AI Task–Skill Atlas — public task-skill network

Generated 2026-05-05 from the Atlas public-data build.

FILES
  nodes.csv    28388 rows. One per node (task or skill).
  edges.csv    56878 rows. One per retained task-skill edge.
  README.txt   this file.

NODES.CSV COLUMNS
  slug            stable url slug used throughout the Atlas.
  type            "task" or "skill".
  name            display name.
  degree          number of incident edges in the retained backbone.
  cluster_id      0..8. Spatial territory assignment from
                  k-means on the 2D UMAP layout.
  cluster_label   short editorial label for the territory.
  x, y            2D UMAP coordinates (OpenAI text embeddings → UMAP).

EDGES.CSV COLUMNS
  task_slug             slug of the task endpoint.
  skill_slug            slug of the skill endpoint.
  importance_points     integer 0-100. How central the skill is to the task.
  importance_role       one of:
                          judgment_analysis
                          physical_execution
                          communication_coordination
                          information_processing
                        The role the skill plays in the task.
  bottleneck_points     integer 0-100. How hard this skill is to replace in
                        the task (higher = more bottleneck-like).
  replaceability_type   one of:
                          easily_replaceable
                          partly_replaceable
                          hard_to_replace

  Blank cells mean the weight was not measured for that edge.
  56878/56878 edges carry importance + bottleneck weights.

PROVENANCE
  The retained backbone of 56878 edges comes from an LLM-adjudicated
  prune of candidate task-skill pairs (majority retain rule across three
  independent votes). The importance and bottleneck points are the
  skill-conditioned weighting runs from the Atlas extension pipeline,
  aggregated across three runs per edge.

  This is one of two directions the weighting was run in. The public graph
  uses the skill-conditioned direction because its candidate pool covers the
  full retained backbone; the task-conditioned direction is used as a
  robustness check in the paper.

SCOPE
  The Atlas measures task-level exposure to automation and the skill
  structure underneath it. It does not measure realised adoption, wages,
  or displacement. Treat this file as a descriptive map of what the model
  says about task-skill structure, not a forecast.

CITATION
  If you use these files, please cite the Atlas paper (see
  https://github.com/prashgarg/AutomationAtlas for the current reference)
  and link back to https://automationatlas.org/task-skill-graphs/.

LICENSE
  Data released under CC BY 4.0. You may share and adapt with attribution.
