Explore Prompts

Define guardrails that block large architectural changes unless an explicit human-approved RFC exists. Include heuristics to detect “too big” diffs.

Create an agent that builds a “system map”: modules, ownership, dependencies, and critical paths. Use it to constrain edits to low-risk areas first.

Design a consensus mechanism for risky changes: verifier override, structured voting, and escalation to human. Include disagreement logs and rationale storage.

Build an eval harness for self-edits: deterministic tool mocks, seeded randomness, replayable runs, and stored artifacts for auditing decisions.

Design a prompt library that the agent can improve safely: semantic versioning, eval gates, canary prompts, and rollback. Include prompt linting rules.

Create a sandbox spec: network isolation, filesystem constraints, resource limits, and safe temp dirs. Include how MCP tool wrappers enforce the sandbox.

Design secret-handling: vault integration, redaction in logs, and automated leak scans on diffs, logs, and artifacts. Include fail-closed behavior.

Create a safety case template per capability: hazards, mitigations, tests, monitoring, rollback, and residual risk. Require evidence links and sign-off rules.

Design a reviewer UX: diff summaries, risk flags, test results, and “approve with conditions.” Include prompts that help humans spot hidden risks quickly.

Define autonomy tiers for the self-improver. Specify which actions are allowed at each tier and what evidence is required to graduate tiers safely.

Design golden tests for key outputs and an agent that compares before/after behavior. Include tolerance rules and how to prevent “golden drift” over time.

Create a tracing and logging spec for the self-edit loop: prompts, tool calls, diffs, test results, deploy events, and user impact metrics.