How we use Claude for product management at MarsBased

At MarsBased, we've been gradually integrating AI into how we work, not just in engineering, but in how we manage products and projects. This post covers the specific workflows our PMs and engineers use today with Claude, what's working well, and how it's starting to shift the role of product management at the company.

The tools and the setup

We use Linear as our task board. Linear has a solid MCP server that exposes practically everything, you can read issues, create them, update labels, change assignments, filter by project, team, status, you name it. Connecting Claude to that MCP server turns your AI assistant into a full Linear client.

Most of what we describe below works in Claude Code and OpenAI’s Codex. The underlying concept is the same: skills (or custom instructions) that encode our internal processes and templates, so the model knows how we want things done.

Skill 1: Functional definition of tasks

One of the most time-consuming parts of product management is writing well-structured user stories and functional specs. We've built a skill, essentially a prompt template loaded into Claude, that encodes our internal standard for functional definitions.

The skill instructs the model to behave like an experienced PM and write issues following a fixed structure:

• User story: As a [persona], I want to [action] so that [outcome]
• Acceptance criteria: Explicit pass/fail conditions
• Flows: Happy path, empty state, loading state, error states
• Edge cases: No data, bad input, concurrent actions, permissions issues
• Out of scope: What's explicitly excluded
• Open questions: Unresolved decisions

The model asks clarifying questions if the description is vague, drafts the issue, iterates based on feedback, and only creates the Linear issue once explicitly approved. It never creates anything without confirmation.

The result: PMs spend less time staring at a blank Linear editor and more time thinking about the actual product problem. The output is consistently structured, which also makes life easier for engineers picking up the task.

Skill 2: Project status reports

We generate client-facing and internal status reports by connecting Claude to Linear via MCP and running a report skill against it.

You give it a project name and a date range. The skill:

1. Fetches all issues from that team updated in the period
2. Filters by meaningful activity, status changes, new comments, assignment changes
3. Groups issues by project → status → last update date
4. Writes a markdown report with a short qualitative summary per issue, based on the assignee-relevant comments

The report structure maps to our internal standard for client reporting. Issues without significant activity are noted explicitly. Nothing is invented, the skill is explicit about only reporting what's visible in Linear.

This has cut the time to produce a weekly client report significantly. What used to be a manual copy-paste exercise across Linear, Notion, and email is now a few prompts.

How PMs use Claude day-to-day

Beyond skills, PMs at MarsBased use Claude for two practical things:

1. Running the project locally and asking questions

PMs install the project on their local machines and use Claude Code to ask functional or code-level questions, "what happens if a user does X without completing step Y?", "where is this validation handled?", "what does this API response look like?". This gives them independence to investigate without blocking engineers.

To be clear: we don't use AI to replace engineering at MarsBased. We believe vibe-coding by product managers is not production-ready, it's not reliable enough, not secure enough, and not scalable enough to substitute for an engineer who understands the system. The PM uses Claude to understand, not to build.

2. Managing Linear via conversation

This is where we've seen the most tangible improvement recently. PMs can create issues, update labels, change assignees, move statuses, and reorganize the board entirely through a Claude conversation connected to the Linear MCP server. No clicking through UI, no switching context.

Example: "Create an issue in the payments project for the third-party API integration, assign it to [engineer], label it as backend, and set it to In Progress", done in one message.

This works for bulk operations too. Cleaning up a sprint, relabelling a set of issues, or generating a list of everything blocked or in review, all conversational.

How the PM role is evolving

This is shifting the PM role at MarsBased in a specific direction.

Previously, our PMs carried a significant QA and project management load, tracking what moved, what got stuck, chasing status updates. A lot of that is now handled more efficiently through AI + Linear MCP.

The role is moving toward product thinking: defining what to build and why, working with clients on requirements, and maintaining the product vision. The project management layer, tracking, reporting, board hygiene, is becoming more of an AI-assisted task than a human one.

Meanwhile, the engineer working with Claude to implement a feature is naturally taking on more of the "is this working as intended?" validation loop. They're in the code and in the context, so the QA feedback cycle happens closer to the source.

We think this is the right direction. Not because PMs are less valuable, quite the opposite, but because their time is better spent on problems that require judgment, stakeholder communication, and product intuition, rather than administrative overhead.

What's next

We're continuing to refine these skills and expand the use of MCP integrations. Linear is just one example, the same approach applies to other tools we use. The pattern is consistent: encode your team's process in a skill, connect the model to your data via MCP, and let the AI handle the mechanical parts of the workflow.

If you're building something similar or want to talk through how we've set this up, we'd love to hear from you.