Keynote takeaway · Berlin, 21 October 2026

Our AI brain, file by file

This is the setup I showed on stage: the plain markdown files my product team and I work from every day at Leadfeeder. Copy the structure, put your own content in it, and start with one file today.

Mansi Acharya Sheth VP of Product, Leadfeeder
The file tree

Everything lives in one folder

We stayed tool agnostic. Each folder is one layer of the brain, and each file has a person who keeps it current.

  • ai_brain/plain markdown, in git
    • context/what it knows
      • me.mdmy voice, tone, principles
      • stakeholders/*.mdC level, peers, my team
      • strategy/company goals, key decisions
      • product/vision, strategy owner: me
      • competitors/one file per rival
      • discovery/per area: calls, tickets, research
    • connections/how it reaches
      • mcps/Slack, Linear, Figma, Amplitude, Canny and more
    • skills/what it can do
      • prd_writer/drafts, then the expert lens
      • product_brief/
      • stakeholder_update/
      • competitive_card/
      • design_system/components PMs design with
    • memory/what it keeps
      • rules.mdsaved only when I ask
    • automations/what it does on its own
      • competitor_releasesweekly
      • feedback_digestweekly, per PM
      • slipping_prdsweekly, for me to act on
      • meeting_actionsdaily, before the day ends
Plain markdownVersioned in gitWorks with any AI toolEvery file has an owner
Bonus

Understand it before you spec it

The prompt I run when we open a new problem area. It puts the moves from my talk to work: it reads your context and discovery files, asks for a baseline, and ends with one metric that defines success. Change the steps to fit how your team works.

Prompt

You have my AI brain: context, connections and discovery files (calls, tickets, research, usage). I'm opening a new problem area. Help me understand it like a product leader, don't just summarise. Work in order; ask before you assume.

  1. The problem: across the signals, what keeps failing, for whom, in which workflow, and why now? State it without a solution. If I drift into a fix, pull me back.
  2. The evidence: weigh it, don't list it. Label each signal Strong, Medium or Anecdotal, give me at least one hard number, and name the gaps honestly. Don't invent.
  3. Who it hurts: the specific persona, segment and plan; who feels it vs who'd configure the fix, and whether it's a user, team or account problem.
  4. The size: how big it is, directionally, on at least two of ARR, activation, retention, efficiency and risk, and how you got there.
  5. The hypothesis: one testable, falsifiable belief, written as “We believe [problem] causes [behaviour]; fix it and we expect [measurable outcome], because [reason].” Never a feature.
  6. The bet: one leading signal (30 to 60 days) and one lagging one (60 to 90 days), each with a baseline, or say “no baseline yet.” Name the single metric that defines success.

Give it back as: the problem in one line · who it hurts · evidence, confidence and gaps · the size · one hypothesis · the primary metric. Free of solutions, tight, in my voice. Where the evidence and I disagree, flag it once, then let me decide.