m2ai-safety-verification

Plugin m2ai-safety-verification from marketplace m2ai-skills-pack in m2ai-portfolio/m2ai-skills-pack. Commands read from the repo's manifest — 2 days ago.

Skills (28)

• action-class-policy

  Interview the user about their agent usage patterns and risk tolerance, then generate a tiered action-class policy document. Each tier gets classification criteria, required approvals, rollback procedures, and monitoring expectations. Use when someone says "/action-class-policy", "define action clas

• agent-authority-map

  Audit and visualize what an agent is actually allowed to do — parses Claude Code settings, hooks, MCP server configs, and tool permissions to surface authority gaps where the agent takes actions no human explicitly approved. Use when the user says "agent authority map", "what can my agent do", "perm

• agent-blast-radius

  Inventory everything an autonomous agent has actually touched over a time window — files modified, commits authored, APIs called, processes spawned, scheduled tasks created — and produce a blast-radius report. Use when the user says "blast radius", "what has my agent done", "agent footprint", "audit

• artifact-contract-validator

  Validate a directory of claimed deliverables for file-type authenticity, formula presence in spreadsheets, embedded media in presentations, and PDF metadata sanity. Catches AI-generated fakes (HTML renamed as .pptx, empty spreadsheet cells passed off as data). Use before shipping any AI-generated de

• classify-loop

  Classify an agent workflow by the loop it operates in (code-generation, operational, supervisory, or fully-autonomous) and map each loop type to appropriate controls, approval requirements, and rollback paths. Use when designing a new agent workflow, auditing an existing agent deployment, or when so

• claudex-plan

  Enforced multi-round planning loop with a stop hook that blocks code execution until the plan survives adversarial review and receives explicit approval. Use when a task is high-stakes enough that a wrong plan is expensive to undo — the stop hook is the \"bouncer\" that prevents jumping straight to

• control-map

  Walk any agent workflow through a 7-row control map (runtime, governed data, identity/principal, action authorization, payment authority, observability, kill switch) to identify which row fails first and generate concrete remediations. Use when designing a new agent workflow, auditing an existing on

• describability-gate

  Eight-field readiness gate that blocks any automation project until inputs, outputs, decision rules, exceptions, owner, success metric, failure mode, and rollback are explicitly defined. Use before kicking off any automation build, as a pre-dispatch check, or when reviewing automation specs. Trigger

• destructive-op-guard

  Design or audit a PreToolUse hook that intercepts high-blast-radius shell operations (DROP TABLE, rm -rf, git push --force to main, kubectl delete, docker system prune, etc.) and requires explicit authorization before they execute. Use when adding a safety hook to a Claude Code project, reviewing an

• info-judgment-boundary-auditor

  Audit an agent system, workflow, or AI implementation to map where information retrieval ends and judgment calls begin. Flags implicit judgment that looks like retrieval, preventing the silent degradation of decision quality. Use when reviewing agent designs, deploying AI into decision workflows, or

• institutional-taste-encoder

  Formalize implicit quality judgments, style preferences, and "taste" into structured constraint specifications that agents can follow. Turns the unwritten rules in someone's head into grep-able, enforceable constraints. Use when the user says "encode my taste", "formalize quality rules", "taste enco

• judgment-gate

  A judgment classifier for planned tool calls or agent actions. Scores reversibility, cost boundary, blast radius, and user observability, then routes to ACT / ASK / QUEUE / ABORT with reasoning.

• judgment-layer

  Add a judgment classifier to any agent or Claude Code skill to decide when to act autonomously vs. ask for confirmation. Evaluates each candidate action against three axes -- reversibility, bounded cost, and user observability -- and routes to act-now, ask-first, or queue-for-review. Outputs a PreTo

• kill-switch-audit

  Verify that a named agent in production has a real, one-step kill path at each of 5 layers (runtime cancel, credential revocation, gateway block, payment freeze, workflow interrupt). Use before shipping a new autonomous agent, during a security review, or any time you need to answer "can we actually

• knowledge-work-tests

  Generate evaluation criteria, acceptance checks, and scoring rubrics for knowledge work outputs -- the missing "test suite" for non-code deliverables. Given a task spec (brief, memo, analysis, report, strategy doc), produce a rubric an agent can self-evaluate against before delivering. Use when defi

• management-function-audit

  Takes an org change description and classifies which management functions (routing, sensemaking, accountability) were removed, retained, or weakened. Predicts failure modes and recommends mitigations based on historical precedents.

• multi-artifact-work-package

  Turn a messy business situation into a full deliverable set with an explicit artifact contract — Word, PowerPoint, Excel, PDFs, press releases, FAQs — and verify each file is real (not HTML with a wrong extension). Use when a client engagement, proposal, or agent task requires multiple coordinated o

• plan-stop-hook

  Installs a \"bouncer\" hook in Claude Code that blocks all file edits until an explicit planning gate is passed. Prevents the common failure of jumping from idea to implementation before the plan has survived adversarial challenge. Use when a task is non-trivial, prior attempts drifted, or the cost

• policy-gen

  Generate the policy agent YAML security policies for agent sandboxing. Inspects a project's structure and produces deny-by-default policies for filesystem, network, and process capabilities. Use when the user says "the policy agent policy", "agent sandbox", "generate security policy", "sandbox confi

• pretty-but-wrong

  Final hostile-reviewer pass on any AI-generated document, deck, or workbook before sharing. Scans for unsourced claims, undated numbers, untraceable charts, inconsistent formulas, and logic gaps. Outputs a ranked must-fix/should-fix/polish issue list — never rewrites. Use before sharing any AI-gener

• promotion-packet-auditor

  Adversarial reviewer that scores a promotion packet or performance self-review for real judgment signal versus AI-generated polish. Surfaces unproven claims, impact assertions without decision context, and output descriptions that lack the reasoning a skeptical reviewer would demand. Use when the us

• recognizable-quality-test-generator

  Generate an LLM-as-judge quality rubric from a workflow description and historical examples. Use when building an agent that needs verifiable success criteria — converts "I know good when I see it" into a prompt-based pass/fail test the agent can run on its own output.

• source-packet

  Builds a structured source inventory from a set of files or documents before any artifact is created. Assigns each source an ID, status, and reliability tier; produces a conflict log where sources disagree; flags human-decision points. Use before starting any AI-generated deliverable (deck, report,

• spec-driven-dev-enforcer

  Enforce a spec-first development workflow that gates code generation behind explicit specification approval. Before any code is written, forces creation of (1) a requirements doc, (2) a design doc, (3) a task decomposition. Each phase produces a reviewable artifact. Code generation only begins after

• tie-breaker

  Use when Claude Code and another agent (Hermes, Gemini, Codex, or any second model) disagree on how to clear a blocker and a human needs a plain-language tie-breaker page. Trigger on \"tie-breaker\", \"/tie-breaker\", \"break the tie\", \"Claude and Hermes disagree\", \"two models recommend differen

• trust-score

  Score any AI-generated document, spreadsheet, or deck on a 0–100 trust scale across five verifiable dimensions. Outputs a per-dimension scorecard and a composite score. Use as a final gate before sharing any AI-generated artifact.

• vendor-pressure-test

  Run a vendor pitch, RFP response, or internal proposal through the 7-row agent control-layer lens to identify which rows are answered, dodged, or hidden behind buzzwords, and what each dodge costs in production. Use when evaluating a vendor for an agent deployment, responding to an RFP, or stress-te

• workbook-doctor

  Audits an existing or AI-generated Excel workbook for hidden risks before it is used in decisions. Runs four passes — structure map, formula risk scan, repair plan, verification memo — and produces a pass/fail checklist on formula consistency, source traceability, assumption clarity, and data freshn

Similar plugins