The moment
01 / 13The forecast is information. The user still owns the difficult translation.
Maya plans a four-hour cycle outside Berlin. One app shows rain probability, another shows wind gusts, and an alert mentions thunderstorms for part of the region. None knows her route, exposed sections, latest return time, rain tolerance or whether a one-hour delay would solve the problem.
Scattered evidence
Three sources, no shared decision
Hourly forecast, radar movement and official warning describe different parts of the same risk.
Personal conditions
The same weather means different things
A short park ride, exposed ridge and child-carrying trip cannot share one recommendation.
Changing state
A good answer expires
A decision at 07:00 may be unsafe at 08:30 if the storm timing or route changes.
Job to be done
02 / 13Help me protect the activity, not merely understand the weather.
Success is not opening the forecast or reading a summary. Success is making a timely, confident and appropriately cautious decision, then avoiding a preventable surprise during the activity.
Initial focus
03 / 13Start with the decision conditions most likely to change the plan.
Each candidate is scored on frequency, consequence if missed and whether current evidence can support a useful recommendation.
Solution choice
04 / 13Earn the agent. Compare simpler ways to solve the job first.
Explains conditions but cannot connect them to the route, constraints or changing plan.
Good for hard warning thresholds, but brittle across activity, exposure, timing and user tolerance.
Works when the decision path is known, but forces users through every question and does not re-plan.
Useful because it gathers missing context, calls several tools, compares plan variants and monitors change.
Experience
05 / 13One decision, its reasons and the next moment it will be checked.
The experience opens with a decision rather than a weather paragraph. Users can inspect evidence, change a constraint and see the plan recompute.
What matters most for this ride?
Start at 08:00 and shorten the exposed final section.
This returns you before the storm window while preserving most of the ride. Recheck at 07:30; cancel if the official warning expands west or lightning probability enters the route window.
Scope line
06 / 13Be explicit about what is real, inferred and still simulated.
Real and designed
Decision model and boundaries
JTBD, state model, decision ownership, evidence contract, safety rules, evaluation design and recheck logic are product decisions.
Simulated here
Feeds, route and user history
Weather values, radar movement, route exposure and preference memory are scripted. No live safety claim is made.
The next proof step is one consented route connected to timestamped forecast, radar and official-warning feeds, run in shadow mode against decisions made by experienced outdoor users.
System design
07 / 13State and safety rules constrain the agent; tools supply current evidence.
Activity, route, time window, transport, companions, equipment, tolerance and latest acceptable return.
Product stateNormalises forecast, radar, official alerts and timestamps; exposes disagreement and freshness.
ToolOfficial severe warning, lightning proximity and stale evidence can block or withhold a recommendation.
RuleCompares start times, shorter routes and cancellation against the user's stated job and constraints.
ModelRechecks only the evidence that could change the accepted plan, until the user ends monitoring.
OrchestratorDecision ownership
08 / 13Every decision has an owner-and a reason the model does not own it.
| Decision | Owner | Justification |
|---|---|---|
| Is an official severe warning active on the route? | TOOL | A current external fact; it must be fetched, not generated. |
| Is evidence too stale to recommend? | RULE | A testable freshness threshold with a safe default. |
| Does lightning breach the safety boundary? | RULE | A deterministic risk gate, not a language judgement. |
| Which route sections are exposed during the risk window? | TOOL | Geospatial intersection over route and weather cells. |
| Which plan variant best preserves the user's job? | MODEL | Requires comparing incomplete trade-offs across timing, comfort and route. |
| Which missing question would change the recommendation? | MODEL | Contextual judgement over current uncertainty and available options. |
| Should monitoring begin? | PERSON | Requires explicit consent for notifications and location-linked processing. |
| Should the activity proceed despite a hard warning? | PERSON | The system can advise or refuse support; the real-world decision remains the user's. |
Evidence and refusal
09 / 13A recommendation expires with its evidence.
Every weather fact names the provider or official authority.
Forecast, radar and warnings carry independent timestamps.
The answer states the route area and time window it covers.
The system names the specific change that would alter the decision.
Missing
Ask one question
If latest return time changes the plan, ask for that-not a full questionnaire.
Stale
Refresh or withhold
If radar or warnings exceed their freshness limit, do not present a current recommendation.
Contradictory
Show the disagreement
When sources diverge materially, choose the safer plan and expose why confidence is lower.
Evaluation
10 / 13Test the decision, the explanation and whether monitoring catches change.
The evaluation set crosses activity type, route exposure, user tolerance, forecast disagreement, missing constraints and sudden warning changes. Safety and usefulness have separate release gates.
Value and learning
11 / 13The value is a better decision; the advantage is learning which evidence changed it.
A general forecast can describe weather. This product can observe which recommendation was accepted, whether the plan changed again and which condition caused that change-with consent and aggregated measurement.
Illustrative outdoor plans where weather affects a stated decision.
Retention, subscription or avoided support value per useful decision.
Leaves €28k monthly contribution before one-off build cost.
Rollout and stop conditions
12 / 13Increase influence before authority-and make rollback specific.
Evaluate past forecast changes and expert-labelled plan decisions.
Generate silently; compare with user choices and actual weather.
Show recommendation, evidence and recheck trigger. No monitoring by default.
Notify only when the accepted plan's decision state changes.
Stop customer recommendations
Safety or evidence fails
A hard warning is missed, a weather claim lacks current evidence, or plan-class performance falls below its gate.
Stop the investment
The decision does not improve
Users still cross-check manually, timely decisions do not increase, or monitored plans cost more than their validated value.
Decision record · 13 / 13
Ship a bounded decision companion, not an autonomous safety authority.
The product earns its place by assembling context, comparing plan variants and monitoring a decision. It does not invent weather, override official warnings, begin monitoring without consent or claim that an outdoor activity is safe.