Unfurnished

Cursor-native workflow pack. Don't attach Claude furniture to Cursor. The room looks empty. That is the point.

This is not a token-golf optimizer. It points Agent at Plan, Ask, Debug, Task, Grep, Browser, and compact — the Cursor plan you already pay for — instead of cloning graph, memory, or router furniture. Skills fill gaps the IDE does not ship. Install it and talk; slash names are optional. Other plugins may run alongside. Not a Claude Code pack.

A few skills, a few knobs, a few short priors. Nothing that wraps something Cursor already ships. Stack-agnostic: same pack on a CLI, a SaaS, or a pipeline.

Start here

1. Customize → Plugins → Import marketplace, paste GitHub Troy-LL/Unfurnished. Each push to this repo updates the install. Re-import if an older install still says Cursor-Maxxing or cursormax.
2. Chat. The pack is on. Docs, features, kickoff, verify, and ship can pull themselves. A job with no checkable done-line gets an interview. You do not need /unfurnished first.
3. /unfurnished off mutes pack slots in this workspace (scratch/unfurnished-off; do not commit). /unfurnished on turns them back on. Customize → disable plugin is the hard off. Other plugins may run alongside; use soft-off only when you want Unfurnished quiet.

Open this repo only to edit the pack. If you already imported the plugin, do not also open this repo as the project — every skill loads twice.

Do not npx a Claude pack to get these files. Do not copy .cursor/ by hand unless you cannot install plugins.

Catalog: .cursor-plugin/marketplace.json (what Customize import reads). Plugin: .cursor-plugin/plugin.json.

/unfurnished if you want it

You type	What happens
/unfurnished alone	Orients. Reports if this workspace is soft-off. Fine. No job required.
/unfurnished plus a job	Takes the paste (messy and meta prompts are fine), strips fluff, names the Cursor native, then works or asks one fork.
/unfurnished off / on	Workspace pack knob. Default is on (no file). Soft-off only — plugin skills may still list.

What fires without you typing it

The test: could the agent usefully reach for this on its own? If yes, it is model-invoked. You can still type the / if you want that session by name. These fire from skill descriptions, not from the always-on prior.

Skill	When it should pull
sdd	Product docs: README, AGENTS.md, architecture, design, eval, one ADR
sdd-eng	A feature, fix, or refactor against that map
grill	No checkable done-line, align / stress-test a plan. /grill docs then sdd
blueprint	New project or whole-job kickoff. Not mid-task. No model picking unless they asked
verify	Multi-file work about to be called done
pre-flight	About to ship
ticket	Feature/fix with no path or no named failing test. Local scratch pointer, not GitHub
evals	Pipelines, traces into fixtures, cost class for this repo. Do not pin a vendor model
write-skill	Authoring a skill
after-compact	Chat was compacted. Opt out with /keep off
unfurnished	They asked what this pack is, or pasted a job after /unfurnished

tdd is the same idea as a prior: feature or fix, not a throwaway script. Adversarial review is not a skill here: use the bugbot / security-review Task types Cursor ships (008).

What only you start

These must not run unprompted. Every skill above is also /<skill>; there are no command twins (008).

Type this	What it does
/deepen	Skill, user-only. List deepening candidates, wait, grill the pick
/voice	plain | ste | off
/keep	Opt out of after-compact (/keep off). Default is on
/unfurnished off / on	Mute or restore pack slots in this workspace

What actually blocks

Two fences are hooks, not sentences (007). .cursor/hooks/fence.py denies a full-file write to a living durable owner (patch in place instead) and any git add / git commit that would put scratch/ in history. It needs python on PATH and fails open without it.

Product docs are sdd. Implementation is sdd-eng. Both ship in this pack, adopted from Troy-LL/troysdd (see .cursor/skills/sdd/UPSTREAM.md). You do not need a second plugin for them. If you already installed the troysdd plugin, turn one of them off so /sdd is not doubled.

Priors in .cursor/rules/: yagni-bias (~40 words) and unfurnished-bias (kernel: natives, one slot, no extra probes — not a slot map). Both always on. @yagni, tdd, and @blast-radius attach when the task matches, or when you @ them.

Why it looks like this

Four decisions do most of the work.

• Native first. Claude Code packs add graph indexes, memory, session files, routers, and overnight loops because that host needs them. Cursor already searches, remembers, routes, checkpoints, and compacts. Codegraph went out for that reason. A new add has to name a gap the IDE does not cover. → 001
• Few first shots, not token golf. Efficiency here means the first few attempts are accepted, out of the box, on whatever model the host picked. Not one-shot perfection. Cursor often does not show tokens — we count shorter sessions and fewer extra tool calls. → 002
• Workflow, not stack. We do not know what you are building and do not need to. Everything works the same on a Rust CLI, a Next.js SaaS, or a data pipeline. Stack guides and framework integrations stay out. → 003
• Kernel, not a slot map. Always-on is constraints. Which skill runs is per-task, from its description. We do not clone Claude’s JS workflow runtime; fan-out is Task. → 006

Two more keep it honest: hard fences are hooks, not prose (007); no command twins of skills and no review skill of our own (008).

What we deliberately leave out — if Cursor ships it, we do not wrap it

If it assumes your stack, it belongs in your project rules instead.

Leave out	Cursor already ships / Where it belongs
Graph indexes, codegraph, extra grep	Instant Grep, embeddings, Explore
Session or memory plugins	Chat resume, rules, Automation Memories
Auto-routers / named-agent orchestrators	Task, /create-subagent, Plan/Ask/Debug, best-of-n
Claude dynamic workflows / JS orchestrator	Task, /create-subagent, /best-of-n (006)
Command stubs that only point at a skill	/<skill> already invokes it (008)
Hook frameworks, hook observability dashboards	hooks.json, /create-hook
Checkpoint / rewind wrappers	Agent checkpoints, /rewind
ccusage-style cost CLIs	Usage dashboard, statusline token %
Caveman / rtk compact packs	Product compact, /summarize
Output-style plugins	User Rules, Agent/Plan/Ask/Debug modes
Statusline theme packs	cli-config.json statusLine
Permission MCP / safety-net plugins	Auto-review, permissions.json, sandbox
Worktree desktop managers	/worktree, /best-of-n, git worktree
Review-agent packs	Bugbot, Security Review
Ralph overnight loops	/loop, Cloud Agents, Automations
A second skill marketplace	Plugins, skills.sh, cursor.directory
Rule-authoring playbooks	/create-rule, Customize → Rules
Slash-command authoring playbooks	/commands, Customize → Commands
Browser / Playwright verify MCP	Native Browser
Canvas / design-to-code taste packs	/canvas, Design Mode
Plan / /autoplan wrappers	Plan mode
PR babysit loops	/babysit
Stack skills, framework guides, service integrations	App-specific .cursor/rules/, cursor-directory
Rewrites of battle-tested community skills	Adopt verbatim with upstream attribution (003); /sdd is the troysdd adopt

What an MDC is — and when to write one

An .mdc in .cursor/rules/ is a constraint injected into Agent context. It is not a skill, not a slash command, not AGENTS.md, and not a User Rule.

Cursor's rules docs give four attach modes. Pick one. The frontmatter is the mode.

Mode	Frontmatter	When it loads
Always	alwaysApply: true	Every chat. Globs and description ignored. Expensive.
Files	alwaysApply: false plus globs	Matching files are in context. Default for stack conventions.
Intelligent	alwaysApply: false plus description, no globs	Agent pulls it when the description matches the task.
Manual	alwaysApply: false, no globs, no description	Only when you @rule-name.

Official bar: keep them short, one concern, concrete examples or @ a file instead of pasting it. Add a rule when Agent makes the same mistake repeatedly. Do not copy a style guide the linter already owns. Do not document npm and git. Plain .md in .cursor/rules/ is ignored. Use AGENTS.md if you want unscoped markdown.

Write one when

• The fact is project-specific and Agent will violate it without a prompt
• It belongs on a glob (**/*.tsx, src/api/**) or on @ mention
• It is not already in README, AGENTS.md, User Rules, or a shipped skill

Do not write one when

• /create-rule or Customize → Rules is the authoring UI (do not wrap that)
• The text is routing (AGENTS.md) or personal voice (User Rules)
• The text is a workflow (skill) or a user-invoked / (command)
• The text is a hard stop (beforeShellExecution hook). Rules do not block git reset --hard
• You want it in every chat and it already lives in README. Point at the file.

This guidebook ships only the priors in .cursor/rules/. When an app needs a stack convention, type /create-rule or add it in Customize. Reach for an existing one before writing from scratch:

npx cursor-directory rules add <slug-or-url>

Skills: skills.sh / npx skills. Use write-skill when the add is a workflow, not a constraint. Do not install a Claude pack to get a marketplace.