Context plane
Approved behavioural, commercial, customer and repo sources become one persistent, source-backed understanding of the product and its users.
The product
Product intelligence that follows the action.
Sightspool joins the assumptions, evidence, expert judgement, product actions and outcomes that normally fragment across analytics, research repositories, meetings and somebody’s memory.
The architecture of the function
Five layers. One continuous learning loop.
“Decision OS” is the governance spine inside Sightspool. The broader product also understands context, validates signals, routes specialist work and accumulates what the team learns.
Signal layer
Sightspool turns beliefs into questions and validates the event, threshold, sample and evidence required before treating a number as a read.
Investigation
One agent reads the question through research, interaction-design and service-design lenses from the same context, instead of producing disconnected AI opinions.
Action governance
Recommendations keep their evidence, uncertainty, downside and client-owned decision attached. Consequential actions remain human-gated.
Learning memory
Product changes re-arm the right questions. Outcomes, including null and negative results, improve the next piece of work.
Live now
This is working product, not a proposed org chart.
Sightspool already runs as a multi-tenant application with connected evidence, one governed agent, shared memory, human-gated actions and a coding-agent surface.
Connected evidence
PostHog, Stripe, GitHub/repo context, first-party Signals and approved customer material enter through source-specific tools with provenance.
Truthful measurement
Unknown events, measured zero and insufficient samples remain distinct. An unavailable signal cannot silently become a product verdict.
Persistent context
The agent carries sources, memories, tasks, personas and journeys across runs rather than beginning every session from a blank prompt.
Governed action
Findings require proof. Interventions require approval. Verdict history is append-only, and an agent can never declare an assumption verified.
Product change awareness
Repository changes can re-arm the assumptions that depend on them, so yesterday’s answer does not survive a materially different product.
Outcome continuity
The action, intended result and later observation remain one record, with digests and coding-agent write-back keeping the learning in circulation.
Evidence methods
The question decides the method, not the tool you bought.
Most tools own exactly one method and bend every question toward it. Sightspool asks what could actually settle this question, gets that, and refuses to claim more than it got.
Measured reads
A question with a runnable signal is measured against your own analytics — an event funnel, a friction rate, a resurrection or scroll-depth read — with the sample size stated and the threshold set in advance.
Senior expert reads
When there is not enough traffic to measure anything, a routed lens gives a qualitative read, and the hard calls go to a senior human. This is evidence, never a measured verdict — and it is graded later against what the numbers eventually say.
Interviews and surveys
When behaviour cannot explain itself, an approved research brief recruits from the relevant cohort and runs a micro-interview in your product, or a survey through your own PostHog. A human designs the study; the agent frames it.
Deep-capture studies
A specific written question can justify bounded, cohort-scoped session recording in your own PostHog — time-boxed, sample-limited, and torn down automatically. Never blanket capture, and never without an assumption behind it.
Demand probes
Test whether anyone wants a thing before building it. The probe is visibly unbuilt, the click is counted as a vote, and everyone who clicks is told immediately that it does not exist yet. That disclosure is not configurable.
Repo and commercial context
What shipped, what it touched, what it costs and who churned. Behaviour becomes far more useful once it can be related to the change that caused it and the revenue standing behind it.
Three surfaces
It meets the team where the work already happens.
The same governed register is reachable from the app, from the coding agent your engineers already have open, and from your product itself.
The web app
- The register, evidence feed and decisions
- Ask the agent from any page
- Approve actions and resolve escalations
- A weekly digest of what changed
Your coding agent
- Ask what is assumed about the file you are editing
- Get the honest state, never an inflated verdict
- Record assumptions as the code is written
- Settled verdicts written back into your context file
Your product
- Capture intent and effort at the moment of friction
- Run approved interviews and probes in-product
- Optional — Sightspool works without it
- Open source, Apache-2.0, provenance-signed
Clear product boundaries
Continuous UX capacity without pretending to be everything.
Sightspool operates customer-facing product uncertainty. It does not quietly expand into roadmap authority, production delivery or formal compliance work.
Sightspool owns
- Source-backed customer and experience context
- UX question routing and specialist investigation
- Evidence quality, uncertainty and downside
- Recommendation, action and outcome continuity
Your product team owns
- Goals, roadmap and priorities
- Requirements and delivery coordination
- The final product decision
- Implementation in the customer product
Human assurance owns
- Consequential or ambiguous recommendations
- Sensitive research and participant safety
- Unresolved specialist tensions
- Formal review where professional judgement is required
Start with a real question
Bring one question. See the intelligence become action.
Start with a real customer or experience question. You will see which sources Sightspool can use, how it routes the work, what it can honestly measure today, and where a senior human enters.