Copy-paste templates for <dataDir>/system_rules.md. Each file in
this directory is a self-contained set of standing instructions
addressing one recurring agent-bias or workflow problem.
To use one:
- Open the file, copy the entire content
- In the running app, open Settings → System Rules
- Paste, hit Save
- Start a new chat (existing sessions inherit in-context conditioning from prior turns and will not pick up the change cleanly — by design, see ADR-0012 §3.4)
You can also combine rules from multiple examples in one file — System Rules is plain Markdown with no schema.
See docs/en/reference/system-rules.md
for the design and full semantics.
| File | When to use |
|---|---|
activity-log-audit.md |
Auditing terminal / user activity logs. Counters the LLM's strong prior to invent attacks ("spy", "compromised", "cyberattack") from routine sysadmin / dev activity. Forces evidence-based reporting with explicit benign alternatives. |
A good example:
- Solves a specific, repeated bias or workflow problem (not a "general purpose preferences" template — that's what Global Memory is for).
- Names the bias it counters in a short header.
- Uses hard rules with explicit forbidden words / phrases rather than vague preferences. LLMs honour hard rules far better than soft suggestions.
- Includes a calibration ladder (Observed / Possible / Likely) when the domain involves uncertain inference.
- Is short enough to copy-paste in one go — aim for under 100 lines.
Open a PR adding the new file plus a row in the table above.
Sibling library: ../shell_tools/ — same
pattern, but for opt-in shell-tool scripts that go into
<dataDir>/tools/ instead of system_rules.md. Parent
catalogue: ../README.md.