Run a notebook data analysis with Claude Code's NotebookEdit without losing outputs

You are working in analysis/prompt_engagement_by_category.ipynb through the NotebookEdit tool, editing at the cell level, not regenerating the notebook file as a whole. A notebook's cell outputs, especially ones that took a long time to compute, are themselves valuable artifacts — treat re-running a cell that does not need to change as a real cost, not a free action.

ANALYSIS GOAL
Whether prompts with an exampleOutput field get meaningfully more time-on-page than prompts without one, across all tier-1 categories.

DATASET
Page-analytics export covering the last 60 days; sample size for some smaller categories is under 200 sessions, too small to treat a small difference as significant.

WHICH CELLS ARE EXPENSIVE TO RE-RUN
Cell 3, which joins the full analytics export against the prompt-library metadata and takes about four minutes to run.

WHEN AN OUTPUT MUST BE PRESERVED VERSUS RECOMPUTED
Any cell whose output depends on cell 3’s join must be re-run if cell 3 changes; cells that only reformat an already-computed result do not need cell 3 to change to stay valid.

WORKING RULES
1. Before editing any cell listed in Cell 3, which joins the full analytics export against the prompt-library metadata and takes about four minutes to run., confirm whether this specific change actually alters that cell's inputs or logic. If it does not, do not re-run it — the existing output is still valid evidence and re-running it purely out of habit wastes real time or compute for no new information.
2. Before writing and running a new exploratory cell, state the specific hypothesis it is testing — what result would confirm it, what result would disconfirm it — before executing it, not after looking at the output and narrating a story that fits whatever appeared. A cell run with no stated expectation beforehand invites explaining away a result that does not actually support the analysis's conclusion.
3. When adding a new cell, insert it at the position where it belongs in the analysis narrative, not appended at the end of the notebook regardless of where its result is actually relevant — a notebook read top to bottom should tell a coherent story, and cells scattered out of narrative order defeat that for the next person who opens this file.
4. If a change to an early cell would invalidate the output of a later cell that depends on it, per Any cell whose output depends on cell 3’s join must be re-run if cell 3 changes; cells that only reformat an already-computed result do not need cell 3 to change to stay valid., say so explicitly and either re-run the dependent cell or flag its now-stale output rather than leaving a stale result sitting uncorrected next to a change that quietly invalidated it.

FINAL CELL
End the notebook with a markdown cell summarizing the actual findings in plain language, citing the specific cells whose output supports each claim — a finding stated without pointing back to the cell that produced it is not distinguishable from a guess to anyone reviewing the notebook later.

CONSTRAINTS
- Do not silently delete an existing cell's output when editing that cell's code, unless the edit means the old output no longer corresponds to anything real; a notebook's history of what was actually tried is itself useful context for whoever reviews this analysis next.
- Cite Page-analytics export covering the last 60 days; sample size for some smaller categories is under 200 sessions, too small to treat a small difference as significant.'s actual known limitations when interpreting a result, rather than presenting a finding as more conclusive than the data underlying it can actually support.

OUTPUT
The notebook edits made, cell by cell, and the final summary cell's content.