Step-by-step workflow for scaffolding a new Dunn Lab project from scratch. Use whenever starting a new analysis, pipeline, tool, or package — including when the user says "new project", "set up a repo", "start a new analysis", "initialize a project", or asks to create project structure, directory layout, or boilerplate for a new codebase. Also trigger when the user describes a new analysis goal (e.g., "I need to analyze some RNA-seq data", "let's build a phylogenetics pipeline") and the working directory is empty or lacks project structure (no README.md, no CLAUDE.md, no src/ or scripts/ directory).
Step-by-step workflow for scaffolding a new Dunn Lab project from scratch. Use whenever starting a new analysis, pipeline, tool, or package — including when the user says "new project", "set up a repo", "start a new analysis", "initialize a project", or asks to create project structure, directory layout, or boilerplate for a new codebase. Also trigger when the user describes a new analysis goal (e.g., "I need to analyze some RNA-seq data", "let's build a phylogenetics pipeline") and the working directory is empty or lacks project structure (no README.md, no CLAUDE.md, no src/ or scripts/ directory).
Dunn Lab New Project Setup
Follow these steps when starting a new project from scratch. This skill references conventions from the dunnlab-defaults skill — apply those standards throughout.
Progress tracking
This skill persists its progress to .claude/new-project-progress.yaml so it can resume after /clear or a new session.
On first invocation (no progress file exists)
Inspect the repo to detect which steps have already been completed. Use these heuristics to mark steps as completed or skipped in the initial task list:
.git/ exists → skip git init in Step 2
.claude/settings.json exists → skip permissions setup in Step 2
.gitignore exists → skip initial .gitignore in Step 2 (still update it in Step 6)
.devcontainer/ exists → skip Step 3
README.md exists → mark as pre-existing in Step 4 notes (still review/update it)
CLAUDE.md exists → same as README
dev_docs/ exists → same as README
environment.yml, renv.lock, or Cargo.toml exists → skip Step 7
Build the initial task list and write it to .claude/new-project-progress.yaml with this format:
# Dunn Lab new-project progress — do not edit manuallyupdated:2024-01-15T10:30:00devcontainer:true# whether user opted for devcontainerin_devcontainer:false# whether currently running inside a devcontainertasks:-id:step-1name:Definetheprojectscopestatus:completed# completed | in-progress | pending | skippednotes:"Python RNA-seq pipeline"-id:step-2name:Initializerepoandconfigurepermissionsstatus:in-progressnotes:""# ... remaining steps
Load the task list into TodoWrite so it is visible during the session.
On subsequent invocations (progress file exists)
Read .claude/new-project-progress.yaml.
If devcontainer: true and in_devcontainer: false, check whether you're now inside a container (e.g., /.dockerenv exists or $REMOTE_CONTAINERS is set). If so, update in_devcontainer: true in the progress file — this confirms the user successfully reopened in the devcontainer and you can continue to Phase 2.
Load it into TodoWrite.
Resume from the first task that is not completed or skipped.
Keeping progress up to date
When you finish a step, mark it completed in both TodoWrite and the progress file.
When you start a step, mark it in-progress.
If the plan changes (steps added, removed, reordered, or revised based on user feedback), update the progress file to reflect the new plan. The progress file is the durable source of truth; TodoWrite is the in-session view.
Use the notes field to capture key decisions (e.g., chosen language, project type) so they survive across sessions.
Phase 1: Bootstrap
The goal of this phase is to get the repo initialized and permissions configured as quickly as possible so that subsequent steps can run with minimal user intervention.
Step 1: Define the project scope
Before writing any code, understand what the user wants to build. Start by asking about the goal — what scientific question or engineering problem are they solving? Then follow up conversationally based on their answer to fill in any gaps:
Language: Default to Python unless the user mentions R-specific packages (e.g., Seurat, DESeq2) or performance needs that suggest Rust. If the answer is obvious from context, confirm rather than ask.
Inputs and outputs: What data goes in, what results come out?
Project type: One-off analysis, reusable tool, or package? Infer from context when possible (e.g., "analyze some RNA-seq data" → analysis; "build a CLI" → tool).
Devcontainer: Ask whether they'd like one. Briefly explain that devcontainers provide a reproducible, isolated environment via Docker and VS Code, but are optional. Record the choice in the progress file.
If the user gives a brief prompt (e.g., "new project for phylogenomics"), infer sensible defaults (Python, analysis, no devcontainer) and confirm them in a single message rather than asking each question individually.
If there are problems with the user's plans (e.g., they suggest inappropriate tools, there are missing or unnecessary steps, there are better approaches, the data can't be used for this purpose, etc.), point out the issues and suggest better approaches.
Use the answers to guide decisions in the following steps.
Step 2: Initialize repo and configure permissions
Do these in order — settings.json first so all subsequent tool calls benefit from the permissions:
Initialize git with git init (skip if already initialized).
Create .claude/settings.json with reasonable permissions for local development. Use acceptEdits as the default mode so file edits don't require individual approval — this lets Claude work fluidly for reading and running code while still requiring confirmation for file modifications, package installs, and git mutations. Read references/settings-permissions.md for the full permission rules and JSON format, then generate the settings file.
Create a minimal .gitignore with .DS_Store and other common ignores. This will be expanded in a later step once the language and project type are known.
Step 3: Set up devcontainer (optional)
If the user opted for a devcontainer in Step 1:
Use the dunnlab-devcontainer skill to add a .devcontainer/ directory with the standard Claude Code devcontainer configuration.
Commit all bootstrap files (settings.json, .gitignore, .devcontainer/).
Tell the user to reopen the project in the devcontainer now — all remaining work should happen inside it. Update the progress file with in_devcontainer: false so that on the next invocation it can detect the transition.
Stop here. The user will start a new session inside the devcontainer and resume with Step 4.
If the user did not opt for a devcontainer:
Commit the bootstrap files (settings.json, .gitignore).
Continue directly to Step 4.
Phase 2: Plan
Step 4: Create project planning documentation
Before writing any code, create the following:
README.md with a project title and placeholder sections for the overview, setup instructions, usage examples, and development notes. If devcontainer: true in the progress file, include a "Development container" section explaining how to use it (install Docker and the VS Code Dev Containers extension, then reopen the project in the container).
dev_docs/overview.md outlining the scientific question or engineering goal, key data sources and their formats, and the planned analysis workflow or architecture. This serves as a reference that can be loaded into context by Claude Code when working on relevant parts of the project.
Additional dev_docs/ files as needed to document the project (e.g., data model, analysis workflow, interpretation notes). Keep each file atomic and focused on a single topic — this way Claude Code can load only the relevant file into context rather than pulling in the entire project's documentation, which helps stay within the context window on larger projects.
CLAUDE.md with a brief project summary, links to the above documentation, and any project-specific instructions for using Claude Code. Also specify to use the dunnlab-defaults skill for coding conventions and project structure.
Ask the user any clarifying questions needed to fill in these documents.
Step 5: Review and finalize the plan
Once the planning documents are drafted:
Review the plan to make sure there aren't better options for the project structure, environment, or documentation based on the project scope and goals. This is a good time to catch any potential issues before scaffolding the project.
Prompt the user to ask if they would like to make any changes. If they want changes, help them iterate until they're satisfied.
Once the user confirms the plan is finalized, commit the planning documents to git.
Phase 3: Scaffold and build
Step 6: Create directory structure and update .gitignore
Scaffold the project following the dunnlab-defaults project structure. Always prefer idiomatic structures for the language and project type.
Update .gitignore as appropriate for the languages in the project. Always exclude:
Add language-specific ignores (e.g., target/ for Rust; .Rhistory, .RData, .Rproj.user/ for R).
Scaffold test infrastructure alongside the directory structure:
Python: Create tests/ with an empty conftest.py and a placeholder test file.
R: Set up tests/testthat/ with a testthat.R runner and a placeholder test file.
Rust: The default cargo init includes a test module; add a tests/ directory for integration tests if appropriate.
Step 7: Set up the environment
Python: Create environment.yml with the project name, Python version, and initial dependencies. If running inside a devcontainer, conda/mamba is already installed — create the environment directly. Otherwise, ensure conda or mamba is available first. Run conda env create -f environment.yml or mamba env create -f environment.yml. If creation fails (usually dependency conflicts), read the error, adjust versions or channels in environment.yml, and retry — don't skip environment setup.
R: Initialize renv with renv::init(). Install initial packages and snapshot with renv::snapshot(). If the project needs Bioconductor packages, configure the Bioconductor repository in renv before installing them. If renv::restore() fails on a package, check whether it needs a system library (common with spatial/genomics packages) and note the dependency in README.md.
Rust: Cargo.toml is created by cargo init. Add dependencies as needed.
Include instructions for environment setup in README.md.
Step 8: Enter development mode
Read dev_docs/overview.md and the project scope notes from the progress file, then break development into atomic tasks tailored to the project type. The decomposition depends on what's being built:
Analysis/pipeline: first task gets a minimal end-to-end pipeline running (read input → stub processing → write output), then subsequent tasks fill in each processing step. Aim for 3–6 tasks for a typical pipeline — one per major processing stage.
CLI tool: first task sets up argument parsing and the entry point, then subsequent tasks implement each subcommand or feature. Aim for one task per subcommand plus one for integration tests.
Package/library: first task defines the public API with stub implementations and a test file, then subsequent tasks implement each function. Aim for one task per public function or logical group of functions.
Present the proposed task breakdown to the user for review before starting implementation — they may want to reorder, merge, or split tasks based on their priorities.
Write the finalized tasks into the progress file and TodoWrite. Then implement them one by one. After each task:
Run tests to verify functionality.
Run linters and formatters to maintain code quality.
Update documentation to reflect new functionality or changes.
Commit changes with descriptive messages.
Do not move on to the next task until the current one is fully implemented, tested, and documented. Each task should be small enough to complete in a single session — this prevents context overload and keeps diffs reviewable.
After completing each task, commit your changes and then run /clear before starting the next task. When the user re-invokes this skill after clearing, it will automatically resume from the progress file — no need to start over.
Step 9: Final verification
Run through this checklist when wrapping up. For each item, actually run the relevant command rather than just eyeballing it:
Environment from scratch: delete and recreate the environment from the config file (environment.yml, renv.lock, or Cargo.toml) to confirm it builds cleanly
Starter script runs: execute the main entry point or pipeline with sample/test input and verify it completes without errors
Tests pass: run the full test suite (pytest, cargo test, testthat, etc.)
Linters and formatters clean: run the project's linter and formatter (black --check . && flake8 ., cargo clippy, etc.) and fix any issues
Code review: look over the project for performance issues, security concerns, or potential bugs — if a refactor is needed, break it into a new task and implement it before moving on
README accurate: follow the setup instructions in README.md as if you were a new user — do they actually work?
CLAUDE.md and dev_docs/ current: verify these files reflect the final state of the project, not the initial plan
