New Project

Skill: new-project

Stand up a reproducible project skeleton in one pass: a DataLad dataset (Tracking + Modularity), a plain BIDS layout (Self-containment), and a pinned container recipe (Portability + Ephemerality). You own the what/why; you delegate every DataLad operation to the datalad doer subagent.

When to use

• The user is starting a new study / analysis project and has no dataset yet.
• Do NOT use to add an analysis to an existing project — that is analyze/propose-comparison.

Steps

1. Gather project info (ask; keep it short):

• short_name (kebab-case, used as the directory name)
• one-line study description
• primary language/stack (Python / R) — for the container recipe
• (optional) expected preprocessing pipelines (fMRIPrep, MRIQC, …) — recorded as notes only

in v1 (Nipoppy wiring is deferred).

1. Create the YODA dataset — delegate to the datalad doer:

> "Create a YODA DataLad dataset at ./ with the text-to-git procedure > (datalad create -c text2git -c yoda --description '')." YODA gives code/, inputs/, outputs/, README.md. The text2git procedure is important: under YODA's default .gitattributes everything except code/** and *.md/*.txt is annexed — which turns project.yaml and BIDS metadata (dataset_description.json, *.tsv, *.json sidecars) into read-only annex symlinks the planner cannot append to. text2git keeps text files in git (writable + diffable) while binary data (NIfTI, etc.) still annexes. Confirm .datalad/ exists before continuing.

1. Overlay a plain BIDS skeleton at the dataset root (write these yourself; they are small

text files that YODA stores in git):

• dataset_description.json — `{"Name": "", "BIDSVersion": "1.9.0",

"DatasetType": "raw"}`

• participants.tsv — header row participant_id (+ any known columns)
• README — study summary + how the dataset is organized
• .bidsignore — ignore non-BIDS helpers (code/, outputs/, containers/)
• derivatives/ — where BIDS-derivative outputs land (keep a .gitkeep)

Raw BIDS data lives at the dataset root; outputs//derivatives/ hold results; inputs/ holds linked source subdatasets (YODA P1). Do not copy raw data in by hand.

1. Scaffold a basic analysis container recipe in containers/ (recipe only; building/

registration happens on the first analyze/run-comparison):

• Python → environment.yml (pin interpreter + core scientific stack) and a minimal

Dockerfile/Apptainer.def built from it.

• R → an renv.lock stub + equivalent container def.

Note in the README that analyses run via datalad container-run (Portability + Ephemerality).

1. Initialize the project ledger — write project.yaml at the dataset root using the shape

below, with one new-project log entry and empty products: / obligations: lists (later skills populate them). Follow the conventions in docs/project-ledger.md; the structure is validated by schemas/project.schema.json.

1. Save everything — delegate to the datalad doer:

> "Save the new-project scaffold: datalad save -m 'scaffold YODA+BIDS project '."

1. Report — dataset path, what was created, and the suggested next step

(analyze/propose-comparison, or link input data with the datalad doer).

project.yaml shape (the ledger)

Validated by schemas/project.schema.json; conventions in docs/project-ledger.md. new-project writes the header + first log entry and seeds empty products/obligations (populated later by analyze/manage-product, disseminate/*, and govern/*).

project:
  name: 
  description: 
  created: 2026-07-10T14:30:00Z
  dataset_root: .
  stack: python            # python | R | other
products: []               # named deliverables — grouped from kept comparisons later
obligations: []            # Manage & Comply commitments (pre-registrations, DMP/ethics/funder)
log:
  # append-only; never rewrite or reorder prior entries
  - { ts: 2026-07-10T14:30:00Z, op: new-project, stage: initialize,
      note: "scaffolded YODA+BIDS dataset + container recipe", branch: main }

Constraints

• Never run DataLad commands yourself — delegate creation and save to the datalad doer.
• Keep project.yaml append-only and schema-valid (schemas/project.schema.json). The

new-project entry is the first log line; products/obligations start empty.

• BIDS files are the only thing this skill writes directly; everything provenance-related goes

through the doer so the very first commit is properly tracked.

• Container building is deferred to first run — scaffold the recipe, don't build it here.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: neuroanalytics
• Source: neuroanalytics/data-science-harness
• License: MIT

Install and usage instructions live in the source repository linked above.