Scientific Agent Skills

Pick a task and get started

Each task's installation command selects one complete skill directory at the reviewed commit. Run it in a system terminal inside the research project, then paste the prompt into Claude Code in that same project. Replace bracketed fields in later tasks with actual files or objectives.

1. Install the local data-inspection skill first

   Copy install command

   Run in a new research project. This pins installer 1.7.4 and the reviewed source, selects the complete exploratory-data-analysis directory, and uses copies for Claude Code. Keep project scope, inspect the installation summary, then reload that project.

   npx [email protected] add https://github.com/K-Dense-AI/scientific-agent-skills/tree/5f2bc0341354fd5590aef76a533ada1e1e5b4e62/skills/exploratory-data-analysis -a claude-code --copy

   Read the author's installation and skill documentation

   Confirm the directory, version and actual entry

   Current entries live under skills/<name>/; scientific-skills/<name>/ was the directory before v2.43.0. Preserve scripts, references, assets and other companion files, and inspect each SKILL.md's own version. The collection is 2.72.0, which removed the previously bundled docx, pdf, pptx and xlsx skills. Install those separately through their author's current instructions if needed.

   Read the current layout and version notes

2. Give your AI this first task

   Copy prompt

   These are editorial task examples, not research runs performed by Atlasnote. The small table is synthetic teaching input; other routes require authorized real materials. Expected deliverables are not completed results.

   Use exploratory-data-analysis. Read the installed SKILL.md and general_scientific_formats.md first, and confirm Python 3.11+ is available. Create eda-first-run inside this project. If it already exists, choose a new directory and update --root in both commands. Save the following verbatim as data.csv inside that new directory (eda-first-run/data.csv by default). This is synthetic practice data, not an experiment: each row is one independent example specimen, material is a group label, pressure_kpa is in kPa, NA explicitly means missing, and check is an unresolved nonnumeric entry. sample_id,material,pressure_kpa S01,A,10 S02,A,12 S03,A,NA S04,A,11 S05,B,14 S06,B,15 S07,B,16 S08,B,check From the project root, inspect the format first and proceed only when its tier is automated_core: python3 .claude/skills/exploratory-data-analysis/scripts/capability_manifest.py inspect data.csv --root eda-first-run --output capability.json python3 .claude/skills/exploratory-data-analysis/scripts/tabular_profile.py data.csv --root eda-first-run --missing-token NA --reveal-identifiers --output profile.json The supplied column names are nonsensitive, so revealing them is permitted. Do not print raw rows. Save commands, exit status, scanned scope and the missing-value convention in run-notes.md inside that directory, then write review.md alongside it from the actual JSON. Distinguish numeric, missing and nonnumeric entries. Do not remove rows, impute, modify the original table or run significance tests. If a step fails, preserve the error and explain the missing requirement; never manufacture profile.json from the expected counts.

3. Check what you receive

   • Expect data.csv, capability.json, profile.json, run-notes.md and review.md inside eda-first-run/. The capability record should identify CSV, and the scan should cover these eight rows.
   • Compare against the small input: pressure_kpa should have six numeric entries, one missing entry and one text entry, with inferred type mixed. These are editorial check expectations, not results from an Atlasnote run.
   • For real data, supply units, sampling hierarchy and missing codes. Defaults inspect the first 100,000 rows of a file up to 64 MiB; quantiles can use a further bounded sample. Read row_limit_reached and related flags before generalizing to the whole dataset.
   Read the author’s notes for this task