New Project

Scaffold a new research project

You are creating a new research project repository under projects/.

Finding the workspace root

The workspace root contains CLAUDE.md alongside projects/, pipelines/, $ROOT/data_catalog/, research/. If the current directory is inside a project or pipeline, search upward to find the root. Use $ROOT for all paths below.

Step 1: Identify the project folder and read existing materials

The user has already created a folder under projects/<slug>/ and may have placed materials in it (PDFs, notes, descriptions, data files, drafts, etc.).

1. Infer the slug from the folder name. If the skill was invoked with an argument, use that as the slug. Otherwise, look for recently created folders under projects/ that lack the canonical structure (no docs/ directory).
2. Read every file in the project folder. These are the user's seed materials — project descriptions, research proposals, paper drafts, notes, data documentation, etc.
3. Extract context from these materials:
   • Research question and motivation
   • Current focus / what to work on first
   • Methods and identification strategy
   • Data sources
   • Relevant literature
   • Institutional background
   • Any other information that maps to the docs/ files
4. Check workspace context for connections:
   • $ROOT/research/meta/data_linkages.md and $ROOT/data_catalog/ for existing datasets
   • $ROOT/research/meta/variable_dictionary.md for known variables
   • $ROOT/research/meta/research_map.md for related projects
   • $ROOT/ideas/ for an idea entry that originated this project

Only ask questions if critical information cannot be inferred from the materials and workspace context. If you do need to ask, ask at most one or two targeted questions — not a full interview.

Step 2: Create the directory structure

Follow the canonical structure from $ROOT/research-kit/rules/project_docs_contract.md. Do not overwrite any files the user already placed in the folder. Create only what is missing.

$ROOT/projects/<slug>/
  README.md
  CLAUDE.md
  .claude/
    settings.local.json
  docs/
    summary.md
    thinking.md
    todo.md
    done.md
    meetings.md
    feedback.md
    literature.md
    institutions.md
    data.md
    methods.md
    decisions.md
    findings/
      index.md     # folder-mode findings — the default per project_docs_contract.md §864
  source/
  build/
    .gitkeep
  paper/
    tables/
    figures/
    references.bib
  talk/          # create only if user requests
  references/    # create only if user requests

Step 3: Populate files

Populate docs/ files using context extracted from the user's materials. The more the user provided, the richer the initial documentation should be. Files for which no relevant information was found get their header and empty template sections only.

CLAUDE.md

Follow the pattern of existing projects (see any $ROOT/projects/*/CLAUDE.md for examples):

# <slug>

Read `docs/summary.md` and `docs/todo.md` (if available) before making changes.

**Current focus:** <inferred from materials>

**Identification:** <inferred from materials, or omit if unknown>

**Shared code:** Use `diarios` for court parsing and cleaning. Do not modify `diarios` without asking.

**Data policy:** Never read raw data files. Dataset metadata is in `../../data_catalog/`.

Adapt sections based on what the materials actually contain. Don't include sections that have no content.

.claude/settings.local.json

{
  "permissions": {
    "allow": [
      "Bash(python3:*)",
      "Bash(git add:*)",
      "Bash(git commit:*)"
    ]
  }
}

Add domain-specific WebFetch permissions if the materials mention specific data sources or websites.

docs/summary.md

Populate with research question, motivation, and conceptual framework extracted from the user's materials. Use the format from project_docs_contract.md.

docs/todo.md

# TODOs

## Setup
- [ ] Initial data exploration
  - created: <today's date>

## <topic from current focus>
- [ ] <first task inferred from materials>
  - created: <today's date>

Other docs/ files

• data.md — populate if the materials describe datasets
• methods.md — populate if the materials describe identification strategy or methods
• literature.md — populate if the materials cite papers
• institutions.md — populate if the materials describe institutional context
• thinking.md — place any speculative or exploratory content from the materials here
• findings/index.md — seed the folder-mode findings page even for a new project, so the first finding lands compliant instead of drifting into an ad-hoc flat file. Copy the ## How to read the confidence tags block (🟢/🟡/🔴 Replicated/Single source/Provisional; Strong/Plausible/Speculative) verbatim from project_docs_contract.md §864, then leave an empty ## Findings overview (with ### Empirical findings / ### Interpretations subheads). Individual findings become sibling docs/findings/<slug>.md files with the required four-class Sources footer. Do NOT create a flat docs/findings.md.
• Remaining files: create with header and empty template sections per the contract formats in project_docs_contract.md section 6.

README.md

# <Project Title>

<One-line research question>

See `docs/summary.md` for details.

Step 4: Handle the user's seed materials

After scaffolding, the user's original files are still in the project folder. Ask the user whether to:

• Move them to a subfolder (e.g., references/ or docs/notes/)
• Leave them where they are
• Delete them (now that their content has been absorbed into docs/)

Step 5: Register the project

1. Check $ROOT/research/rules/workspace.md — note the new project in the projects/ listing if the user wants to update it.
2. Check $ROOT/research/meta/research_map.md — suggest where it fits in the project clusters.
3. If the project originated from an idea in $ROOT/ideas/, update the idea's frontmatter (status: project, related_project: $ROOT/projects/<slug>).
4. Suggest creating a project brief in $ROOT/research/project_briefs/.

Step 6: Initialize git

cd $ROOT/projects/<slug> && git init

Create a .gitignore:

build/
*.pyc
__pycache__/
.ipynb_checkpoints/

Gotchas

• Never overwrite files the user placed in the folder. The user's materials are the source of truth.
• Do NOT create optional docs files (theory.md, hypotheses, results.md, etc.) unless the user asks or the materials clearly warrant them. When hypotheses are warranted, create folder mode (docs/hypotheses/index.md + one <slug>.md per hypothesis), never a flat docs/hypotheses.md — folder mode is the default per project_docs_contract.md §5 (same as docs/findings/).
• Do NOT create talk/ or references/ directories unless requested.
• Pipeline repos are different — they have fewer docs files and no paper/. If the user actually wants a pipeline, tell them and use the pipeline structure from the contract.
• Always use today's date (from system context) for created dates.
• Check $ROOT/research/meta/variable_dictionary.md and $ROOT/research/meta/data_linkages.md before populating data.md — the dataset may already be documented.