How to prompt Claude for GTM: 6 rules that actually work

How to prompt Claude for GTM work, not just GTM copy

Most people prompt Claude for GTM the way they'd prompt any chatbot: one task, one message, one output. Write me this email. Summarize this call. That works, but it treats Claude like a faster typist, not like something that can actually run a chunk of your go-to-market motion. The gap between those two uses isn't the model. It's the prompt.

Claude Code and Claude's agentic features (subagents, skills, memory, tool use over MCP) are built to plan, delegate, and act across multiple steps without you babysitting each one. But none of that shows up automatically. You have to prompt for it. Below are six rules that make the difference between "Claude wrote me an email" and "Claude ran the workflow," each with the specific line you can reuse.

Rule 1: give it the goal, not the steps

The default instinct is to hand Claude a task already broken into steps: "search for accounts, then check funding, then draft an email." That caps Claude at exactly the steps you listed, and it can't adapt when step two turns up something you didn't anticipate.

Give it the outcome instead. "I need to book meetings with Tier 1 accounts showing buying signals. Here's my ICP." Let Claude figure out which accounts, which signal, which channel. This matters more in GTM than most domains because the right next step genuinely depends on what the last one found, an account with a recent funding round gets a different approach than one with a new VP of Sales.

Rule 2: stop it from overplanning

Left alone, an agent will sometimes hedge: it lays out four options, weighs tradeoffs, and asks you to pick one instead of just doing the work. That's useful when a decision genuinely needs your judgment. It's wasted motion when you just wanted the task done.

One line fixes most of it: "When you have enough information to act, act. Give me a recommendation, not a survey." This doesn't mean Claude stops checking with you on real judgment calls, changing a review score, touching production data, sending an email on your behalf are all still worth a pause. It means Claude stops asking permission for reversible, low-stakes steps it's already equipped to take.

Rule 3: let it delegate to subagents

Claude Code can spin up subagents, separate instances that work on a piece of a task in parallel and report back. For GTM work this maps naturally onto how a real pipeline runs: one subagent researches an account, another checks it against your ICP, another drafts the outreach, all at once instead of in sequence.

The prompt that unlocks this is explicit: "Split this across subagents, run them in parallel, then hand me the review." Without that instruction, Claude will often do the whole thing serially in one thread, which works but is slower and gives you one long transcript to audit instead of three focused ones.

Rule 4: use skills, not prompts

If you've explained the same workflow to Claude more than twice, that's a sign it should be a reusable skill or slash command, not something you retype from scratch every session. Claude Code supports custom skills that fire from a keyword or command across any conversation, and Claude.ai supports Projects with saved instructions that persist the same way.

The test is simple: anything you've explained twice should be saved once. A GTM team running signal-based outbound might save a /lead-sourcing skill, a /meeting-prep skill, and a /signal-strategy skill, each one encoding the exact steps and judgment calls specific to that workflow, so invoking it takes one word instead of a paragraph of context every time.

Rule 5: give it memory

Starting every conversation from zero means re-explaining your ICP, your scoring rules, and your voice every single time, which is both wasted effort and a source of drift, since the explanation you give on a Tuesday afternoon is never quite identical to the one you gave Monday morning.

Put that context in a Project (Claude.ai) or a persistent memory file (Claude Code) once: your ICP, your tone, your qualification criteria. "Use everything in this Project: my ICP, my voice, my scoring rules." Claude gets more consistent over time instead of restarting cold, and it stops making the same clarifying mistakes it made last week.

Rule 6: audit before reporting

Agents can report work they didn't actually verify, a summary that says "done" when the underlying step silently failed or produced something different than claimed. This is the rule most worth keeping even if you drop everything else, because it's the difference between trusting an agent's output and having to check it yourself anyway.

One line: "Before reporting, check every claim against a real result." Concretely, that means Claude re-reads the file it edited instead of assuming the edit landed, re-runs the query instead of assuming the row count, and opens the sent email instead of assuming the draft went out clean. Skipping this step is how "the campaign is live" turns into a campaign that never actually sent.

Putting it together

None of these six rules require a different model or a paid add-on. They're prompting discipline, the same kind of discipline that separates a rep who runs a tight process from one who improvises every call. Write the goal, not the steps. Let Claude act instead of surveying you. Delegate to subagents when work is genuinely parallel. Save anything you've explained twice as a skill. Give it your ICP and voice once, in memory, instead of every session. And always ask it to check its own work before it tells you the work is done.

The tools underneath still matter. None of this compounds if Claude can't actually reach your CRM, your enrichment data, or your sequencer, which is what MCP is for. A CRM like Attio or a research tool like Clay with a real MCP server is what turns "let it delegate" from a nice idea into subagents that actually read and write your pipeline. If your stack isn't wired up yet, start with the AI-native sales stack guide to get the layers connected, then read seven real GTM workflows run through Claude Code to see these prompting rules applied to actual pipeline work. And check the AI-stack-fit leaderboard before you commit to a tool that Claude can't actually drive.