← ClaudeAtlas

ratemyharnesslisted

Use this skill to rate, audit, red-team, or release-gate a concrete AI agent harness, runtime, orchestrator, runner, repository, deployment, trace set, or configuration. Use for loop correctness, tool dispatch and result correlation, context and memory isolation, permissions and sandboxing, approvals, retry and idempotency, timeouts, cancellation, budgets, termination, observability, cost, recovery, author defense, and same-rubric re-reviews. Trigger for wording such as rate my harness, audit this agent runtime, review my agent loop, red-team this orchestrator, is this agent production ready, harness 上线前挑刺, 给 agent harness 打分, or 这个智能体运行时能上线吗. Require an actual harness artifact or runtime evidence, including one in the current workspace. Do not use for generic agent architecture advice, ordinary code review, reviewing a standalone Skill or prompt, validating only a plugin manifest, debugging one model response, or evaluating a model without its runtime.
AmsonntagChow/ratemyharness · ★ 1 · AI & Automation · score 69
Install: claude install-skill AmsonntagChow/ratemyharness
# RateMyHarness ## Runtime boundary - **Model** proposes the next message, tool call, or action. - **Loop** invokes the model, dispatches tools, returns results, and decides whether to continue. - **Harness** enforces runtime behavior around the loop: context, state, memory, tools, permissions, approvals, budgets, cancellation, tracing, evaluation, and recovery. - **Skill** supplies instructions and resources inside a harness; it cannot grant authority or enforce a sandbox. An isolated loop is reviewable, but label its verdict component-limited rather than presenting it as a complete harness audit. ## Deterministic runtime invariants Use these zero-tolerance invariants as the stable center of every route: | Invariant | Required runtime behavior | |---|---| | Authority | Enforce policy outside model-controlled text and bind approval to the principal, action, arguments, and validity window. | | Correlation | Bind every model step, tool call, attempt, result, checkpoint, and terminal event to the correct run; reject or reconcile late and duplicate results. | | Isolation | Partition session and tenant state, enforce provenance and retention boundaries, and prevent cross-run contamination. | | Idempotency | Give non-idempotent effects stable end-to-end keys, durable state, and reconciliation so retries and resumes cannot silently duplicate them. | | Termination | Enforce independent step, tool-call, wall-clock, token, and cost limits, and propagate cancellation through activ