg found 16 with sales, marketing or growth roles opened in the last 60 days, and also found that 355 of 675 job postings returned were already closed, which is why the date matters.

Job changes need extra care. In treg's study of 148 people who had publicly announced a new role in the past month, a typical contact database didn't show the new employer for 68% of them. A live profile read was far more current, though it found only about a third of the people.

Rule: no source link, no signal. No "why now", no outreach. And test a signal on your own wins and losses before you give it weight: in treg's study, hiring activity didn't separate startups that went on to raise from those that didn't.

7. Research the first line; a person writes it

The last sentence matters. An agent asked for a reason will produce one, real or not. Asking it to say "none found" keeps the notes honest.

8. Tag every row, and measure

Store the source list, the signal, the provider that found the email and the verifier's verdict on every row. Once a month, join it with outcomes in the CRM. Then you can answer the questions that decide next quarter's budget: which list, which signal and which provider produced meetings, and what each deliverable lead cost.

treg skills for sales and recruiting

treg packages two of these workflows as skills your agent can follow end to end:

• lead-signals finds in-market buyers from public signals (complaints, job changes, hiring, funding, tool adoption) and can monitor a topic, a competitor or a list of accounts.
• recruit turns a LinkedIn job post into a ranked shortlist: criteria from the post, a people search, a reason per candidate with cited evidence, and a checked contact, under a per-candidate spend cap.

npx skills add superdesigndev/treg --skill lead-signals
npx skills add superdesigndev/treg --skill recruit

In Claude Code you can also install each as its own plugin, for example /plugin install recruit@treg. If you're new to skills in general, they're just written instructions with the tools attached; our Claude guide covers the basics.

Clay, a stack of seats, or an agent?

These suit different teams:

• Keep Clay if your team wants a shared, visual table that non-technical people edit together.
• Keep your Apollo or similar seat if reps search and sequence inside it all day. You can connect that key to treg and your agent will use it, unmetered.
• Work from an agent with a per-call data layer if one person runs the workflow and wants it written down as prompts and rules that can be rerun, reviewed and versioned like code.

Most teams end up with a mix. The useful question isn't which tool, it's whether your workflow is written down well enough that any of them could run it. If it isn't, start by mapping it: our free Workflow Mapper turns a messy process into steps you can automate.

Rules before you send a single email

• Business contacts are still personal data. A named person's work email is personal data under GDPR in Europe and the UK, and under Sri Lanka's Personal Data Protection Act, whose main duties apply from 1 January 2027. Know your lawful basis, honour opt-outs and keep records. Our PDPA guide is a good starting point; for your own case, ask a lawyer.
• A person approves every send. Run new automations in shadow first, then on a small segment, and only then at volume.
• Verify before sending. Bounces damage your sending domain for everyone on your team.
• Cap the spend. Give the agent its own treg token with a cap, and have it state the expected total before any loop.
• Count your own market first. Data coverage varies a lot by country and provider. If you sell in Sri Lanka, Singapore or the Gulf, run the free counts for your market before trusting numbers from a US-heavy run like the ones above.

Learning GTM engineering

The tools in this guide are the easy part. The hard part is the judgement: writing an ICP a machine can follow, knowing when a signal is noise, and knowing when a message is good enough to send. If you want to try the outbound side without setting anything up, Cocoon's free Prospect Finder and Cold Email Writer walk through the same steps in a browser. If your next job is organic search rather than outbound, the same agent setup powers Claude Code for SEO.

Frequently asked questions

Sources

treg

• The GTM engineering playbook (2026), treg
• Work email finding benchmark, treg
• Find and verify a lead list (recorded run), treg
• Lead signals skill, treg
• treg.to homepage