Reference

Quick-access reference for ulk commands, paths, variables, and terminology.

Essential Commands Cheatsheet

Session lifecycle

Command	Effect
`/ulk:bruce go`	Detect project state, take the most logical action
`/ulk:bruce status`	Full project diagnostic
`/ulk:bruce next`	Execute next task from docs/todo.md
`/ulk:bruce fix`	Find and fix errors
`/ulk:bruce audit`	Code quality audit
`/ulk:bruce audit all`	4 parallel audits (code, perf, a11y, SEO)
`/ulk:bruce review`	Pre-delivery completeness review
`/ulk:bruce ship`	Pre-release checks
`/ulk:bruce sync`	Sync with Notion/Linear
`/ulk:godspeed`	Quick project diagnostic (30s, cached 30 min)
`/ulk:2b3`	End-of-session: lint → tests → commit → memory capture
`/ulk:xavier`	Verify work context (account, repo, last state)
`/ulk:gandalf`	Context health check
`/ulk:gandalf status`	Quick context zone assessment

Context hygiene

Command	Effect
`/rewind`	Undo last turn, reformulate from clean state
`/clear`	Clear context (use between unrelated tasks)
`/compact`	Compact with preservation instructions
`/compact Preserve: X. Discard: Y.`	Targeted compact
`/effort low`	Minimal reasoning (mechanical tasks)
`/effort medium`	Most prompts — efficient default
`/effort high`	Default for Opus 4.8 agentic coding
`/effort xhigh`	Hardest tasks + long async workflows (was Opus 4.7 default)
`/effort max`	Rarely justified

Audits

Command	Effect
`/ulk:sargeras`	10-axis full project audit
`/ulk:ed209`	OWASP Top 10 security audit
`/ulk:killbill`	Cloud cost audit (Vercel, GitHub, Neon)
`/ulk:blackemperor`	All 4 audits in parallel, consolidated report
`/ulk:perf-auditor`	Core Web Vitals, bundle, database
`/ulk:a11y-auditor`	WCAG 2.1/2.2 accessibility
`/ulk:seo-auditor`	SEO technical audit
`/ulk:vision`	Code quality, 8 dimensions
`/context-audit`	Token health score 0-100
`/ulk:verify`	Spec-to-code conformance (3 dimensions × 3 severities)

Documentation

Command	Effect
`/ulk:shuri`	Generate/update spec.md, todo.md, CLAUDE.md, README.md
`/ulk:strange`	Reverse-engineer spec from code
`/ulk:strange mode=prompt`	Reverse-engineer prompt from AI output
`/ulk:lovecraft memory capture`	Capture MEMORY.md → docs/_memory/
`/ulk:lovecraft memory surface`	Read-only vault summary for current session
`/ulk:lovecraft memory distribute`	Inject vault into CLAUDE.md
`/ulk:lovecraft memory dream`	REM consolidation: dedup, normalize, archive
`/ulk:agamotto`	Reverse-engineer design from Figma URL → docs/design.md

Skills

Command	Effect
`/caveman`	Activate terse output mode for session
`/caveman off`	Deactivate terse output mode
`/symbols list <file>`	List all TypeScript/JavaScript symbols
`/symbols view <file> <symbol>`	View one symbol body
`/symbols refs <file> <symbol>`	Find references across codebase
`/context-mode query <id>`	Retrieve stored verbose output
`/context-mode list`	List N most recent stored outputs
`/context-mode stats`	Interception rate, tokens saved
`/architecture-diagram`	Generate HTML+SVG architecture diagram

Deployment

Command	Effect
`/ulk:deploy-vercel`	Deploy to Vercel
`/ulk:deploy-netlify`	Deploy to Netlify
`/ulk:deploy-cloudflare`	Deploy to Cloudflare Workers/Pages
`/ulk:deploy-aws`	Deploy to AWS
`/ulk:deploy-docker`	Docker build + push
`/ulk:deploy-fly`	Deploy to Fly.io

Installation and maintenance

Command	Effect
`./install.sh`	Install or reinstall ulk
`./install.sh --dry-run`	Simulate without writing files
`git pull && ./install.sh`	Upgrade to latest version
`ulk install --audit-only`	Pre-install audit without writing to ~/.claude/
`ulk clean`	Remove orphan files (renamed/deleted agents)
`ulk clean --dry-run`	List orphans without deleting
`ulk skills list`	List declared external skills
`ulk skills update`	Fetch/install all external skills
`./framework/tools/check-tools.sh`	CLI + skills diagnostic
`node framework/cheatheet/generate-registry.cjs`	Regenerate registry.json + registry.md
`node framework/cheatheet/generate-commands.cjs`	Regenerate commands/ from agents/

Managed Agents (cloud)

Command	Effect
`ulk ma sargeras "<prompt>"`	Run sargeras as a Managed Agent
`ulk ma ed209 "<prompt>"`	Run ed209 as a Managed Agent
`ulk ma 2b3 "<prompt>"`	Run 2b3 as a Managed Agent
`ulk install --audit-only --with-managed-agents`	Audit managed agents setup

RTK (token compression)

Command	Effect
`rtk proxy <cmd>`	Compress a command’s verbose output
`rtk gain`	View cumulative savings
`rtk gain --history`	Command history with per-command savings
`rtk --version`	Verify correct RTK is installed

Environment Variables

Variable	Default	Effect
`CLAUDE_CODE_DISABLE_1M_CONTEXT`	unset	Set to `1` to disable Opus’s 1M context window (4.7/4.8), force 200K (auto-compact at ~155K)
`CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`	unset	Set to `80` to trigger compact at 80% rather than at overflow
`ANTHROPIC_API_KEY`	required for Managed Agents	Anthropic API key (first-party only — not Bedrock/Vertex)
`AUTO_DREAM_MIN_HOURS`	24	Minimum hours between auto-dream cycles
`AUTO_DREAM_MIN_SESSIONS`	5	Minimum sessions between auto-dream cycles
`CHROME_PORT`	9222	Chrome DevTools Protocol port (mb CLI)
`CHROME_BIN`	system	Chrome binary path (mb CLI)

Write environment variables in ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_DISABLE_1M_CONTEXT": "1",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "80"
  }
}

Do not set CLAUDE_CODE_SUBAGENT_MODEL — it overrides all agents’ model: frontmatter declarations.

Important File Paths

Repository structure

Path	Contents
`framework/agents/`	All 90 agent definitions (markdown files)
`framework/agents/_shared/`	Shared protocols, base rules, model policy
`framework/agents/registry.json`	Auto-generated agent registry (count, names, phases, categories)
`framework/cli/`	ulk CLI source (Go)
`framework/packages/core`	Shared TypeScript library (parser, types, GitHub client)
`framework/packages/status-board`	Project tracking dashboard
`framework/schemas/`	JSON schemas (todo, spec, project, dashboard)
`framework/community-skills/`	Vendored community skills (source)
`framework/tools/`	Shell scripts (cli-registry.json, accountability-report.sh, etc.)
`framework/managed-agents/`	Managed Agents orchestrators and triggers
`framework/cheatheet/`	Registry and command generator scripts
`docs/guides/`	Technical guides (token-optimization, large-codebase, managed-agents)
`docs/research/`	Spike documents (spikes are research notes for features)
`docs/audits/`	Archived audit reports (sargeras, ed209)
`site/`	Documentation site (Astro, ulk.regrets.app)

~/.claude/ (installed files)

Path	Contents
`~/.claude/commands/`	Slash commands for each agent
`~/.claude/skills/`	Installed skills (figma-, swift-, kepano-*, etc.)
`~/.claude/hooks/`	Optional hooks (memory loop, auto-dream, accountability)
`~/.claude/agent-memory-local/`	Persistent memory for 9 memory-enabled agents
`~/.claude/state/`	State files (context-mode.sqlite, auto-dream.json, auto-dream.lock)
`~/.claude/settings.json`	Global Claude Code settings
`~/.claude/mcp.json`	MCP server configuration

Project-level files (in your project, committed to git)

Path	Contents
`CLAUDE.md`	Project context for Claude Code (stack, agents, doc-mode)
`docs/spec.md`	Narrative spec (Obsidian frontmatter + sections)
`docs/todo.md`	Kanban board (Obsidian Kanban format)
`docs/design.md`	Design source of truth (palette, typography, tokens, anti-patterns)
`docs/backlog/`	Faru doc-mode: one card per folder (CARD.md)
`docs/_memory/`	Cross-session knowledge vault (managed by Lovecraft)
`docs/audits/`	Archived audit reports
`.ulk-reports/`	Accountability trail, audit reports (gitignored)
`.ulk/managed-agents/ids.json`	Managed Agent IDs (gitignored)

Model and Effort Recommendations by Task Type

Task type	Recommended model	Recommended effort
Full project audit (sargeras, ed209)	Opus	xhigh
Orchestration (bruce, blackemperor)	Opus	xhigh
Architecture decision (tony)	Opus	xhigh
Design system (stark, agamotto)	Opus	xhigh
Error fixing (robocop)	Opus	xhigh
Session checkpoint (2b3)	Sonnet	high
Task execution (task-runner)	Sonnet	high
Doc generation (shuri)	Sonnet	high
Deploy (deploy-vercel, etc.)	Sonnet	high
Frontend implementation (brique)	Sonnet	high
Quick diagnostic (godspeed)	Haiku	low
Context monitoring (gandalf)	Haiku	low
Mechanical reformatting	Sonnet	low
Most day-to-day prompts	Sonnet	medium
Heavy planning, trade-off analysis	Opus	xhigh

Glossary

Agent

A markdown file in framework/agents/ with a YAML frontmatter that defines its role, model, tools, and invocation. When installed, it becomes a Claude Code slash command (/ulk:<name>). Agents have no code dependencies — they are dependency-free markdown files.

Skill

A reusable instruction set installed in ~/.claude/skills/. Skills extend Claude’s capabilities for a specific domain (Figma integration, Swift patterns, Obsidian formatting). They are invoked as slash commands and can be loaded by agents via the loads_skills: frontmatter field.

MCP (Model Context Protocol)

A standardized protocol for connecting Claude to external tools and data sources. MCP servers provide tools (read/write file systems, query databases, call APIs). In ulk, CLIs are preferred over MCPs when available (CLI = 0 tokens vs MCP tool call overhead).

Sub-agent

A Claude Code sub-session spawned by the main session to handle a specific task with a clean context. The sub-agent returns only its synthesis to the main session. Used to prevent context pollution during heavy explorations (reading many files, searching a large codebase).

Phase

A lifecycle stage in the development process. ulk defines 7 phases: define, plan, build, verify, review, ship, orchestrator. Each agent declares its phase in frontmatter. Bruce routes to agents based on the current project phase.

Context rot

Gradual degradation of response quality as the context accumulates irrelevant content — failed attempts, corrected approaches, transitional summaries, repeated information. The 5 context hygiene rules exist specifically to prevent context rot.

Faru

ulk’s default documentation mode since 2026-05-18. A kanban-native, agent-first doc format where each spec is a folder (docs/backlog/YYYY-MM-DD-TYPE-SLUG/) with a CARD.md. One card = one deliverable. The format is git-native and directly dispatched to Claude agents.

Managed Agent

A ulk agent that runs as an autonomous cloud session via the Anthropic Managed Agents API. Triggered by cron, webhook, or CLI. Runs without a local machine. 5 agents are currently ported: sargeras, ed209, lovecraft, 2b3, ci-guard.

A high-scrutiny operating mode activated by Bruce on critical signals (security finding, scope creep, context >80%, audit warning). In Banner Mode, Bruce triple-checks every routing decision against three questions: why / risks / downstream impact.

Context Manifest

A frontmatter convention that lets agents declare their token budget, loaded files, loaded skills, and MCP usage. Bruce reads these declarations before routing to prevent context overflow. Key fields: context_budget, loads_files, loads_mcps, mcp_cap_strategy.

CLAUDE.md

The project-level instruction file read by Claude Code at session start. It contains the project context, stack, agent configuration, and doc-mode. Committed to git, it is the shared source of truth for team-wide AI behavior in a project.

Auto-Dream

A background memory consolidation cycle inspired by REM sleep. Runs when ≥24 hours have passed and ≥5 sessions have occurred since the last cycle. Merges duplicates, normalizes timestamps, archives obsolete entries, rebuilds the vault index. Write sandbox restricted to docs/_memory/**. Never touches code.

RTK

Rust Token Killer — a CLI proxy that compresses verbose command outputs by 60–90%. Invoked as rtk proxy <cmd>. Included in the base ulk installation.