65 lines
3 KiB
Text
65 lines
3 KiB
Text
# Extra LLM instructions for Home Assistant analysis
|
|
|
|
This is a generic template for optional owner instructions.
|
|
|
|
Copy it to `llm_instructions.md` and customize that local file:
|
|
|
|
```bash
|
|
cp llm_instructions.md.sample llm_instructions.md
|
|
```
|
|
|
|
`llm_instructions.md` is intentionally gitignored. Put private details there, such as real names, exact locations, entity naming conventions, or property-specific notes. Do not commit sensitive personal context to git.
|
|
|
|
The contents of `llm_instructions.md` are appended to the daily analysis prompt.
|
|
|
|
## Suggested article structure
|
|
|
|
- Write the article in three parts:
|
|
1. A short funny blog-style story/commentary in paragraphs, not bullets.
|
|
2. A short visible "Bottom line" or "Conclusion" section.
|
|
3. A concise serious briefing with only the most important data, anomalies, risks, and recommendations.
|
|
- The serious briefing should use short titled subsections so the webpage can collapse/expand them.
|
|
- Avoid overusing bullets. Use bullets mostly in the serious briefing section.
|
|
- Do not write or emphasize "Strong evidence"; assume it by default.
|
|
- Explicitly label uncertainty only when useful, for example: "Possible" or "Wild guess".
|
|
- Keep the whole article concise.
|
|
- Do not repeat observations or recommendations from previous articles unless today's data changes the conclusion or makes it newly important.
|
|
|
|
## Suggested serious briefing subjects
|
|
|
|
Keep roughly the same subjects each day, but title them naturally or humorously:
|
|
|
|
1. What actually happened / key data
|
|
2. Trends vs recent reports and behavior patterns
|
|
3. Privacy leaks, anomalies, and risks
|
|
4. Practical high-value recommendations
|
|
|
|
## Suggested analysis focus
|
|
|
|
- Occupancy and presence patterns
|
|
- Sleep/wake timing signals
|
|
- Lights, doors, windows, locks, motion, climate, media, batteries, and unusual sensor changes
|
|
- Privacy leaks: what could an observer infer?
|
|
- Practical Home Assistant automations or fixes
|
|
- Missing or ambiguous data should be called out honestly
|
|
|
|
## Optional local context examples
|
|
|
|
Replace these with your own private notes in `llm_instructions.md`:
|
|
|
|
- Entities with prefix `secondary_` belong to a secondary property. All other entities belong to the primary property.
|
|
- The primary property should be treated as higher priority than the secondary property.
|
|
- `person.example_vehicle` is a vehicle, not a person at home.
|
|
- `sensor.example_backup_age` is important because stale backups are a risk.
|
|
- Group observations by property when multiple properties are mentioned. Write the property label once, then list relevant bullets underneath.
|
|
|
|
## Optional custom questions
|
|
|
|
1. Did anything look unusual overnight?
|
|
2. Are any batteries, devices, or sensors acting suspicious?
|
|
3. Could the home infer when I am asleep, away, or busy?
|
|
4. What would make this setup more private or secure?
|
|
|
|
## Optional style example
|
|
|
|
Write in a dry, calm, slightly ominous deadpan tone: observant, factual, mildly sarcastic, and not emoji-heavy. The tone should flavor the report, not replace useful analysis.
|