Operating rules · compact Markdown output
Build agent rules for the job, not a generic mega-prompt
Choose a profile, then select the individual safeguards the agent needs. Each selected category adds one specific rule line to the generated file.
Purpose and audience
For people configuring an AI coding, research or automation agent that needs explicit outage, recovery, authority and reporting behaviour. Use it to assemble a concise rules file; then review the output against the actual job contract before use.
How to use it
- Choose one of the four named presets to load its checkboxes, or choose Custom to preserve your current selection.
- Read each safeguard’s benefit, token impact, autonomy impact, risk and exact generated rule.
- Select or clear safeguards; the preview updates immediately.
- Copy the Markdown or download
AI_AGENT_OPERATING_RULES.md.
Profiles
- Lean Default: checkpoint, reversible recovery, human gates, idempotency, progress, ownership, lean workflow and telemetry.
- Maximum Autonomy: all ten safeguard categories, while retaining human-only boundaries.
- Maximum Token Efficiency: six rules aimed at preventing repeated context, blind retries and unnecessary model/tool work.
- Development Safety: recovery, write fencing, crash/reboot handling, isolation and telemetry.
- Custom: the current manual selection; changing a checkbox switches the profile label to Custom.
Output behavior: each checkbox adds or removes the exact “Generated output change” sentence shown on its card. Profile selection changes the checked set; the common boundary rules remain in all outputs.
Method and limitations
The builder composes fixed text in the browser; it does not call an LLM or validate your agent’s runtime permissions. A generated instruction cannot grant credentials, authority or technical capability. Keep essential safety constraints and adapt them to the job.
Related: token optimisation rules, context cost calculator, methodology, experiments.
Configure operating rules
Choose a profile, then select the safeguards that fit the job. The preview produces compact Markdown you can review, copy or download. It is a starting point, not a substitute for task-specific policy.
Rule notes explain the trade-off: benefit, token impact, autonomy impact and risk. Generated rules omit this explanatory text to keep the downloaded instructions short.