AI Agent & Automation Workflows

Use when designing, reviewing, or bootstrapping a source-backed memory layer for AI agents. Focus on durable facts, provenance, retrieval quality, privacy boundaries, update proposals, and eval-backed search rather than raw transcript storage.

Use when a task needs local authentication routing, API token loading, secret references, project environment files, password-manager CLI use, keychain-backed secret access, service account setup, or troubleshooting repeated auth prompts. Choose the least-exposure path by ownership: project-defined helpers first for project workflows, reusable local secret caches for repeated agent commands, and explicit user confirmation before creating new durable auth storage.

Review recent agent work, memory summaries, rollout notes, repeated failures, and workflow friction to decide which lessons should become durable skills, instructions, helper scripts, automations, or follow-up work. Use when a team wants to convert repeated agent operating lessons into reusable workflow improvements without publishing private logs or bloating global instructions.

Use when an agentic task needs clear boundaries for what context belongs in the parent prompt, child agents, sidecar models, memory, docs, evals, public artifacts, or final answers. This skill prevents context bloat and private-context leakage by classifying sources, sensitivity, provenance, retention, and allowed use before work is delegated or published.

Identify when agent, harness, skill, workflow, or AI system behavior needs deterministic controls instead of model judgment, and recommend production-grade implementation approaches. Use when designing, reviewing, or hardening agentic workflows, tool harnesses, skills, evals, routing, parsing, validation, permissions, retries, state machines, audit logs, CI gates, policy enforcement, or AI reliability controls.

Use when a user starts, reviews, or rescopes a complex agentic project that needs multiple nested harnesses, workstreams, sub-agents, eval loops, or public artifact tracks. This skill composes a parent harness and child harnesses with clear dependencies, success evidence, integration gates, model-role choices, and stop conditions before execution begins.

Coach and critique Harness-First Project Design for substantial AI or software starts, ambiguous initiatives, team training, reusable skill or workflow design, and project bootstrap reviews. Use when the work should design the operating harness before implementation: goal altitude, source of truth, context boundary, minimum project scaffold, first execution lane, verification loop, and reusable learning.

Use when the user wants to improve, simplify, sharpen, restructure, or rewrite their own writing, documents, messages, updates, proposals, memos, presentations, or communications so the point, story arc, ask, evidence, examples, and influence path are clearer, more persuasive, more concise, and still accurate.

Use when building, reviewing, deploying, or extending Model Context Protocol servers, especially TypeScript MCP servers, stdio transports, Streamable HTTP transports, remote MCP servers, hosted agent platforms, or tools that expose external systems to AI clients. Emphasize schemas, narrow tools, auth boundaries, smoke tests, and direct-access risk review.

Use when a task needs multiple independent model perspectives, sidecar model review, adversarial synthesis, benchmarkable ensemble reasoning, or comparison against a Fusion-style multi-model answer. This skill defines a deterministic council workflow with explicit worker roles, isolated prompts, structured outputs, a separate synthesizer, and verifiable routing through local CLIs or Vercel AI Gateway.

Use when a task may benefit from sub-agents, sidecar models, parallel reviewers, delegated implementation, independent verification, or nested agent workflows. This skill decides whether to delegate, defines each agent contract, limits context and permissions, and keeps the parent agent responsible for integration, verification, and final claims.

Use whenever the user starts a new project, initiative, repo, app, tool, research effort, reusable skill, local integration, or other substantial build by giving a goal, desired outcome, constraints, requirements, success criteria, or what good looks like. Design the smallest effective project operating harness before execution: intent, success evidence, failure modes, work mode, structure, verification loop, and first execution path.

Assess an SMB team, function, or workflow against the SMB AI Maturity Model and recommend the next capability move. Use when the user asks where a business, team, function, or workflow sits in AI maturity, what is blocking the next level, or what to do next.

Use when an agent needs to prove a claim, choose tests or evals, verify a public artifact, validate a workflow, check launch readiness, or decide what evidence is enough. This skill routes work to the smallest credible verification harness: tests, evals, smoke checks, screenshots, read-backs, diff review, secret scan, sidecar critique, or user review.

Use for high-stakes or uncomfortable decisions that need structured advisor personas, adversarial challenge, weighted tradeoff scoring, forced budget allocation, agreement/disagreement tracking, and an auditable decision ledger. Trigger when the user asks for a war council, advisor council, ruthless CFO/operator/customer-rep review, decision debate, forced prioritization, tradeoff call, uncomfortable recommendation, or a $100/$1000-style allocation across options. Do not use for ordinary brainstorming, simple summaries, copyediting, or low-stakes preference lists.

Use when the user wants to evaluate information, documents, proposals, updates, strategy notes, research, or jargon-heavy writing and distill the real point into plain-language, decision-ready signal while preserving accuracy, nuance, uncertainty, and the ability to expand detail incrementally.