slide-fit-check

Slide Fit Check

Typst does not error when content overflows a slide. The element bboxes still exist in the PDF page stream past the page boundary (page.rect.y1), so they are detectable with pymupdf (a.k.a. fitz) — but only after the PDF is built. This skill builds the affected session, locates the edited content-block(s) in the PDF, and reports any overflow.

When to run

• After editing a content-block's summary: in a slide-emitting module (modules/Session_*/*.typ). The hook will remind you.
• After adding an image(), large grid(), extra bullets, or a wider-aspect asset to existing summary content.
• When the user reports "spillover" or "content cut off" on a slide.

When to skip

• Edit only touches a details: block (document-only — never on a slide).
• Edit is inside a content-block-doc-only(...) (no slide rendered).
• Edit is a typo fix or other change that cannot affect layout (e.g., tweaking prose inside details:, renaming a variable).

When skipping, briefly say so in your reply ("skipping slide-fit-check — edit is doc-only") so the user knows you considered it.

Slide canvas conventions

The dual-format submodule renders slides at 1152 × 648 pt (16:9). The nominal summary: block is dy: 65pt, height: 470pt, but in practice slides routinely use the full canvas — the page-number footer alone sits at y≈639, so any threshold tighter than the page boundary itself produces noise.

The reliable check is therefore: does any element's bbox extend past the page edge (y1 > 648 or x1 > 1152)? That is what Typst silently allows and what visually appears as "spilled over" content in a PDF viewer.

Workflow

1. Identify the session and edited block titles

From the edited file path, extract the session number (e.g., modules/Session_4/1_forward_fx_market.typ → session 4).

From the diff of this turn's edits, list the content-block(...) titles whose summary: was modified. Skip content-block-doc-only titles.

2. Build the session slides PDF

make build/Session_<N>_*/Session_<N>_slides.pdf

If the build fails, stop — fix the build error first; overflow check is moot.

3. Locate target page(s) and check overflow

Run a single Python script that takes the session number plus one or more title strings and reports per-page status. Use python3 with pymupdf (already installed):

python3 - <<'PY'
import fitz, glob, sys

SESSION = 4                                    # ← set
TITLES  = ["Forward FX Market Overview"]       # ← set; one or more

PDF = glob.glob(f"build/Session_{SESSION}_*/Session_{SESSION}_slides.pdf")[0]
doc = fitz.open(PDF)

# Slide canvas (from dual_format submodule)
PAGE_W, PAGE_H = 1152, 648
TOL = 2.0                          # ignore sub-pixel overruns

def collect_bboxes(page):
    bboxes = []
    for b in page.get_text("rawdict")["blocks"]:
        for L in b.get("lines", []):
            for s in L.get("spans", []):
                bboxes.append(("text", fitz.Rect(s["bbox"]), s.get("text", "")[:60]))
    for img in page.get_image_info():
        bboxes.append(("image", fitz.Rect(img["bbox"]), ""))
    for d in page.get_drawings():
        bboxes.append(("draw", fitz.Rect(d["rect"]), ""))
    return bboxes

for title in TITLES:
    hits = [i for i, p in enumerate(doc) if title in p.get_text()]
    if not hits:
        print(f"[!] '{title}' not found in any page — title may have changed")
        continue
    for i in hits:
        page = doc[i]
        bboxes = collect_bboxes(page)
        overflow = [b for b in bboxes if b[1].y1 > PAGE_H - TOL or b[1].x1 > PAGE_W - TOL]
        max_y = max((b[1].y1 for b in bboxes), default=0)
        max_x = max((b[1].x1 for b in bboxes), default=0)

        if overflow:
            status = f"OVERFLOW (max y1={max_y:.1f}/{PAGE_H}; max x1={max_x:.1f}/{PAGE_W})"
        else:
            status = "OK"
        print(f"page {i+1:>2}  {status}  — '{title}'")
        if overflow:
            for kind, r, txt in overflow[:5]:
                print(f"           {kind:5s} bbox={tuple(round(v,1) for v in r)}  {txt!r}")
PY

This single script:

• finds each title in the PDF,
• collects every text/image/drawing bbox on its page,
• flags OVERFLOW when any element extends past the page boundary (y1 > 648 - TOL or x1 > 1152 - TOL),
• prints the offending element rects when there is overflow.

4. (Optional) Visual confirmation for borderline cases

If you suspect a visual issue the bbox math missed (overlapping siblings, image cropped by a container with a clip path, near-edge layout that won't project well), render the page to PNG and view it:

python3 - <<'PY'
import fitz
doc = fitz.open("build/Session_<N>_*/Session_<N>_slides.pdf")
for p_num in [<page>]:                  # 1-based
    pix = doc[p_num - 1].get_pixmap(dpi=80)
    pix.save(f"/tmp/slidefit_s<N>_p{p_num}.png")
PY

Then Read the resulting PNG (Claude can view images directly).

5. Report

For each checked page, report one line:

page <N>  — <title> — OK   |   OVERFLOW: <numbers>

If overflow, propose a concrete fix:

• shrink font: #set text(size: <X>em) inside a presentation-only[ ... ] wrapper
• reduce image width (width: 80%) or restructure SVG aspect to be flatter
• restructure the grid (drop a column, move bullets to details:)
• split into two slides via separate content-block calls

If everything passes, say so in one line. Do not fabricate issues.

Why programmatic, not visual-only

The original draft of this skill assumed visual inspection of rendered PNGs, similar to how a human checks a slide. The bbox-based check above is more reliable: Typst writes the off-page text/image coordinates into the PDF stream verbatim, so a numerical comparison against page.rect deterministically detects overflow without depending on the model's image-reading judgment. Reserve PNG inspection for the borderline cases the bbox check cannot catch (overlapping siblings, intra-page collisions, asymmetric layouts).

Evolving this skill

When a slide overflow occurs that this check missed, add a new detection rule here (e.g., new collision class, new dual-format-region constants if the submodule's slide template changes). The tolerance constants (TOL, INNER_BOT, PAGE_H) are the natural place to evolve.