| name | solve-cv-task |
| description | Turn a plain-language CV goal into the correct measurable build route. TRIGGER when: user asks to detect/count vehicles, read serial numbers, track people through zones, use a factory camera to catch mistakes, solve a CV problem from sample images, test feasibility, or build an end-to-end vision capability without knowing the modality. SKIP when: request clearly matches a specialist; asks managed deployment cost, annotation/training cost, or other economics only; asks current platform how-to; invokes estimate-economics; or says an evaluated model works and only needs package/export/deployment/delivery.
|
| allowed-tools | Bash, Read, Write, Edit, Glob, Grep, AskUserQuestion |
Be Sentinel's low-barrier front door. Translate the user's operational outcome into the correct CV output, independent acceptance gate, fastest safe proof, and explicit next route—without requiring them to know CV terminology.
Roboflow platform knowledge lookup. Read ../../resources/roboflow-platform-lookup.md. Sentinel owns stable solution-building discipline. Installed official Roboflow skills or current MCP resources own exact product/model/tool truth. If neither is available, continue provider-neutral work and use a scaffold; never copy or guess a volatile platform recipe.
1. Read before asking
Inspect the user's files, sample images/clips, code, README, labels, and existing evaluation artifacts. Summarize the observable problem in plain language. Ask at most three questions that artifacts cannot answer:
- “What should happen when the camera sees the target?”
- “Out of 100 real cases, how many misses or false alarms are acceptable for a first proof?”
- “Where will images come from, and may they leave this machine/site?”
For a novice, define only the next needed term. Use “catch rate” before introducing recall and “false-alarm rate” before precision.
2. Route by required output
| User needs | Route | Key discriminator |
|---|
| boxes, object count, crops, per-object metadata, per-person PPE | detect-and-analyze | one output per visible instance |
| one verdict for the whole image, including whole-image compliance | classify-or-flag | one label/flag per image |
| masks, contours, area, crack width, calibrated physical measurement | segment-and-analyze | pixel geometry and calibration |
| persistent identity, crossings, paths, dwell, video events | track-and-count | association across frames |
| text, numbers, serials, forms, codes, meters | read-text | character/field extraction |
| keypoints, joint angles, posture, gestures, actions | recognize-pose-or-gesture | skeleton/keypoint semantics |
| several perception/rule/aggregation stages | decompose-to-pipeline | end-to-end staged decision |
| annotation/training/deployment economics or crossover | estimate-economics | economic decision, not build |
| accepted model/pipeline needs package, integration, monitoring, deployment | deliver-cv-project | capability already measured |
| setup/authentication health | check-sentinel-setup or auth-setup | environment/connection issue |
If two outputs are required, name a primary acceptance owner and compose the second stage explicitly. Do not hide a multi-stage system inside one skill.
3. Write the proof brief
Before routing, state:
- operational decision and action owner;
- input source and representative sample;
- required output schema;
- costlier error (miss or false alarm);
- independently produced gold evidence needed;
- first metric/threshold in plain language;
- privacy/data boundary;
- target runtime/latency when known;
- routed skill and why.
If the user cannot choose a metric, propose a clearly labeled draft based on their operational statement and ask for confirmation. The specialist freezes the acceptance revision before baseline work.
4. Handle feasibility honestly
Separate three states:
plausible: signal appears observable; measure it;
blocked: independent evidence shows resolution, lighting, viewpoint, timing, or sensor limits;
unverified: available inspection cannot decide; improve capture or obtain a specialist/domain review.
Do not call a project impossible because a general model failed to see it. Do not call it feasible because a demo looked convincing.
5. Hand off with continuity
Invoke or recommend the selected skill with the proof brief and inspected artifact paths. The specialist owns frozen acceptance, candidate measurement, artifact, and decision. A passing build routes to deliver-cv-project; a material scale/build-vs-buy question routes through estimate-economics.
- Independent human/sensor evidence owns acceptance; candidate output and pseudo-labels do not.
- Obtain explicit consent before upload/data movement and before paid actions.
- Never request a token or key for plugin installation.
- Never invent provider model IDs, commands, API fields, prices, or deployment paths.
- High-stakes medical, safety, or legal decisions require qualified human ownership.
Follow ../../resources/ledger-protocol.md. Record a routing/proof-brief event only when a durable brief is written. Let the owning specialist record acceptance and evaluation; never duplicate hook-covered MCP rows.
<stop_rules>
- Output/action remains ambiguous after three focused questions → write the ambiguity and smallest discriminating sample/check; do not guess.
- User asks only a current platform how-to → delegate upstream and return to Sentinel when a delivery decision exists.
- Existing accepted capability only needs integration → route directly to
deliver-cv-project.
</stop_rules>