Color Summary Slide

Color-Coded Summary Slide

Overview

Use this skill to produce short, dense, color-coded summary slides (typically 1–4 slides) that explain a reasoning process — most often a clinical diagnostic process: how a diagnosis was reached from presentation → imaging → labs/CSF → differential → conclusion. The signature move is that color encodes the type of information, not decoration, so a reader can scan a slide and immediately tell clinical facts from imaging from lab data from "ruled out" from "confirmed".

This is different from /journal-reading and /case-report (full decks). Reach for this skill when the user wants a tight visual summary of a specific question ("summarise how we diagnosed X", "make a 2-page slide on the workup", "add a slide with the MRI and the findings").

Triggers: "色彩總結 / 顏色分類 slide", "diagnostic process slide", "summarise the workup", "differential diagnosis slide", "compact summary slide", "add the imaging screenshot with findings".

Prerequisites

pip3 install python-pptx pillow pymupdf

Design system (the rules)

1. Color = information type. Assign each information category a semantic key and keep it stable across every slide. Defaults in scripts/colorsummary.py:

   key	color	use for
   clinical	amber	symptoms, history, triggers, teaching points
   imaging	blue	CT / MRI / X-ray / imaging findings
   csf / lab	purple	CSF, lab panels, fluid data
   infection / danger	red	infection, "negative / excluded", red flags
   tumor / good	green	tumor/positive evidence, normal values, "supports Dx"
   diagnosis	navy	conclusions, final diagnosis, takeaways

   Rename/extend the TYPES dict per domain, but never use a color for two meanings in one deck.

2. Always show a legend on the first slide (deck.legend) so the color mapping is explicit.

3. Compact + large font. Prefer few words, big type. Body 13–15 pt, values 14–16 pt bold, headers 26–28 pt. Cards taller, whitespace tighter.

4. Components (all in scripts/colorsummary.py):

   • header(title, sub, page) — navy bar + accent underline + page number.
   • legend(keys) — color→type key.
   • banner(kind, runs) — one accented strip (e.g. the clinical "trigger").
   • timeline(steps) — step cards with colored headers + arrows between them.
   • kv_panel(title, kind, rows) — key/value data table with flag chips (e.g. ↑, ↑↑, normal, "malig (–)"); ideal for CSF/lab panels.
   • card(kind, title, body, solid=False) — an interpretation card; solid fills the accent (use for the final diagnosis / conclusion).
   • takeaway(runs) — navy bottom bar for the one-line "so what".
   • image(path, caption) — de-identified image panel with border; returns the right edge x so you can place annotation cards beside it.

5. Rich text format used by every component:

   runs = [ line, line, ... ]              # one paragraph per line
   line = (text, size, bold, color)        # single-run line, OR
   line = [ (text, size, bold, color), ... ]  # multi-run line
   # color = a TYPES/COLORS key string ("imaging","tumor","dark","white","sub","mint") or an RGBColor
   

De-identification (REQUIRED for images)

Clinical screenshots (PACS/EMR) almost always contain PHI — patient name, full chart number, DOB — in the top bar and side panels. Before placing any image on a slide:

1. Open it and locate identifier regions (usually the top strip and the left study-list panel).
2. Crop to just the region of interest (e.g. the brain) so identifiers are physically removed — do not merely cover them.
3. Re-open the crop and confirm no text identifiers remain.

from PIL import Image
im = Image.open("pacs_screenshot.png").convert("RGB")
im.crop((left, top, right, bottom)).save("crop_deid.png")   # brain only

Never version-control the output or the source images (they contain PHI) — see "Output" below.

Workflow

1. Gather the narrative

Read the source records (progress notes, imaging reports, lab/CSF data). Pull out, per step: the date/tag, the finding, and which type it is (clinical / imaging / csf / infection-negative / tumor-positive / diagnosis).

2. Plan the slides

Typical shapes (mix as needed):

• Timeline slide — presentation → workup steps → diagnosis set.
• Data + differential slide — kv_panel (CSF/labs) on the left; stacked cards on the right (supports-Dx / against-alternative / final diagnosis).
• Image slide — de-identified image on the left; findings + conclusion cards on the right.

3. Generate

Copy the example and adapt the content dicts:

cp .claude/skills/color-summary-slide/scripts/example_diagnostic_process.py /tmp/build.py
# edit the content dicts, then:
python3 /tmp/build.py path/to/deid_image.png /path/to/out.pptx

Or import the library directly:

import sys; sys.path.insert(0, ".claude/skills/color-summary-slide/scripts")
from colorsummary import Deck
deck = Deck()
s = deck.header("...", "...", page="1 / 3")
deck.legend(s, ["clinical", "imaging", "csf", "infection", "tumor"])
# ... banner / timeline / kv_panel / card / takeaway / image ...
deck.save("out.pptx")

4. Verify layout

This environment cannot always render PPTX to preview. Check programmatically:

from pptx import Presentation
from pptx.util import Emu
p = Presentation("out.pptx"); SW, SH = p.slide_width, p.slide_height
for si, s in enumerate(p.slides, 1):
    for sh in s.shapes:
        if sh.left is None: continue
        r, b = sh.left + (sh.width or 0), sh.top + (sh.height or 0)
        if r > SW + 9144 or b > SH + 9144:
            print("overflow slide", si, sh.shape_type)

No output = nothing runs off the slide. If a text box overflows, shorten the text or enlarge the card (compactness comes from fewer words, not smaller font).

Output

• Save the .pptx to the working directory (or the relevant {date} JR/ folder).
• PHI: if the content contains patient data, do not commit it to git — deliver the file directly. The repo's .gitignore already excludes *.pptx, *.pdf, and *.docx.

Files

• scripts/colorsummary.py — the helper library (Deck + components + palette).
• scripts/example_diagnostic_process.py — runnable 3-slide worked example (timeline → CSF panel + differential → image). Run it to see the output, then copy and adapt.