Quickplot

quickplot

Turn raw numbers into a compact unicode sparkline histogram + box-and-whisker plot, plus a full numeric summary, printed in the terminal. For quick-and-dirty "what does this distribution look like" checks — real/paper plots happen elsewhere.

Group name on the left, graphics (histogram over box-&-whisker) on the right, all on a shared fixed x-axis; the full numeric summary is a column-aligned list below. The histogram is 2 character rows tall by default (16 levels of vertical resolution, linear — bar heights stay proportional), which makes low-count tail bins visible without distorting anything; --rows 1 drops to a compact single row.

200 samples each

Normal                            ▃ ▂█
                     ▁  ▁▂▁▃▄▆▆█▆██████▄█▅▁▂▅ ▃ ▁▁ ▁         <- histogram (2 rows)
                     ├──────────[==┃==]───────────┤          <- box & whisker
Exponential  █
             █▃▆█▅▅▄▃▄▄▃▃▁▂▁▁▂  ▁▁ ▁      ▁
             ├─[==┃===]─────────────────────┤
Uniform       ▃ ▁ ▄▃ ▁   ▃   ▆   ▁▃ ▁▃  ▁█        ▁ ▃▁▁ ▁▃▄
             ▅█▇█▄██▅█▇▇▇█▄▄▇█▇▅▅██▇██▇▇██▅▅▇▅▂▇▄▅█▄███▅███
             ├──────────[===========┃==========]──────────┤
             └──────────┴──────────┴───────────┴──────────┴  <- shared axis (once)
             0.16       24.6       49          75.6     100

Normal:      n=200  min 18.6  q1 41.5  med 49.9  q3 56  max 82.6  mean 49.5  sd 11.2
Exponential: n=200  min 0.16  q1 4.31  med 10.8  q3 21  max 69.9  mean 14.5  sd 13.6
Uniform:     n=200  min 1.22  q1 24.6  med 50.1  q3 75  max  100  mean 50.2  sd 29.5

Box & whisker reads ├ min · [ q1 · ┃ median · ] q3 · ┤ max. Stacking several groups lets you eyeball-compare spreads since every box sits on the same axis.

The contract — why this is a script, not eyeballing

plot.py is the only thing allowed to turn the data into a picture or a summary. It computes the histogram, the box plot, and the full seven-number summary itself. You run it and paste its output verbatim.

Do NOT, separately, describe the distribution in prose ("looks roughly normal", "Olmo is worse", "tightly clustered"), round the numbers, or report a subset of them. The user explicitly does not want that layer — the script's output is the answer. The plot is only a visual aid; the authoritative numbers are the text summary line, which prints n, min, q1, median, q3, max, mean, and std for every group. Unparseable values are counted and flagged (⚠ N unparseable), never silently dropped. If they ask a follow-up about the numbers, answer from that printed summary.

How to run

Self-contained uv script (only dep: numpy) — nothing to install. Pass a file or pipe via stdin:

~/.claude/skills/quickplot/plot.py data.json
echo '{"Gemma":[...],"Olmo":[...],"Qwen":[...]}' | ~/.claude/skills/quickplot/plot.py --title coherence
cat scores.csv | ~/.claude/skills/quickplot/plot.py

Typical flow: the data already exists (a list in the conversation, a column in a CSV/parquet, an array in a script). Get it into one of the input formats below — usually by writing a small JSON object to /tmp or piping — then run the script and show what it prints.

Input formats (auto-detected)

| Input | Result | | ---------------------------------------- | -------------------- | | JSON object {"Gemma":[..],"Olmo":[..]} | one panel per key | | JSON array [1,2,3,...] | a single panel | | CSV/TSV with a header row of names | one panel per column | | plain whitespace/newline numbers | a single panel |

For the common "models × metrics" case (Gemma/Olmo/Qwen each scored on coherence + alignment), the cleanest input is a flat JSON object with one key per group, e.g. {"Gemma coherence":[...], "Gemma alignment":[...], ...}. All groups share the same fixed x-range so the panels are directly comparable.

Flags

• --width N — plot width in characters (default 46)
• --rows N — histogram height in character rows (default 2; 1 = compact)
• --title TEXT — title printed once above the panels
• --summary-only — print just the numeric summary lines, no plot

Out of scope

Scatter, line, or time-series plots, and anything headed for a paper or slide deck. This is terminal distribution-at-a-glance only.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: beyarkay
• Source: beyarkay/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.