Contenu du pack

Un fichier SKILL.md (le guide pratique) + un modèle de départ. Aperçu :

Always-On Privacy Audit

When to use

• You're deciding whether a camera-and-microphone smart display belongs in your kitchen, living room — or bedroom.
• A vendor claims "on-device processing," "no ads," or "private by design" and you want to verify what those words actually cover.
• You're setting house rules for a shared AI display: what it may hear, see, store, and share.
• You're a reviewer or installer who needs a repeatable privacy checklist for smart-home gear.

Workflow

1. Inventory the sensors (input: spec sheet + device → output: sensor map)
   List every sensor: cameras (resolution, field of view, indicator light behavior), microphones (count, far-field range, mute behavior — hardware switch or software toggle?), and ambient sensors (temperature, humidity, presence, sound recognition for alarms). For each, note: can it be physically covered or disconnected, and what breaks if you do.
2. Trace the data path (input: sensor map + privacy policy → output: data-flow diagram)
   For camera/mic data, answer in plain sentences: where is it processed (on-device chip? which enclave?), what leaves the house (voice snippets? face embeddings? nothing?), where it's stored (how long, where, encrypted how), and who can access it (vendor employees? contractors? law enforcement process?). If the policy doesn't say, write "undisclosed" — that's a finding, not a gap in your work.
3. Test the recognition gates (input: device → output: gate test log)
   With each household member: stand in front of the display and check what a stranger would see — personal content with no one recognized, with the wrong person recognized, with the right person but no phone nearby. Try the adversarial cases: a photo of a member held up to the camera, a recorded voice sample played at the mics. Log pass/fail per case.
4. Audit the mute and the light (input: device → output: trust verdict)
   Press mute. Then verify independently: does the indicator light reflect a hardware state or a software promise? Check the vendor's documentation for what "mute" actually cuts (mics only? camera too? ambient sensors?). A mute you can't verify is a preference, not a protection — say so in the scorecard.
5. Check the money trail (input: privacy policy + terms → output: monetization verdict)
   Find the answers: is sensor-derived data used for advertising or ad measurement? Is it sold or shared with data brokers? Does deleting your account delete the stored voice/face data, and how do you verify? Any "yes" or "undisclosed" here caps the score no matter how good the engineering is.
6. Write the house rules (input: all findings → output: scorecard + rules)
   Score each dimension 1–5 (sensors, data path, gates, mute, monetization) and write the household rules that follow: room placement (never bedrooms for camera devices scoring under 4), who gets enrolled, guest protocol, and a quarterly re-audit date — because firmware updates change privacy properties silently.

Aperçu du modèle

Always-On Privacy Audit Worksheet

Work through with your AI before the device comes out of the box.

1. Sensor map

2. Data path (plain sentences)

• Where is camera/mic data processed?
• What leaves the house, exactly?
• Where is it stored, for how long, encrypted how?
• Who can access it (employees, contractors, legal process)?
• Mark every "undisclosed" — each one is a finding.

3. Recognition gate tests

4. Mute & indicator light

• What does "mute" actually cut (mics? camera? ambient?):
• Hardware switch or software toggle (evidence):
• Indicator light reflects hardware state or software promise:

5. Money trail

• Used for ads or ad measurement?

Le pack de compétence lui-même (SKILL.md + modèle) est rédigé en anglais.

Acheter le pack complet — $4.99 Pro — les 511 compétences