annotate-findings

Skill: Annotate Findings

Take the notebook at $ARGUMENTS (which must already have been executed) and replace every placeholder observation cell with real findings derived from the executed outputs of the surrounding code cells.

This skill closes the loop opened by /analyze-dataset (which writes placeholders) and /execute-notebook (which produces real outputs). It eliminates the most common failure of generated notebooks: markdown that claims results that the code never actually proved.

Principle

A finding is only valid if it traces back to an executed output. This skill never invents numbers. If the preceding code cell has no output, or the output cannot be interpreted (binary blobs, image data only), the placeholder gets a clear note saying so — never fabrication.

Step 1 — Validate input

1. Verify the notebook file exists.
2. Verify it has been executed: at least one code cell must have non-empty outputs and a non-null execution_count. If the notebook looks pristine (no executions), fail with: Notebook does not appear to have been executed. Run /execute-notebook first.
3. Locate all markdown cells containing the literal marker <!-- annotate-findings: pending -->. If none are found, report No placeholder cells found — nothing to annotate. and exit cleanly.

Step 2 — For each placeholder, gather context

For each placeholder markdown cell at index i:

1. Extract the "Qué buscar" bullets from the placeholder (they list what to inspect — use them as the question prompt).
2. Walk backwards from cell i collecting code cells until you hit either a section header (## N.) or another placeholder. These are the code cells whose outputs feed this observation.
3. From each gathered code cell, extract:
   • The cell's source (so the skill can read what was computed).
   • All outputs of types stream (stdout/stderr text), execute_result (text/plain representation of last expression), and display_data (descriptions of figures — read the text/plain fallback if present, otherwise note "figure produced").
4. Concatenate sources + outputs into a context block per placeholder.