Init Project

Skill: init-project

What it does. Bootstraps a new book from the p10t template: interviews the author for the essentials, populates project.yaml, resets project-specific files to blank templates, and verifies the generic machinery is intact. The first skill run in any new project.

Triggers

• "init the project"
• "/init-project"
• "set up this book"
• "start a new book here"
• Equivalent phrasing in any language

Input. None — everything is gathered by interview.

Output. A configured .project/, a filled project.yaml, an adjusted root CLAUDE.md, and a suggested first-steps plan.

Core principle

Ask only what cannot be defaulted. Configure everything the answers imply.

A bootstrap that asks twenty questions never gets finished. The interview is seven questions or fewer; everything else is derived or left at sensible defaults, explicitly listed at the end so the author knows what to adjust later.

Execution protocol

Step 1 — Detect the starting state

Check what exists:

• Fresh clone (config files still contain {placeholders}) → full bootstrap
• Existing manuscript present (files in manuscript/ or elsewhere) → bootstrap + offer the extraction path (Step 5)
• Already initialized (project.yaml has a real title) → stop; offer to update settings instead of re-initializing. Never reset an active project.

Step 2 — The interview

Seven questions, one message:

1. Title (working title is fine)
2. Output language — the language of the manuscript and of all generated output. If it is not pt-BR or en, say so plainly: the framework's detection signals are calibrated for those two, and other languages inherit the definitions, ceilings, and treatments but need their signals adapted. Note it in style-guide.md → Density ceilings → Language adaptation so analyses stay comparable
3. Genre and audience (one line)
4. Structure — chapters grouped in acts/parts? Interludes? Rough planned count?
5. Manuscript location and layout — use manuscript/ or an existing folder; and flat (every chapter file in one directory) or chapter (one directory per chapter, named the chapter id)? Default flat, and say why: it is right below ~15 chapters, and migrating later is a mechanical move because the chapter filename never changes. Recommend chapter when the planned count from Q4 is large and the author expects outlines and analyses per chapter. See .project/templates/layout.md
6. Analyses next to chapters, or centralized in .project/reports/technical/?
7. AI-use stance — the declaration line for project.yaml (offer the default: "human curation item by item")

Step 3 — Apply

• Fill project.yaml completely — including derived fields (naming pattern from the structure answer, paths.manuscript/paths.layout/paths.analyses from Q5–6). Zero-pad the numbers in paths.naming: ordering is lexicographic, so unpadded ids sort 10 before 2
• If an existing manuscript was found in Step 1, check it against the declared layout and report any mismatch — never move the author's files during init
• Update the root CLAUDE.md: title, structure section, paths
• Verify project-specific files are blank templates; if this is a copy from a previous book, list any file still carrying old content (previous persona, old preserve list) and confirm before clearing each — never silently wipe
• Confirm generic machinery intact: .claude/skills/, .project/templates/
• Offer to remove the template's own artifacts, which belong to p10t rather than to this book: examples/, manuscript/README.md, and this repo's CHANGELOG.md. Offer, do not delete — some authors keep examples/ as a reference

Step 4 — Report the configuration

Show what was set, what was defaulted, and where to change each thing later. One screen.

Step 5 — Route to the right first step

The fork that matters:

Existing manuscript → suggest the extraction sequence:

1. define-persona (bootstrap mode — point it at the chapters you consider most yours)
2. build-worldbuilding + create-character (extract mode)
3. analyze-chapter on the first chapter → the revision cycle begins

Blank page → suggest the creation sequence:

1. define-persona (bootstrap mode — bring earlier writing produced without AI, the most valuable corpus)
2. define-references
3. build-worldbuilding + create-character (create mode)
4. outline-chapter → draft-scene

Either way, close with: the system starts weak and compounds. The first chapter's cycle teaches the next one.

What to avoid

Re-initializing an active project. The one destructive mistake. Detection first, always.

Silent wiping. Leftover content from a previous book is listed and confirmed file by file.

Interview sprawl. Seven questions. Anything else has a default and a note.

Configuring beyond knowledge. Word-count targets, phase tracking, density targets — leave at defaults; they calibrate during use.

Relationship to other skills