Ask in plain English — run-on and all. Columnly cleans, reshapes, computes, writes real formulas, and ships hosted analyst briefs and dashboards where every figure traces back to the cell it came from. It never edits your original sheet unless you ask, and every write has full undo.
Before an expensive build, Columnly asks only the questions your prompt left genuinely open — never more than four. Each comes with 3–4 concrete options grounded in your actual data, a recommended one marked, and a free-text escape. Then it shows a plan card to confirm.
It never asks what your prompt already settled or what the data shape forces — a clear dominant figure means it leads with that; a funnel shape means it builds a sequence.
Skip the questions and it proceeds with the recommendations. The questions help; they never get in your way.
The questions are the only stochastic part — all the gating logic is deterministic, so the same inputs always yield the same plan.
Pure conversation is answered directly — no wasted verification. A "clean this sheet" request becomes one verified sub-goal per real problem found, and follows a preserve-never-destroy rule: an unfixable value is left exactly as-is; it never guesses.
Normalize variant values with a completeness guarantee — every distinct value accounted for, nothing guessed, a cell never blanked unless you ask.
Writes real, recalculating Sheets formulas, each cross-checked against an independent computation before it lands. Data-exfiltration functions are blocked.
Pandas-backed pivots, unpivots, crosstabs, group-aggregates, splits, explodes, windows and ranks — chained in one pass through a safe op vocabulary.
Fonts, fills, borders, banding, number/date/accounting formats, plus full conditional-formatting rules on number, text and date predicates.
Columnly's model drafts a structured plan; a pure-Python engine executes it and returns an honest diff of exactly what changed — and it never edits your sheet directly. A separate step applies the verified result. No arbitrary code runs on your sheet, and failures are loud, never silent.
Columnly writes the artifact directly, and five deterministic verifiers gate it: every figure must carry a citation to a stored, machine-computed value and match it exactly. A number that doesn't trace is deleted before ship. It writes two candidates and a skeptical critic — seated as a busy decision-maker — picks the better one and can force a revision.
Analyst brief — a glimpse of a real one. Thesis, a scorecard, the evidence — every figure one tap from its source.
Dashboard. A dense instrument — KPI scorecard, verified figures, the trace one tap away.
The author can only display stored values. If it needs a figure that doesn't exist, it asks the engine to compute it — verified — rather than making one up.
Four dark instrument themes and four light themes, all WCAG-AA contrast-checked, one accent each. A brief is an argument; a dashboard is a dense instrument.
A pure-Python stats engine computes correlations, distributions, trends and segments before the model runs — so analysis is grounded in the whole dataset, not a sampled prompt.
Most real questions live across several tabs. Columnly is learning to read the entire workbook and reconcile it into one answer — checking before it combines, so a cross-tab number is reconciled, never silently merged.
Orders here, costs there, a legend over there — stitched into a single brief or dashboard.
Pivot one dimension against another — counts, sums, shares — into a clean, labelled grid from a plain-English ask.
Point at a cross-tab formula and ask "what does this do?" — it explains it and traces every input back to its tab.
Add Columnly to Google Sheets and get an analyst that proves its work — right where your data already lives.
