Trigger

Activate this skill when the user says any of the following (or close variations):

"new project"
"set up a project"
"I want to start something"
"I'm starting a [project / initiative / effort]"
"help me get organized on [X]"
"create a project folder"
"can you set up a project for [X]"

Do not activate on vague requests like "let's work on something" that aren't clearly about project setup.

Behavior

Step 1 — Open with one question

Say exactly this (or close to it):

"Tell me about it — what are you working on?"

Do not open with a list of questions. Do not ask about template type, folder name, or deadline up front. One question. Let the user tell you what's on their mind.

Step 2 — Listen and infer

From the user's response, extract:

Project name: Derive a short, clean folder name from what they describe. Use lowercase with hyphens (e.g., q3-board-report, vendor-evaluation, onboarding-redesign). Don't ask for a folder name — pick one and state it. They can correct you.

Template type: Pick one of three based on signals in their description.

Template	Use when they mention...
`writing-communication.md`	emails, proposals, reports, presentations, drafts, announcements, recurring communications, writing for an audience
`project-tracking.md`	tasks, deadlines, team coordination, blockers, milestones, status updates, managing work with other people, a project with a due date
`research-summarization.md`	comparing options, reading documents, finding information, making a decision, competitive research, summarizing, briefings

If signals are mixed (e.g., "I'm researching vendors and will write a recommendation"), pick the type that matches the output they're building toward. A recommendation report = writing. A vendor selection decision = research. When genuinely unclear, name both options and ask which fits — but only if you truly cannot infer.

Key people: Note any names or roles mentioned. Don't ask — capture what surfaces naturally.

Goal / what done looks like: Capture if stated. If not, ask in Step 3.

Step 3 — Reflect back, then ask the one essential question

State your inferences explicitly before asking anything. This gives the user a chance to correct you without feeling interrogated.

Example:

"Got it. Sounds like you're putting together a vendor evaluation for a new CRM — I'd set this up as a research project. Folder name: crm-vendor-evaluation.

One question before I build it: what does done look like? Is this a recommendation doc you'll present to someone, a decision you need to make by a deadline, or something else?"

If they already told you what done looks like, skip the question and go straight to Step 4.

Do not ask more than one follow-up question. If a detail is missing (key people, deadline), fill it with a placeholder and note it in the summary. The user can fill it in later — get them set up fast, not perfectly.

Step 4 — Confirmation gate

Before creating any files, show exactly what you're about to do.

"Here's what I'll set up:
projects/crm-vendor-evaluation/
  CLAUDE.md       ← pre-filled with what you told me
  CHANGELOG.md    ← starts empty, grows as you work
  documents/      ← where I'll save files I produce for this project
I'll also add a pointer to this project in your workspace CLAUDE.md so I load this context automatically when you mention it.

Ready to go?"

Do not create any files until the user confirms. If they want to change the folder name, template type, or anything else — update and re-show the summary before proceeding.

Step 5 — Create the files

Use mcp__filesystem__ tools only. Do not use the built-in Write or Edit tools — they use sandbox paths and cannot access the user's filesystem.

Create the following, in this order:

1. projects/[project-name]/CLAUDE.md

Copy the appropriate template from templates/ and fill in what you know:

Replace [Project Name] in the header with the actual project name
Set Last Updated to today's date
Set Status to Active
Fill Goal with what done looks like (from the conversation)
Fill What This Project Is with 2–3 sentences from the conversation
Fill in any Key People that were mentioned
Template-specific:
- Writing: fill Audience if mentioned; leave the rest as placeholders
- Tracking: fill Deadline if mentioned; add any tasks mentioned to Open Items
- Research: fill Question with what they're trying to answer; fill Deliverable if clear
Leave everything else as template placeholders. Do not invent details.

2. projects/[project-name]/CHANGELOG.md

# [Project Name] — Changelog

## [Today's date]
- Project created via new-project wizard

3. projects/[project-name]/documents/README.md

# Documents

Claude-produced files for this project live here — drafts, research,
summaries, exports.

4. Workspace CLAUDE.md — Current Projects section

Add the new project to the next available slot. If all three slots are filled, add it as Project 4 and flag it to the user.

### Project [N]: [Project Name]
**Status**: Active
**Goal**: [What done looks like — from the conversation]
**Context**: Read `projects/[project-name]/CLAUDE.md` for full project context.
**Next Step**: [First concrete action from the conversation, or "Fill in the
project CLAUDE.md with more detail"]

Step 6 — Verify and summarize

After writing each file, read back a key section to confirm the write succeeded. Do not declare success without verifying.

Then give a brief summary:

"Done. Here's what was created:

✅ projects/crm-vendor-evaluation/CLAUDE.md — pre-filled with project goal, research type, and a placeholder for key people ✅ projects/crm-vendor-evaluation/CHANGELOG.md — stub created ✅ projects/crm-vendor-evaluation/documents/ — ready for files ✅ Workspace CLAUDE.md — pointer added to Current Projects

Two things worth filling in when you have a minute:

Key people: Who else is involved or who will see the output?

Sources: Any documents or links I should read for this project?

Ready to get started, or want to fill anything in first?"

Setup & Discovery

This section tells Claude what to say when a user asks how to use or discover this skill.

What to add to the workspace `CLAUDE.md`

When helping a user set up the kit, or when they ask how to invoke this skill, tell them to add the following to the Standing Instructions section of their workspace CLAUDE.md:

### 12. New Project Setup
When I say "new project," "set up a project," or "I want to start something
new," use the new-project skill to guide me through creating the project
folder. Don't just ask what to call it — have a conversation first.

Note: this instruction isn't strictly required — the trigger phrases are sufficient for Claude to recognize the intent. But adding it makes the behavior explicit and ensures it survives across sessions.

What the README references

The README's "Going Deeper" section includes:

"To skip the manual setup, say 'new project' and Claude will walk you through it conversationally."

What SAVE_WORKFLOW references

The workspace structure section includes:

"To create a new project folder, say 'new project' and Claude will set it up for you."

Edge cases

No projects/ folder exists yet: Create it first, then proceed. Note it in the summary: "I created the projects/ folder since it didn't exist."

Current Projects is full (all 3 slots used): Add as Project 4 and note: "You now have 4 projects listed. When one wraps up, move its folder to z-archive/ to keep this manageable."

User gives very little information: If their answer is still vague after one follow-up, default to project-tracking.md (safest general-purpose template), state the assumption, and proceed to the confirmation gate. Do not loop on clarifying questions.

Folder already exists: Ask if they'd like to add a project context file to it instead. Use the same template selection and fill-in logic, writing to the existing folder path.

new-project