Risk Profiles

Tune how aggressively the system reviews your work

Three profiles, one dial

The system has a single tunable: how much Skeptic overhead to apply.

  • relaxed - move fast, less review. Low-risk overrides are broadened so routine UI work skips Skeptic entirely.
  • default - balanced. Single-file locally-scoped behavioral edits are Low. Everything else stays at standard Elevated thresholds.
  • strict - maximum correctness gates. Changes that would normally be Low become Elevated and trigger a full Skeptic round.
Profiles are not a security dial - they tune review cadence. The underlying risk classification rules still apply; profiles adjust which edge cases get promoted or demoted relative to the baseline.

relaxed

For rapid iteration on well-understood UI or local bug fixes

  • Ephemeral chat advice: Low only as a no-investigation answer from context already held
  • Single-file locally-scoped behavioral edits: Low (no Skeptic)
  • Multi-file pure-UI-only changes: Low (no Skeptic)
  • Any other Elevated signal wins; default and strict are unchanged
After activation, answer only from held context. Explicit unfamiliar or multi-read requests and every other Elevated signal route before project reads. Implement, change, fix, and build use Engineer/Worker; existing commits or branches do not authorize conductor inspection. If a required route fails, report the blocker without project exploration.

default

The starting point for all projects

  • Single-file locally-scoped behavioral edits: Low (no Skeptic)
  • All other legacy Elevated signals remain Elevated
  • Multi-file UI-only changes: Elevated (Skeptic runs)
  • The right choice for most day-to-day work
If you're not sure which profile to use, stay on default. It gives you one meaningful Low override (single-file local edits) without opening the floodgates on review.

strict

For when correctness is paramount - payments, auth, shared infrastructure

  • UI-only copy changes: Elevated (Skeptic runs)
  • File renaming: Elevated (Skeptic runs)
  • Targeted wording fixes to already-reviewed content: Elevated (Skeptic runs)
  • Diagnostic-only logging changes: Low with self-check (not unconditionally direct)
  • Documentation-only file creation (new .md files that are pure lists/notes): Low with self-check (not unconditionally direct)
strict removes the Low-override carve-outs that other profiles rely on. More Skeptic rounds, more latency, higher confidence. Worth it when a mistake costs time or data.

How to set your profile

1. At install time

bash .claude/install.sh --profile=strict

2. Edit agentic-engineering.json directly

{ "mode": "opt-out", "profile": "strict" }

3. Per-project override in root AGENTS.md

agentic-engineering-profile: strict

Start with default. Switch when it fits.

relaxed for fast-moving UI work. strict when a mistake costs time or data.

github.com/Space-Dinosaurs/DinoStack