An AI product manager that knows where to stop. Built by RampStack.
Krine is an AI product manager built by RampStack. It does the convergent product work, research, specs, roadmaps, and experiment analysis, on its own, and stops at the decisions that need human judgment, surfacing them as clear choices rather than making them. It is built on RampStack's open bimodal framework.
The operated product lives at rampstack.co/krine. This repository documents the framework Krine runs on and shows worked examples of how it behaves.
Most attempts at an autonomous AI product manager fail the same way. They push the agent to decide everything, so it produces confident output exactly where confidence is not warranted: a clean spec, and in the same breath a positioning call it had no business making alone.
Krine treats product work as two different kinds of task and refuses to blur them.
Convergent work has a knowable-correct answer and a known process to reach it. Writing a spec from a decided feature. Computing a sample size. Ranking a backlog against stated inputs. Two competent practitioners given the same inputs produce substantially the same output. Krine runs this work.
Divergent work has no knowable-correct answer, only judgment, taste, and stakes. Which positioning to pursue. Whether a creative direction lands. Whether to ship or kill. Two competent practitioners reasonably disagree. Krine stops here and hands the decision back, already researched and framed as a clear choice.
The skill is knowing which is which, and stopping at the second kind instead of faking it. That restraint is the point. An agent earns trust by being honest about its boundary, and the boundary is where the product work meets the decisions only a person should own.
The full framework, including the classifier, the overrides that force a stop, the prioritization gates, and the data-or-silence rule, is written up in the whitepaper: rampstack.co/whitepaper.
Five surfaces of product work, each with the point where it stops.
Conversion testing. Reads a funnel, generates ranked, evidence-grounded test ideas, and stops at which hypothesis to run. When the data is too thin to support a trustworthy test, it declines to recommend one and says why, rather than dressing up noise as a result.
Feature definition. Turns a decided feature into a complete spec: proposal, user stories, acceptance criteria, out-of-scope, dependencies. It does not guess effort, and it will not write a spec around a missing answer; if the inputs that justify the feature are absent, it stops and asks for them.
Roadmap planning. Takes a set strategy and a backlog, ranks within each theme, maps dependencies, fits the plan to capacity, and produces a sequenced plan with an explicit Not-now list. It does not set the strategy, and it surfaces the plan as a recommendation rather than committing it.
Readiness check. Before any run, it confirms there is a defined surface, a set goal, and enough data to produce a meaningful result. When something is missing, it declines to start and names the condition that would clear it.
Experiment evaluation. Defines an experiment and stops so a human deploys it, then evaluates the measured result and stops so a human ships or kills it. It never calls a winner before the result is statistically significant. A large lift on a small sample is reported as still accruing, never as a win.
The through-line: Krine does the work, and the honest stop is where it hands you the decision.
This repository is the documentation for the Krine framework and a set of worked examples. It is not the engine itself; the engine is operated by RampStack and runs the product at rampstack.co/krine.
What you will find here:
- The framework, documented. The convergent and divergent distinction, the classifier that decides which is which, and the stops that protect the decisions a person should own.
- Worked examples. White-labeled, copyable prompts showing how to use each capability, what it returns, and where it stops. See
examples/.
What you will not find here: the operated engine's internals. The framework is open and documented; the running system is the product.
Each example is a template you can read and adapt: the prompt, what comes back, and the honest stop.
Every example pairs a successful run with the case where the workflow declines, because the decline is the part that is hard to find anywhere else.
Krine is built by RampStack, which also maintains the public claude-skills catalog. The thinking behind Krine is published openly in the bimodal product manager whitepaper.
- Product: rampstack.co/krine
- Whitepaper: rampstack.co/whitepaper
- RampStack: rampstack.co
Krine is the name of RampStack's AI product manager, in active use since 2026. The name and its associated branding are used in commerce by RampStack. This documentation, dated and public, is part of that record.