How I AI: Amplitude's Viral Internal AI Tool for Product Development

Amplitude built an internal AI tool that product managers use to search company knowledge, synthesize customer feedback, draft PRDs, and generate prototype instructions in the middle of a meeting. Wade Chambers says the first version of Moda came together in roughly three to four weeks of spare engineering time, then spread across the company within days after launch.

What makes Moda interesting is not just the feature list. Amplitude deliberately put the tool inside Slack so employees could see each other’s prompts, borrow workflows, and learn by example. The result is part enterprise search layer, part AI assistant, and part internal behavior change project. Moda now supports company-wide search, feedback analysis, PRD generation, and prototype handoffs into tools like Bolt, Figma, Lovable, and v0.

Amplitude decided to build around its own systems and workflows instead of stitching together separate AI apps for different teams. That gave employees one place to work with company context, but it also meant Amplitude had to own permissions, governance, retrieval rules, logging, evaluation, and ongoing support. Wade is clear that the speed of the initial build does not remove the operational burden afterward.

A product idea that once moved through weeks of research, PRDs, design mockups, and engineering review can now reach a first prototype in a single collaborative session. In this episode of How I AI, Wade explains Moda’s initial build, Slack launch, adoption, and product-development workflows.

How the first internal version came together

Wade says the project started after seeing another company demonstrate an internal Slack agent. Instead of treating it as a large platform initiative, Amplitude used a few engineers working part-time to build an early version quickly and learn by doing. The team leaned on existing APIs and enterprise search infrastructure, including Glean, instead of rebuilding every layer from scratch.

That constraint shaped the entire approach. The goal was not to create a perfect long-term architecture on day one. It was to unlock internal company knowledge fast enough that employees would actually use it. Wade repeatedly emphasizes that they were willing to throw parts away later if needed. The immediate value was getting employees closer to company context, customer evidence, and reusable workflows.

Why launching inside Slack accelerated adoption

Moda launched first inside Slack because employees already spent their day there. That decision mattered as much as the model or retrieval stack. Shared channels allowed employees to watch coworkers ask questions, inspect outputs, and copy prompts that worked.

Wade describes this as a social engineering decision as much as a product one. Someone could scroll through the channel, notice a respected sales leader or product manager using Moda heavily, and immediately understand the tool had practical value. The visibility created lightweight peer training without formal onboarding.

At the same time, the setup required careful handling of permissions and sensitive data. Public experimentation only works when the system respects access controls, keeps restricted retrieval private, and avoids exposing protected information through logs or shared outputs.

Patterns that made the Slack rollout work

• Slack interface: The first version kept interaction simple. Employees could ask questions in a familiar chat workflow instead of learning a separate enterprise application.
• Visible examples: Shared-channel usage became a live gallery of prompts, mistakes, and useful outputs. Employees borrowed prompts from one another and quickly learned what kinds of questions produced strong results.
• Familiar interaction model: Moda behaved like other Slack tools employees already understood, which lowered the barrier to experimentation and reduced workflow switching.
• Observability and iteration: The team could monitor recurring questions, weak outputs, and friction points. Wade says this helped them identify missing data sources, retrieval issues, and places where prompts needed stronger guidance.
• Fast adoption: Wade says Moda spread across the company within about a week. Product, engineering, sales, marketing, support, and leadership teams all began using it, but the company still treats usage as separate from proof of quality. Accuracy, usefulness, and failure handling still need evaluation.

Turning scattered customer feedback into usable product signals

Moda's feedback synthesis is useful because it brings evidence from several approved systems into one reviewable thread. The internal feedback-analysis workflow shows how a PM narrows from themes to source records without treating the summary as final truth.

How the feedback synthesis workflow operates

1. Data sources: Moda retrieves information from systems including Slack, Zendesk, Productboard, customer-call transcripts, Confluence, Jira, Salesforce, Google Drive, and GitHub, depending on the employee’s permissions and the company’s rules for approved access.

1. Theme generation: A product manager can ask Moda to identify recurring requests or pain points across a defined set of records. During the demo, Wade starts broadly to see where customer energy appears strongest.
2. Narrowing the investigation: After spotting a trend around connecting session replay with funnel analysis, Wade drills into that specific request to gather more detailed evidence and examples from customers.
3. Evidence inspection: Moda returns excerpts, references, and linked context so the PM can verify whether the retrieved feedback actually supports the proposed theme. Wade repeatedly checks the source material instead of treating the summary itself as final truth.

Tools shown: Moda, Glean, Zendesk, Productboard, and Slack.

The workflow dramatically reduces the time spent gathering and organizing information, especially across fragmented systems. The harder product questions still remain human ones: whether the evidence is representative, whether the opportunity is strategically important, and whether the proposed solution fits the broader product direction.

From one customer sentence to a PRD and prototype

Wade then demonstrates the workflow that will probably stand out most to product teams: turning a single customer insight into a structured PRD and prototype flow.

The starting input is intentionally minimal. The request is essentially that customers want session replays linked directly to funnel steps so they can inspect where users convert or drop off. Moda expands that short prompt into problem framing, solution exploration, requirements, and prototype instructions.

How Moda generates PRDs and prototype instructions

1. Starting request: The workflow begins with a short customer need about connecting session replay data with funnel analysis so teams can inspect drop-off and conversion moments more directly.
2. PRD generation: Moda expands the request into structured sections covering the problem, potential solutions, detailed requirements, and prototype planning. Wade describes these outputs as working drafts, not approved specifications. Product, design, engineering, legal, and security review still happen afterward.
   
   Behind the scenes, the system uses orchestrated prompts, retrieval from company knowledge sources, and multi-step generation flows. Wade shows YAML-based prompt structures inside GitHub and explains that the system can ask follow-up questions if it needs more context before generating downstream artifacts.

1. Prototype handoff: Moda generates instructions that teams can paste into prototyping tools such as Bolt, Figma, Lovable, and v0. In the example shown, the same instructions were tested across multiple tools to compare how each interpreted the product direction. Wade notes that only approved and non-sensitive context should leave the internal environment for external generation services.
2. Revision and review: Teams iterate on the generated artifacts instead of accepting them blindly. Wade says reviewers frequently challenge assumptions, ask for additional evidence, compare multiple design directions, and regenerate sections that feel weak or incomplete. Employees can comment directly on parts of the generated document and rerun downstream sections from that point.

Tools shown: Moda, Confluence, Bolt, Figma, Lovable, v0, and GitHub.

Moda does not need to create a perfect PRD to change the workflow. The single-prompt PRD workflow shows how a customer sentence becomes a structured draft and prototype brief while product, design, and engineering retain approval.

The real build-versus-buy question

Moda works because it is tightly connected to Amplitude’s internal data, workflows, and employee habits. Another company could reproduce parts of the experience with off-the-shelf tools, but the moment you centralize enterprise search and AI generation internally, you inherit responsibility for permissions, retrieval quality, governance, maintenance, support, and evaluation.

Wade is notably pragmatic about this tradeoff. The initial prototype was fast, but sustaining an internal AI system is operational work.

The reusable pattern is not necessarily the exact stack Amplitude built. It is the combination of shared company context, permission-aware retrieval, and workflows designed around specific employee tasks.

The strongest outputs in this episode come from systems that can reference real internal evidence instead of generic model knowledge. Faster drafting only matters if employees can inspect the supporting sources, challenge weak reasoning, and improve the prompts and retrieval logic over time.

The part worth copying is the narrow, workflow-specific approach. Moda succeeds because it helps employees complete concrete tasks inside tools they already use, not because it tries to become a universal AI layer overnight.

Start with one painful internal workflow, define the approved data sources, make the outputs inspectable, and keep humans responsible for the final decision. This works especially well for research synthesis, early-stage product exploration, and fast collaborative prototyping. It still requires human judgment anywhere evidence quality, customer impact, security, or product direction matter.