💡 AI-driven inspection completeness coaching with adaptive observation prompts

[github-actions[bot]]

Summary

After each inspection, score its completeness against the recommended observation set for that hive's current state and season. Surface a supportive coaching card: "Great coverage on 8 observations. Next time, try noting honey stores too — it helps the AI give you better feeding recommendations." Adapt the prompted observations based on colony risk profile, season, and beekeeper skill level. This closes the loop between inspection data quality and recommendation confidence.

Market Signal

Beekeeping Diary claims "first AI-powered analysis of individual hive condition" but analyzes the hive state, not the inspection quality. No beekeeping competitor coaches beekeepers on what observations they should be capturing. In the broader field inspection industry, AI-powered audit completeness checks are emerging as table stakes: Inspectly360 (agriculture) and Clappia (field teams) both launched completeness scoring in 2026. HiveSense ships voice capture but provides no guidance on what to observe — it captures what beekeepers say but doesn't guide what they should look for.

User Signal

PRD FR35 defines milestone-based skill progression tied to user activity and learning progress, but lacks concrete progression signals beyond badge milestones. The architecture's recommendation contract requires action + rationale + confidence + fallback — but recommendation confidence directly depends on inspection data quality. If beekeepers consistently miss key observations (e.g., never logging stores estimates), the recommendation engine operates with partial information and systematically lower confidence scores. Existing idea #499 (contextual skill-building micro-lessons) teaches during decisions; this teaches during inspections — complementary, not overlapping.

Technical Opportunity

The observation schema (observations table with JSONB structured_data and observation_type column) already supports checking which observation types were logged per inspection. Recommended observation sets can be generated from the existing recommendation engine's context assembly (Cloud Memorystore). Scoring is purely algorithmic — compare logged observation_type values against the recommended set for that hive's state, season, and the beekeeper's skill level. The coaching card is a standard Gluestack Card component with tva() variants for positive/constructive framing. No new infrastructure needed.

Assessment

Dimension	Score	Rationale
Feasibility	high	Algorithmic comparison of logged vs. expected observations; standard UI card; no ML required
Impact	high	Directly improves recommendation quality by improving input data; provides concrete skill progression signal
Urgency	med	Most valuable after the inspection flow (Epic 8) and recommendation engine are functional

Adversarial Review

Strongest objection: Beekeepers may feel judged or pressured, undermining Broodly's "calm, supportive, recovery-oriented" emotional design principle. Scoring inspections risks making the app feel like a test rather than a tool.

Rebuttal: Framing is everything. Not "You scored 60%" but "Nice work on your brood observations! For even better feeding recommendations next time, try noting honey stores too." This aligns with the PRD's emotional arc (anxious → guided → confident) and follows the same pattern as fitness apps that celebrate streaks while suggesting improvements. The coaching card is dismissible and never blocks workflow. Advanced beekeepers can disable it entirely. The core insight: the score exists to improve recommendations, not to grade beekeepers.

Suggested Next Step

Define the "recommended observation set" algorithm: given hive state (from last inspection), season (from region + date), risk profile (from recommendation engine), and skill level (from user profile), which observations should be prompted? Start with a static mapping (e.g., spring inspection always recommends: brood pattern, queen sighting, stores estimate, varroa indicator check) and make it adaptive post-MVP.