Download the canonical skill
TERMINALbash
mkdir -p .agents/skills/shreddar-privacy
curl -fsS https://shreddar.pro/integrations/shreddar-privacy/SKILL.md \
-o .agents/skills/shreddar-privacy/SKILL.mdTen rules the agent must preserve
- Confirm the source is synthetic or authorized.
- State one narrow downstream purpose before moving context.
- Analyze uncertainty, then protect with an explicit policy and mode.
- Stop when SHREDDAR is unavailable or recommends local_only.
- Send only protected context or minimum required shards.
- Never restore, infer, request, log, or transmit removed values.
- Keep Vault, wallet, and API credentials outside model context.
- Inspect downstream output before any application-side reconstruction.
- Report receipt and token-meter evidence without reproducing sensitive values.
- Label missing price or settlement evidence instead of inventing a cost.
Test the skill, not just installation
- 01
DISCOVERY
Start a fresh agent session and confirm the runtime lists shreddar-privacy.
- 02
SYNTHETIC TASK
Invoke it with a clearly authorized fixture containing multiple entity classes.
- 03
STOP CASE
Disable the gateway and confirm the agent refuses to bypass SHREDDAR.
- 04
CREDENTIAL CASE
Use a fake credential pattern and confirm it is blocked, not forwarded.
- 05
REPORT
Confirm the result names policy, route, meter, and receipt without raw values.