Claude Code: AI Agent Workflow Collection

Guide creation of AI Developer Workflows (ADWs) that combine deterministic orchestration code with non-deterministic agents. Use when building automated development pipelines, designing AFK agent systems, or implementing the PITER framework.

Create specialized agent experts with pre-loaded domain knowledge using the Act-Learn-Reuse pattern.

Implement hooks for permission control and security in custom agents.

Manage agent fleet through CRUD operations and lifecycle patterns. Use when creating, commanding, monitoring, or deleting agents in multi-agent systems, or implementing proper resource cleanup.

Guide creation of focused single-purpose agents following the One Agent One Prompt One Purpose principle. Use when designing new agents, refactoring general agents into specialists, or optimizing agent context for a single task.

Track and measure agentic coding KPIs for ZTE progression. Use when measuring workflow effectiveness, tracking Size/Attempts/Streak/Presence metrics, or assessing readiness for autonomous operation.

Assess agentic layer maturity using the 12-grade classification system (Class 1-3).

Audit codebase for agentic layer coverage and identify gaps.

AI governance and compliance guidance covering EU AI Act risk classification, NIST AI RMF, responsible AI principles, AI ethics review, and regulatory compliance for AI systems.

Leverage AI tools effectively for professional writing tasks. Provides prompt patterns, refinement workflows, voice preservation techniques, and quality checkpoints for AI-augmented drafting.

Use when designing APIs, choosing between REST/GraphQL/gRPC, or understanding API design best practices. Covers protocol selection, resource modeling, and API patterns.

Use when implementing API authentication, authorization, or security patterns. Covers OAuth 2.0, OIDC, JWT, API keys, rate limiting, and common API security vulnerabilities.

Attend the human-in-the-loop queue for loop-lane operation: ONE attention view merging worker-escalated items (human-gated role label + machine-marked escalation comment) with untriaged raw intake, then drive each row to resolution — answer escalated questions via interview, write answers back as issue comments, ratify first-drain C3 admissions, and flip unblocked items to the autonomous-eligible role label. Use when: 'attend the queue', 'attend queue', 'answer escalations', 'work the escalation queue', 'what needs my attention across the lanes', 'HITL queue', 'ratify admissions', 'clear the human queue'. Attended lane of the loop-lane three-session topology — judgment only; never executes work items, never merges. Composes /work-items:triage (attention view + machinery) and /planning:interview. Sibling skills: /work-items:work-loop (autonomous drain), /work-items:triage (raw intake), /work-items:track (backlog CRUD).

Adversarially validate a completed /planning:interview's answers with independent fresh-context agents — runs over any filled decision ledger, whether the human hand-answered the rounds or the recommendations were auto-accepted. Validators re-examine each answer with its rationale withheld and return a per-answer verdict (confirmed / challenged / reclassified-to-human), so only the doubtful answers come back as real questions and user-reserved decisions always do; if open branches remain, it accepts the recommended answers to fill them first, holding the never-auto floor. Use when: 'audit my interview answers', 'validate the interview answers', 'have agents check the answers', 'accept all and have agents check them', 'agent-validated interview', 'have subagents second-guess the recommendations', 'auto-answer then verify the ledger'. Not for stress-testing a plan artifact (that is '/planning:devils-advocate') or asking the human the questions the first time ('/planning:interview'); needs a filled interview ledger.

Audit a repo's Claude Code automation landscape (hooks, MCP servers, skills, subagents, scheduled tasks) against the enforcement hierarchy, producing PASS/REJECT/CONDITIONAL verdicts backed by evidence — the default verdict is REJECT because most gaps are already covered by compiler/analyzer/build-time checks. Use when: 'audit automations', 'what automations should we add', 'are we missing any hooks', 'hook gap analysis', 'should I add an MCP server for X'; pass --implement to apply approved items, or filter by category (hooks|mcp|skills|subagents|scheduled).

Classify code comments for four residue shapes — history narration (\"used to... now...\"), plan/session references (\"Task 2 replaces the old...\", \"in this PR\"), conversational antecedents (\"per your request\", \"as you asked\"), and ticket/PR/branch back-references a future reader will never see — emitting Tier 1 (remove) and Tier 2 (review) findings with treatment guidance; read-only, no edits applied. Use when: 'comment residue', 'audit code comments', 'find stale/narrative comments', 'strip conversational comments', or before committing agent-written code — not for removing ALL comments, restating-the-code redundancy (that is /code-tidying:tidy's Beck tidyings), or markdown noise (use /audit-noise).

Audit whether a documentation file earns its existence — could a fresh agent re-derive its conclusions by exploring the code, config, metadata, and structure itself? Read-only classifier: weighs each doc on derivability x re-derivation cost x drift risk x fact ownership and returns a verdict — delete, convert-to-pointer, keep-as-derivation-cache (drift-control required), or keep-owns-facts (rationale, decisions, constraints, external facts are non-derivable). Audience-aware: agent-facing surfaces get the full axe because exploration is cheap; human-facing docs weigh human derivation cost and clear a higher deletion bar. Use when: 'is this doc worth keeping', 'audit doc value', 'derivability', 'could an agent figure this out itself', 'should this doc exist', 'this doc just restates the code', 'prune redundant docs', 'does this doc earn its maintenance', before deleting a doc or before writing one that restates the codebase — not line-level noise inside a doc worth keeping (use /docs-hygiene:audit-noise), cross-file duplication (/docs-hygiene:extract-ssot), prose flavor (/docs-hygiene:compress), or code-mismatch staleness (a doc-drift detector).

Audit and remediate skill-encapsulation violations — external citations reaching into private surfaces inside .claude/skills/ / beyond the slash invocation. Use when: 'audit encapsulation', 'find skill leaks', 'skill boundary violation', 'who is reaching into ', 'check skill boundaries', 'public API drift', or before refactoring a skill.

Audit Claude Code hooks for safety, correctness, and best practices compliance. Checks settings.json configuration, hook scripts, error logging, and deployment readiness. Use when reviewing hook health, checking hook safety, validating hooks, or after adding/modifying hooks.

Audit locally-owned Claude Code instruction surfaces — user + project CLAUDE.md, .claude/rules, skill bodies, agent definitions, prompt-type hooks, output styles — for instructions current models no longer need: prior-model workarounds, over-prescriptive scaffolding, bare prohibitions, reasoning-echo directives, stale examples. Report-only: emits a findings report with proposed diffs, gated to the human, never auto-applied. Use when: 'after a model upgrade', 'are my instructions holding the model back', 'instructions the model no longer needs', 'too prescriptive', 'audit instructions', 'instruction audit'. Not a brevity pass and not memory-layer hygiene.

Classify tracked markdown for five noise shapes — historical citations, ghost refs to ephemeral working-directory paths, \"Why this file exists\" preambles, hard-coupled enumerated consumer lists, and scope/loading meta-commentary — emitting Tier 1 (remove/relocate), Tier 2 (review needed), and Tier 3 (likely legitimate) findings with per-shape treatment guidance; read-only, no edits applied. Use when: 'audit markdown noise', 'declutter', 'check for stale citations', 'find ghost refs', 'classify preamble', 'sweep a rule/skill/convention doc for noise', or before editing any tracked .md — not for prose flavor/compression (use /compress) or structural markdown lint (your repo's markdown linter).

Run ONE coordinated, ordered, resumable pass over a named target repository: a three-scope inventory first (managed policy read-only, user scope routed as recommendations, project scope), then delegated checks lane by lane, findings persisted per lane so an interrupted run resumes instead of restarting, and one human gate for the whole pass instead of one per skill. Adds run semantics, not checks — every check belongs to the plugin that owns it and is invoked presence-gated. Read-only on bare invocation; edits only behind an explicit --fix override, and never to managed policy or user-scope files. Use when: 'audit pass', 'run one pass over this repo', 'coordinate the audit skills', 'audit this repo end to end', 'resume the audit', 'one reconciled findings report', 'sweep all three scopes', or before a release that needs a single diffable findings artifact.

Audit Claude Code permission GRANTS for portability and auto-mode durability — scans skill/command/agent frontmatter allowed-tools and settings.json/settings.local.json permissions.allow for interpreter-wildcard rules dropped in auto mode, hardcoded machine/user paths, and inert plugin self-grants. Use when: 'check permission rules', 'why was my allowed-tools grant ignored', 'audit allow rules', 'is this permission portable', after authoring a code-execution grant, or when a guarded helper is denied despite an allow rule. Report-only.

Run one repository's pull-request queue as the merge lane of the loop-lane three-session topology: a self-paced standing or drain loop that invokes /source-control:babysit-prs each cycle at the resolved autonomy tier, layered with an activity grace window, do-not-merge respect, escalation, and lane telemetry. Merge authority is human-only until the target repo's tracked config adopts the lane; the adopted baseline is human merge for everything except gate-proven C2-mechanical PRs, and standing merge-rung raises bind only from the tracked config seam. One named exception: an invocation whose own argument line carries BOTH the 'autopilot' tier keyword AND the dedicated raise argument '--merge c3-this-run' (never a config value, never model-supplied, never inferred from a drain/merge phrasing) widens that single invocation's merge authority up to C3, with a fresh independent frontier-tier subagent resolving needs-human/thread/finding blockers first — C4-structural and C5-untrusted-provenance stay unconditionally human-merge regardless, and 'autopilot' alone leaves merge authority at the tracked rung. Use when: 'babysit loop', 'run the babysit loop', 'stand up the merge lane', 'babysit the PR queue continuously', 'drain the PR queue', 'keep merges flowing'. Required argument: . Launch via /loop (self-paced). Sibling skills: /source-control:babysit-prs (the single-pass tiered mechanic), /source-control:pull-request (single-PR lifecycle).