← ClaudeAtlas

agentic-guardrailslisted

The rules that let implementation finish without asking the human — assumption ledger, repair budget, stop conditions, the shortcuts that fake a green build, evidence, blast radius. Read it before any autonomous implementation, refactor, migration or bug fix, and whenever a run is about to ask a mid-implementation question.
AlpayC/claude-skill-set · ★ 0 · AI & Automation · score 60
Install: claude install-skill AlpayC/claude-skill-set
# Agentic guardrails Shared reference for every skill that changes code without a human watching. The single source of truth for these five rules — other skills invoke this rather than restating them. Four of them also exist as hooks, installed by `guardrail-hooks`: destructive commands, protected paths and the blast radius, the shortcut list, and the evidence requirement. Where they are installed, those rules are enforced rather than requested. The reasoning stays here, because a block that does not explain teaches nothing. ## The ledger Mid-implementation questions are almost never blocking. They are decisions with an obvious-enough default and a cost of being wrong that a reviewer can absorb. So you do not ask — you **decide, record, and continue**. Every decision you would have asked about goes in the ledger at `docs/specs/<id>.assumptions.md`: ```markdown ## A3 — Error state for expired session **Decided:** Redirect to /login with a toast, matching `OrdersPage` (apps/orders/src/pages/orders.page.ts:88). **Alternative:** Inline error panel with a re-auth button. **Why this one:** Two of three comparable pages already redirect; consistency beats novelty here. **Blast radius if wrong:** One component, ~15 lines. **Reversal cost:** Low. ``` Write the entry *before* you implement the decision, so the ledger reflects reasoning rather than justification after the fact. A decision qualifies for the ledger when a competent reviewer could plausibly pick either option and t