How I AIHow Amplitude built an internal AI tool that the whole company’s obsessed with (and how you can too)
CHAPTERS
- 0:00 – 0:36
Moda goes viral at Amplitude: an internal AI agent people can’t stop using
Wade opens with the story of how Moda spread from a few early demos to seemingly the whole company within a week. The tool is positioned as an internal AI agent that unlocks enterprise knowledge and helps generate artifacts like PRDs from company data.
- •Moda’s rapid internal adoption and “viral” spread
- •Core promise: unlock internal data + answer questions
- •Creates work artifacts (e.g., PRDs) from enterprise context
- •Sets the stage for why this matters to speed and decision-making
- 0:36 – 3:00
Claire’s challenge and show setup: build your own internal AI tool fast
Claire introduces the episode’s goal: showing how Amplitude built an internal enterprise search and product-building assistant. She challenges listeners to prototype an internal tool within a month and frames the episode around practical replication.
- •Episode focus: internal enterprise search + business Q&A + product artifacts
- •“Build your own in a month” challenge to the audience
- •Context: product leaders using AI to build better workflows
- •Expectation-setting for a live walkthrough of the tool
- 3:00 – 5:09
Build vs. buy for internal AI: move fast, don’t overengineer
Claire asks why Amplitude built instead of buying. Wade explains the tool was created in spare time over 3–4 weeks, with a philosophy of leveraging off-the-shelf components and being willing to rebuild if needed.
- •Internal build took ~3–4 weeks part-time by strong engineers
- •Split market view: some leaders buy, others advocate internal build for leverage
- •Leverage existing components (e.g., Glean APIs) to avoid heavy investment
- •Principle: build fast, iterate, and be willing to throw it away and redo
- 5:09 – 5:57
What Moda is: one interface to company knowledge and PRD creation
Wade defines Moda as an internal tool supporting Amplitude’s shift toward becoming “AI native.” Its goal is broad access to enterprise data so more people can discover insights and produce outputs like PRDs with full context.
- •Moda supports AI-native transformation by broadening access to data
- •Unlocks internal sources and synthesizes answers with citations
- •Designed to help people learn from each other’s successful usage
- •Produces artifacts like PRDs and related planning docs
- 5:57 – 9:20
Adoption by “social engineering”: build it where work already happens (Slack)
They discuss intentionally designing for adoption by meeting employees where they already work and making usage visible. Wade credits inspiration from another company’s Slack agent, emphasizing that public, reusable queries create a viral learning loop.
- •Problem: uneven AI fluency across employees
- •Slack as the universal entry point reduces friction
- •Visibility: people reuse prompts/questions and learn by observing others
- •Credibility loop: seeing trusted colleagues use it drives adoption
- 9:20 – 10:41
Slack demo: Moda introduces itself and shows company-wide usage
Wade demonstrates Moda in Slack—asking it to describe what it does and how to access it. He also shows how widespread usage is across departments, using the tool itself to understand who’s using it and what they ask.
- •Slack-based agent responds quickly and explains capabilities
- •Emphasis on citations/verification of sources
- •Usage spans PM, engineering, sales, support, marketing, execs
- •Team uses Moda to evaluate friction and improve prompts/data access
- 10:41 – 13:00
What data Moda can access: broad enterprise connectors with guardrails
They dig into the datasets behind Moda’s answers. Wade lists major internal systems and notes access is limited to enterprise-public sources, not private/restricted data.
- •Connectors include Confluence, Jira, Salesforce, Zendesk, Slack, Drive, Productboard, Zoom, GitHub, etc.
- •Can include meeting transcripts (e.g., Outreach)
- •Guardrails: no private/personal/restricted datasets
- •Goal: comprehensive context for better internal answers
- 13:00 – 19:03
PM workflow: theme mining from customer feedback to find ‘heat’
Wade shows a product-management flow: start broad with qualitative theme analysis, then narrow into a promising topic. The example explores customer desire to connect session replay with funnel analysis, backed by quotes and sources.
- •Start wide: aggregate themes across Productboard/Zendesk/transcripts
- •Narrow down to a specific high-signal request area
- •Validate with direct quotes/evidence from systems of record
- •Use opposing queries/critical checks to confirm real demand
- 19:03 – 21:06
Behind the scenes: architecture, interfaces, and the PRD orchestrator
They switch to the web UI and discuss how Moda works technically. Wade shows configuration/orchestration in GitHub (YAML/prompt-driven flows), built on an internal framework and using Glean for retrieval as part of the RAG approach.
- •Two front doors: Slack bot and a richer web UI
- •Internal framework (“Langley”) handles AI requests/orchestration
- •Glean API used for retrieval; feeds RAG context into generation
- •PRD flow breaks work into problem/solution/requirements + prototype prompts
- 21:06 – 23:24
How they learned to write good prompts and operate it as a shared product
Claire probes prompt quality and ongoing improvement. Wade explains a mix of experienced AI builders plus using AI recursively to improve prompts, and an open contribution model where engineers, PMs, and designers can all participate via GitHub.
- •Prompting skill came from both expertise and iterative learning
- •Use AI to generate/improve prompts, then edit for fit
- •Repo-based workflow: contributions via GitHub, not a closed team only
- •Cross-functional contributions (designers/PMs) and “AI week” experimentation
- 23:24 – 26:48
Live PRD creation from a single insight: from sentence to plan + prototype steps
They generate a PRD from one sentence about linking session replays to funnel steps. Moda outputs structured sections (problem, solution, requirements) and produces prototype-generation prompts that teams can paste into tools like Bolt/v0/Figma Make.
- •Input can be as small as a single-sentence insight
- •Outputs a multi-part PRD: problem exploration, solution exploration, requirements
- •Generates prototype prompts to accelerate UI exploration
- •Teams compare multiple prototype tool outputs to explore variations
- 26:48 – 29:10
Review and iteration: don’t YOLO—comment, regenerate, and apply critical thinking
Claire asks whether anyone reads AI-generated PRDs; Wade describes a review process and encourages critical evaluation. They can comment at any layer, fix upstream framing, and regenerate downstream sections to iterate quickly and safely.
- •Docs are reviewed; PRD is part of decision-making, not just a formality
- •Expect follow-up queries: cons, alternative solutions, evidence checks
- •Comment-driven workflow enables targeted regeneration of sections
- •AI speeds work but still requires human judgment and verification
- 29:10 – 32:37
Impact on velocity: compressing weeks of PM/design/eng into a single meeting
Wade explains how Moda and prototyping tools have changed product development cadence. What used to take weeks—research, PRD, design mockups, engineering translation—can now happen in a single meeting, with reviews anchored on prototypes.
- •Higher velocity: faster evidence gathering, planning, and prototyping
- •Prototypes become the focal point in product reviews; PRDs support them
- •Pace varies by project type (new product vs UI-heavy integration)
- •AI shifts bottlenecks dynamically depending on complexity and research needs
- 32:37 – 34:38
Role swapping enabled by AI: engineers, designers, and PMs cross-execute
They describe intentionally swapping roles in demos—designer coding in Cursor, engineers drafting PRD requirements, PM iterating designs. The exercise is both functional and a way to build empathy and fluency across disciplines.
- •Teams can rotate roles to prove end-to-end workflows are possible
- •AI tools let non-specialists contribute meaningfully across functions
- •Improves empathy, respect, and shared craft understanding
- •Practical team exercise: role-swap session to surface gaps and opportunities
- 34:38 – 40:26
Recap, lightning round, and practical prompting tactics
Claire recaps Moda’s creation and challenges listeners to build their own. In lightning round, Wade shares excitement about AI reducing tech-debt toil and explains his strategy for when outputs miss: trace where it went wrong, fix upstream input, and give targeted feedback.
- •Recap: internal tool built quickly, deployed in Slack + web UI, widely adopted
- •Engineering outlook: AI helps tackle tech debt and expand output per engineer
- •Prompting tactics: revert to last ‘good state’ and correct the upstream framing
- •Give specific feedback (tone, detail, audience, user voice) to steer outputs
