AGENTS_INDEX

Agents Index — Selection Tree for This Machine

The canonical catalog of every Claude Code subagent and skill on this machine, organized so Claude can pick the right one for the job. Updated 2026-05-15.

This file is loaded into the global session context (referenced from ~/.claude/CLAUDE.md). Every project on this machine knows where to find it: C:\Users\IT_Admin\.claude\agents\AGENTS_INDEX.md.

---

How to access agents from any project

Claude Code loads agents from three locations:

Scope	Path	Loaded
Machine-wide	C:\Users\IT_Admin\.claude\agents\*.md	Every session, every project
Project-scoped	<project>\.claude\agents\*.md	Only when cwd is inside that project
Plugin-managed	C:\Users\IT_Admin\.claude\plugins\cache\ruflo\<plugin>\<ver>\agents\*.md	When the plugin is enabled (currently: ruflo-adr, ruflo-browser, ruflo-jujutsu, ruflo-testgen, ruflo-loop-workers — see ~/.claude/settings.json)

Invocation: Task(subagent_type="<name>", description="...", prompt="...") from any agent or main session. Auto-routing also fires when the user’s intent matches an agent’s description field.

Cross-project access: If you’re in project X and need an agent that’s project-scoped to Y, two options:

1. Run from inside project Y’s cwd (preferred — the agent already has its context).
2. Read the agent file directly from path and embed its prompt manually (degraded — no auto-dispatch, no context guarantees).

All agents share persistent memory through the Obsidian knowledge base at d:\Vault\AI Vault — see Memory & learning below.

---

Quick selection — symptom → agent

The user says…	Reach for
”Build me a marketing site / landing page”	marketing-design (orchestrator)
“My computer is slow / something’s broken in Windows”	windows-engineer
”What does IRC §X actually say?”	tax-researcher
”How do I reduce my taxes on Y?”	tax-strategist
”What entity should I form / what’s the S-corp election deadline?”	entity-tax-specialist
”What’s Indiana’s PTET / will I owe California tax if I move?”	state-tax-specialist
”Record this transaction / close the month for Studio 516”	bookkeeper (in TaxProfessional)
“Curate / clean up the tax vault”	tax-knowledge-curator (in TaxProfessional)
“Check for new IRS / Tax Court rulings”	tax-update-watcher (in TaxProfessional)
“Analyze the diff / score this PR’s risk”	git-specialist (plugin: ruflo-jujutsu)
“Drive the browser / test this UI”	browser-agent (plugin: ruflo-browser)
“Write tests for this module / find coverage gaps”	tester (plugin: ruflo-testgen)
“Create / supersede an Architecture Decision Record”	adr-architect (plugin: ruflo-adr)
“Run something in a loop / schedule a worker”	loop-worker-coordinator (plugin: ruflo-loop-workers)

---

The selection tree

agents/
├── system/
│   └── windows-engineer ─────── deep Windows-internals diagnostics, perf, repair, tuning
│
├── marketing-design/  (orchestrator: marketing-design)
│   ├── strategy/
│   │   ├── brand-strategist ─── palette, typography, voice, archetype, logo
│   │   └── copywriter ────────── headlines, value props, body copy, CTAs
│   ├── build/
│   │   └── designer ──────────── picks stack, ships running site (Next/Astro/SvelteKit + Tailwind)
│   └── refine/
│       ├── seo-specialist ───── meta tags, schema.org, OG cards, sitemap, robots
│       ├── conversion-specialist  funnel, CTA placement, friction, trust signals
│       ├── analytics-specialist   one provider, conversion events, UTM, consent
│       ├── accessibility-auditor  WCAG 2.2 AA pass
│       └── performance-engineer   Core Web Vitals + Lighthouse
│
├── tax-and-books/  (project-scoped to d:\Projects\TaxProfessional\)
│   ├── research/
│   │   └── tax-researcher ────── pulls IRC / regs / rulings / cases; verbatim quotes
│   ├── planning/
│   │   ├── tax-strategist ────── legal tax-minimization plays with IRC basis + risk label
│   │   ├── entity-tax-specialist  entity choice, elections, basis, distributions
│   │   └── state-tax-specialist   50-state SALT, PTET, nexus, residency, conformity
│   ├── operations/
│   │   └── bookkeeper ────────── transactions, monthly closes, year-end handoff
│   └── maintenance/
│       ├── tax-knowledge-curator   MOC, wiki-link audit, tag normalization, dedup
│       └── tax-update-watcher ── IRB / Tax Court / JCT / state DOR cadence monitor
│
└── plugins/  (managed by Claude Code plugin system)
    ├── ruflo-jujutsu/
    │   └── git-specialist ────── diff risk, reviewer suggestion, change classification
    ├── ruflo-browser/
    │   └── browser-agent ─────── Playwright automation, scrape, validate, screenshot
    ├── ruflo-testgen/
    │   └── tester ─────────────── TDD London School, coverage gaps, test generation
    ├── ruflo-adr/
    │   └── adr-architect ─────── ADR create / index / supersede / review
    └── ruflo-loop-workers/
        └── loop-worker-coordinator  schedule, dispatch, health-monitor background workers

---

Catalog (full)

windows-engineer — system

• Where: C:\Users\IT_Admin\.claude\agents\windows-engineer.md
• Use when: Windows slowness, software/driver issues, networking flake, registry / services / scheduled tasks, anything needing elevated authority.
• Authority: Pre-authorized for reads & most reversible changes; confirms before BCD / driver swap / disk repartition / Defender disable / mass deletion. See its own file for the full list.
• Memory tag: windows-sysadmin
• Model: opus

marketing-design — orchestrator

• Where: C:\Users\IT_Admin\.claude\agents\marketing-design.md
• Use when: A customer-facing landing page or marketing site is needed for any product (built by you, another agent, or already in the cwd).
• Dispatches: 8 specialists below, in phase order. Specialists pass context through shared files (BRIEF.md, BRAND.md, COPY.md, OPTIMIZATION_NOTES.md), not memory.
• Output: Deployable site in marketing-site/ subdirectory + launch-readiness scorecard.
• Memory tag: marketing-firm

brand-strategist — marketing/strategy

• Where: C:\Users\IT_Admin\.claude\agents\brand-strategist.md
• Use when: Defining customer profile, archetype, palette, typography, voice, logo direction (always Phase A of marketing-design).
• Output: marketing-site/BRAND.md.
• Memory tag: brand-strategy

copywriter — marketing/strategy

• Where: C:\Users\IT_Admin\.claude\agents\copywriter.md
• Use when: Writing every word on the site in the brand voice. Runs after brand-strategist.
• Output: marketing-site/COPY.md.
• Memory tag: copywriting

designer — marketing/build

• Where: C:\Users\IT_Admin\.claude\agents\designer.md
• Use when: Building the actual site from BRIEF + BRAND + COPY. Picks stack (Next.js / Astro / SvelteKit + Tailwind).
• Output: Running site in marketing-site/.
• Memory tag: web-design

seo-specialist — marketing/refine

• Where: C:\Users\IT_Admin\.claude\agents\seo-specialist.md
• Use when: Adding meta tags, OG/Twitter cards, schema.org JSON-LD, sitemap, robots, canonical URLs to a built site.
• Output: Edited site + SEO section appended to OPTIMIZATION_NOTES.md.
• Memory tag: seo

conversion-specialist — marketing/refine

• Where: C:\Users\IT_Admin\.claude\agents\conversion-specialist.md
• Use when: Reviewing built site for CTA placement, funnel friction, social-proof positioning, trust signals, form design. Tags CTAs for the analytics pass.
• Output: Edited site + conversion section in OPTIMIZATION_NOTES.md.
• Memory tag: cro

analytics-specialist — marketing/refine

• Where: C:\Users\IT_Admin\.claude\agents\analytics-specialist.md
• Use when: Wiring one analytics provider (GA4 / Plausible / PostHog / Fathom), conversion events, UTM capture, consent gate.
• Output: Edited site + analytics section in OPTIMIZATION_NOTES.md.
• Memory tag: analytics

accessibility-auditor — marketing/refine

• Where: C:\Users\IT_Admin\.claude\agents\accessibility-auditor.md
• Use when: WCAG 2.2 AA audit — keyboard nav, screen reader, color contrast, focus, motion preferences, forms.
• Output: In-place fixes + a11y report in OPTIMIZATION_NOTES.md.
• Memory tag: a11y

performance-engineer — marketing/refine

• Where: C:\Users\IT_Admin\.claude\agents\performance-engineer.md
• Use when: Final perf pass. Core Web Vitals (LCP / INP / CLS), Lighthouse, image / font / JS / third-party optimization.
• Output: Optimized site + performance report in OPTIMIZATION_NOTES.md.
• Memory tag: performance

tax-researcher — tax/research (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\tax-researcher.md
• Use when: Researching a specific IRC section, Treasury reg, IRS ruling/notice/PLR, court case, form, or “what does the law actually say about X.”
• Output: Cited primary-source notes under TaxProfessional/Federal/... or States/<ST>/....
• Memory tag: tax-research
• Model: opus

tax-strategist — tax/planning (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\tax-strategist.md
• Use when: “How do I reduce tax on X?” — legal tax-minimization strategies grounded in IRC basis, with risk labels.
• Output: Strategy notes under TaxProfessional/Strategies/. Refuses listed/abusive transactions.
• Memory tag: tax-strategy
• Model: opus

entity-tax-specialist — tax/planning (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\entity-tax-specialist.md
• Use when: Entity choice, formation election (Form 8832 / 2553), filing forms & deadlines, basis tracking, distributions, payroll, reasonable-comp, §1202 QSBS, entity-specific planning.
• Output: Per-entity notes under TaxProfessional/Entities/<entity>.md.
• Memory tag: tax-entity
• Model: opus

state-tax-specialist — tax/planning (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\state-tax-specialist.md
• Use when: Specific state, multi-state apportionment, residency / domicile, nexus (income or sales), PTET election, state conformity to federal IRC.
• Output: Per-state notes under TaxProfessional/States/<ST>/.
• Memory tag: tax-salt
• Model: opus

bookkeeper — tax/operations (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\bookkeeper.md
• Use when: Recording transactions, monthly closes, quarterly summaries, year-end tax-prep handoff for Studio 516 / IncomeProject / Personal-Adam.
• Owns writes: d:\Vault\AI Vault\Books\ (everyone else reads only).
• Memory tag: bookkeeping
• Model: sonnet

tax-knowledge-curator — tax/maintenance (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\tax-knowledge-curator.md
• Use when: Vault hygiene — MOC updates, wiki-link audit, tag normalization, dedup, stale “last verified” flagging, re-indexing.
• Memory tag: tax-curation
• Model: sonnet

tax-update-watcher — tax/maintenance (project-scoped: TaxProfessional)

• Where: d:\Projects\TaxProfessional\.claude\agents\tax-update-watcher.md
• Use when: Periodic cadence (weekly/monthly) check of IRB, Tax Court, JCT, congressional tax bills, state DOR newsrooms. Flags affected KB notes for review.
• Memory tag: tax-updates
• Model: sonnet

git-specialist — dev/git (plugin: ruflo-jujutsu)

• Where: C:\Users\IT_Admin\.claude\plugins\cache\ruflo\ruflo-jujutsu\0.1.0\agents\git-specialist.md
• Use when: Analyzing diffs for risk, classifying change types, suggesting reviewers, managing PR lifecycle.
• Backed by: mcp__ruflo__analyze_diff* MCP tools (stateless — no AgentDB writes).
• Skills attached: diff-analyze, git-workflow
• Model: sonnet

browser-agent — dev/automation (plugin: ruflo-browser)

• Where: C:\Users\IT_Admin\.claude\plugins\cache\ruflo\ruflo-browser\0.1.0\agents\browser-agent.md
• Use when: Playwright browser automation — navigate, interact, screenshot, validate UI, scrape structured data.
• Backed by: mcp__ruflo__browser_* MCP tools (per-session, no persistent writes).
• Skills attached: browser-test, browser-scrape, ruflo-browser
• Model: sonnet

tester — dev/testing (plugin: ruflo-testgen)

• Where: C:\Users\IT_Admin\.claude\plugins\cache\ruflo\ruflo-testgen\0.1.0\agents\tester.md
• Use when: Writing tests using TDD London School (mock-first, outside-in), detecting coverage gaps, generating test suggestions.
• Skills attached: testgen, test-gaps, tdd-workflow
• Model: sonnet

adr-architect — dev/architecture (plugin: ruflo-adr)

• Where: C:\Users\IT_Admin\.claude\plugins\cache\ruflo\ruflo-adr\0.1.0\agents\adr-architect.md
• Use when: Creating / indexing / superseding / linking Architecture Decision Records to code. File operations on docs/adr/*.md (the plugin’s prompt mentions AgentDB — that path is denylisted by policy; the filesystem ADR work still proceeds).
• Skills attached: adr, adr-create, adr-index, adr-review
• Model: sonnet

loop-worker-coordinator — dev/automation (plugin: ruflo-loop-workers)

• Where: C:\Users\IT_Admin\.claude\plugins\cache\ruflo\ruflo-loop-workers\0.1.0\agents\loop-worker-coordinator.md
• Use when: Scheduling persistent background workers via CronCreate, dispatching /loop workers, health monitoring.
• Skills attached: loop-worker, cron-schedule, ruflo-loop, ruflo-schedule
• Model: haiku

---

Skills (catalog by source)

User-invocable skills (auto-discovered, currently safe per machine policy)

From ruflo-* plugins (active):

• diff-analyze, git-workflow — git workflows
• browser-test, browser-scrape, ruflo-browser — browser automation
• testgen, test-gaps, tdd-workflow — test generation
• adr, adr-create, adr-index, adr-review — ADR lifecycle
• loop-worker, cron-schedule, ruflo-loop, ruflo-schedule — scheduling

From d:\Projects\ruflo\.agents\skills\ (auto-discovered when ruflo’s tree is in cwd, safe subset only):

• agentic-jujutsu — quantum-resistant version control concepts
• pair-programming — driver/navigator/switch modes
• performance-analysis — bottleneck detection
• verification-quality — truth scoring & quality gates
• skill-builder — create new skills with proper YAML
• sparc-methodology — Specification / Pseudocode / Architecture / Refinement / Completion
• stream-chain — multi-agent stream-JSON pipelines
• github-code-review, github-multi-repo, github-project-management, github-release-management, github-workflow-automation — GitHub-scoped workflows
• claude-api — Anthropic SDK / Claude API helper
• init — bootstrap CLAUDE.md
• review — review a PR
• security-review — security review of pending changes
• loop, schedule — built-in scheduling sugar
• update-config, keybindings-help, fewer-permission-prompts — settings & UX
• simplify — review changed code for reuse / quality

Skills denylisted by machine policy

Per C:\Users\IT_Admin\.claude\CLAUDE.md, the following auto-discovered ruflo skills must not be invoked because they write to a parallel persistent memory store (AgentDB / ReasoningBank / RVF / Hive-Mind / Flow-Nexus tenants):

• All agentdb-* (advanced, learning, memory-patterns, optimization, vector-search)
• All reasoningbank-* (agentdb, intelligence)
• hive-mind-advanced (also hive-mind)
• All flow-nexus-* (neural, platform, swarm)
• All v3-* (cli-modernization, core-implementation, ddd-architecture, integration-deep, mcp-optimization, memory-unification, performance-optimization, security-overhaul, swarm-coordination — these are ruflo’s own dev internals, not for use by other projects)
• worker-benchmarks, worker-integration (ruflo dev internals)
• swarm-orchestration, swarm-advanced, hooks-automation (write to AgentDB by default; native Task + SendMessage covers the ground without the side-effect)

If a skill prompt tells you to “store in AgentDB” or “use mcp__ruflo__memory_*”, route to Obsidian instead via node ~/.claude/bin/obsidian-research.mjs save .... The deny list in settings will reject the AgentDB call regardless.

---

Memory & learning

Every agent — user-level, project-scoped, or plugin — pins to the shared learning protocol at d:\Vault\Compendium\System\agents-learning-protocol.md (also on GitHub at https://github.com/adamhudson777/Compendium/blob/main/System/agents-learning-protocol.md). The protocol moved out of the Obsidian vault on 2026-05-21 along with the rest of the reference libraries.

The contract — full detail in that note — is:

1. Query first before researching: node ~/.claude/bin/obsidian-research.mjs query "<topic>" --hybrid --k 8 --tags "<your-canonical-tag>"
2. Adapt, don’t blindly apply — notes are time-stamped claims; verify before acting.
3. Save non-obvious findings with mandatory outcome:* tag: node ~/.claude/bin/obsidian-research.mjs save --title "..." --tags "<canonical-tag>,outcome:<accomplished|failed|partial>,..." --agent "<name>" --body "..."
4. Log at breakpoints: node ~/.claude/bin/obsidian-research.mjs log "<one line>"
5. Re-index after a batch: node ~/.claude/bin/obsidian-research.mjs index

Canonical tag per agent is fixed in the protocol’s tag table — don’t drift from it. Obsidian is the canonical memory layer on this machine; ruflo’s AgentDB / ReasoningBank / RVF stack is denylisted at the permission layer to avoid a parallel store.

---

How to add a new agent

1. Decide scope:
   • Machine-wide (useful in every project) → write ~/.claude/agents/<name>.md.
   • Project-scoped (only meaningful inside one repo) → write <project>\.claude\agents\<name>.md.
   • Plugin-managed (sharable across machines) → publish via the plugin system.
2. Frontmatter required: name, description (auto-router matches against this — keep specific), tools (subset to lock down, or omit to inherit), model (inherit, opus, sonnet, haiku).
3. Append a ## Memory & learning — firm conventions section that:
   • Cites [[System/agents-learning-protocol]]
   • Pins a unique canonical tag
   • Lists the 5-step loop with the agent’s specific query phrase
4. Update this file with the new entry under the right tree branch AND update the canonical tag table in d:\Vault\Compendium\System\agents-learning-protocol.md.
5. Verify Claude Code picked it up: in any session, the auto-router should match its description on the first relevant prompt.

---

How to deprecate an agent

Move (don’t delete) its .md file to ~/.claude/agents/_deprecated/<name>.md so the version history stays available. Strike its row from this file’s tree and catalog. Add a one-line entry in the deprecation log at the bottom.

Reference / experimental agent libraries

The following are project-scoped agent collections that are NOT auto-loaded outside their own project. They exist on disk but are out of scope for the firm-wide learning loop (no canonical tag wiring, no protocol adherence). Listed here so Claude knows they exist and can choose to read one ad-hoc if a specialist match is needed and no first-class agent fits.

Location	Count	Character	Notes
C:\Users\IT_Admin\Desktop\My First Agent\.claude\agents\	~171	Third-party general-purpose collection (data-engineer, microservices-architect, 3d-artist, ml-engineer, …)	Project-scoped; doesn’t follow the Obsidian memory protocol. Useful as one-off references when no canonical agent fits. Do not invoke them globally — bring their prompt content into a focused Task only when needed.
D:\Projects\IncomeProject\	1 (referenced)	autonomous-income-strategist is referenced in the project’s CLAUDE.md but the .md file is not currently installed at any standard location. Restore from git history if needed.

If any agent from these libraries earns its keep (gets used repeatedly across multiple projects), promote it: move the .md to ~/.claude/agents/, add a ## Memory & learning section pinning a canonical tag, and add an entry to the catalog above and the protocol’s tag table.

Deprecation log

• 2026-05-15 — d:\Vault\DesignFirm\*.md (5 files: brand-strategist, copywriter, designer, conversion-specialist, marketing-design) — stale 4-specialist marketing firm; superseded by the 8-specialist firm in ~/.claude/agents/. Files removed; folder’s README turned into a redirect.