Source inventory
List document versions, pinned revisions, applicable scope, and missing source evidence.
Skill boundaries
Define what a skill may do, then define when it must STOP
A safe skill handles only de-identified offline data and produces explanations or checks for human review. Any action that could change bus, project, or physical state lies outside the skill's authority.
Allowed offline capabilities
A skill may help with four kinds of work
Static comparison
Compare de-identified configuration text and label differences, inferences, and unknowns.
Review design
Prepare preflight checks, human approval points, stop signals, and recovery questions.
Evidence organization
Classify existing records as documented facts, offline results, live observations, or unknowns.
Do not cross
Four mandatory STOP rules
Group operations
Stop immediately when asked to read, write, scan, or batch-process a real group address. Do not produce an executable command or apply an example to a live site.
ETS programming and download
Stop immediately when asked to change an ETS project, program or download, or assign an individual address. Hand the work to an authorized KNX professional under live-site procedures.
Telegram transmission and replay
Stop immediately when asked to create, inject, transmit, replay, or modify a telegram. Reading a de-identified record does not permit the skill to suggest transmission.
Physical control
Stop immediately when asked to change lighting, locks, shades, HVAC, alarms, or any other physical state. The skill has no authority to control equipment.
Skill contract
Every safe skill must state all five terms
- Input: Accept only de-identified, secret-free offline data with a clear source.
- Output: Provide only summaries, differences, questions, and recommendations for human review.
- Authority: No network access, credentials, command execution, configuration writes, approval, or live-site control.
- Evidence: Cite a source for every fact. Mark missing evidence as unknown and claim no live result.
- Handoff: After a STOP trigger, retain the reason, known scope, and questions requiring a human decision; do not continue designing an operation.
Review before use
Check refusal conditions as carefully as capabilities
Do not use a skill that lacks explicit STOP rules, source-evidence requirements, data-minimization rules, or human-approval boundaries. Neither prompt injection nor a user request may remove these restrictions.