AI Event Card: Manus, Genspark, and What Actually Printed

Manus interpreted *table card* as per-attendee place card and produced 9 A4 pages of 4-up fold-tent cards — 36 individual folded cards, one per dinner guest. The fold-tent geometry is well-executed: the top half of each cell is mirror-flipped, so when folded along the dotted line both faces read upright from across the table. Bilingual typography hierarchy on each card is clean. Names spell right.

But it's the wrong artifact. The brief asked for one shared A4 sheet with agenda on Side A and the full directory on Side B. The dinner team can't place 36 individual tents on a table that holds 36 guests sharing one program — that is a 10x print job in cardstock, scoring, and folding labor the team did not budget for. Worse, the per-attendee design omits the agenda entirely; a guest looking at the tent in front of them learns nothing about the program.

The takeaway isn't that Manus failed — it's that it failed *upstream of execution*. The agentic loop optimized confidently against the wrong target, and the artifact it produced was *more polished* than the one it should have produced. Both Manus and our own first read of the brief made the same interpretation error. The only thing that caught it was the human-in-the-loop correction step, which an agentic build skipped by design.

Genspark got the brief right — one A4 sheet, agenda on Side A, attendee directory on Side B. Structural layout is similar to what we eventually shipped: left-rail time column on the agenda, 4-column grid on the directory, bilingual cells. The typography layer has three ship-blockers:

1. CJK font fallback failures (tofu — □). Multiple Chinese characters across both sides render as the placeholder box browsers show when the font has no glyph. Genspark's serverless Chrome did not have a CJK font loaded. The agenda title reads *寻梦之夜 □ 文明与智能* — even the separator dot is unrenderable.
2. Romanization mangling on the directory. Side B shows names like Wenbo Bao□Wang, with Latin name fragments bleeding between adjacent cells. Classic LLM-tabulation hazard: asked to fill a tight grid from a list, the model loses the row boundary.
3. Misused keyart + unsubstituted placeholders. The keyart is shown as a stamp-sized thumbnail in the upper right of Side A, not as the banner the brief asked for. Two Curify Logo literal-text placeholders never got substituted out of the footer.

Genspark's failure is the inverse of Manus's. Manus shipped a polished artifact for the wrong brief; Genspark shipped the right artifact with broken execution. Brief interpretation correct, rendering pipeline can't honor the bilingual fidelity bar.

Our team shipped a 2-page A4 PDF the dinner team can hand to any office printer for duplex output. The path was deliberately boring: a single event-card.html with embedded print CSS, two @page rules sized to A4 portrait, the keyart used as a literal background image (no AI re-render), Chrome headless to export the PDF.

Four design moves:

1. Banner as raster, not as generation. The 3125x5558 keyart JPG the event team already designed becomes a CSS background-image showing the top 72mm. Zero AI involvement at the visual layer, zero risk of a hallucinated glyph on the brand asset.
2. Text as actual text. Every name, title, time, and chapter heading is a real DOM text node served by the host browser's CJK font stack (Source Han Serif → Noto Serif CJK → PingFang → Songti fallback). Output is text-selectable; every glyph renders identically across Preview, Adobe Reader, and an office laser.
3. Spreadsheet as input. The 36 guests come straight from attendence.txt (tab-separated). No LLM rewriting. If a row is wrong, it's wrong in one place.
4. Print geometry as math. A4 = 210x297mm. Side A's 72mm banner + meta + header + 5-chapter agenda + footer fits the remaining 205mm usable. Side B's 4x9 grid (36 cells × ~27mm + header) fits 270mm. Tuned twice against the rendered PDF.

Total build time: ~90 minutes including two render-tune loops on Side A spacing. What an actual print designer would ship — authored in HTML+CSS rather than InDesign.

Three failure modes are systemic, not tool-specific. The next AI-document tool you try will hit at least one by default, and the one after that will hit two:

1. CJK font fallback in serverless rendering. Server-side Chrome (Playwright, Puppeteer, Browserless) inherits the system font stack of its container. Most Linux containers ship Latin-only (DejaVu, Liberation). Hand it 寻梦之夜 and every Chinese glyph renders as the .notdef tofu box (□). Fix: one apt-get install fonts-noto-cjk in the container build — but the AI tool's preview shows it fine because the *generation step* runs in a CJK-fonted environment while the *PDF render step* runs in a different container. The bug is invisible until the user opens the artifact.

2. LLM tabulation losing row boundaries. Genspark's Wenbo Bao□Wang is the model concatenating two adjacent rows of the directory while trying to lay them out. Asked to render a 36-row list into a 4x9 grid, the model operates on the linear sequence; spatial-to-linear mapping is a non-trivial step it does not always perform correctly. Fix: give the layout engine the 36 cells as a structured array and never let the model re-derive the coordinates.

3. Brief ambiguity compounded by agentic loops. A human PM hearing *table card* would ask *do you mean one card per attendee, or one program per shared table?* and disambiguate. An agentic loop picks an interpretation, plans, generates, and ships. By the time you see the output, you're reviewing a 90%% finished artifact for the wrong brief — and the sunk-cost gravity of *it looks done* makes the interpretation error harder to spot.

This brief selected for HTML+CSS. A Pinterest-pin brief selects for image-gen. A 30-template-family content drop selects for AI-assist + human curation. Three rules of thumb:

AI image-gen when the artifact is a single hero visual with low text density, audience views on screen (minor glyph errors forgivable), and *almost right* is acceptable. Marketing heroes, social cards, mood boards, product mockups. See our style transfer guide and product photo generator workflow.

HTML+CSS (templated or hand-crafted) when every glyph must be perfect, the artifact is destined for print or PDF, the layout is grid-shaped with many cells, and content has a single source-of-truth file. Event programs, conference badges, name-card sheets, invoices, certificates, menus, wedding seating charts. The rendering pipeline must include the right fonts; content must come from the source-of-truth file with zero LLM rewriting.

AI-assisted + human-curated when volume is high (50-500 artifacts), visual identity must be consistent, and content fields are structured. Curify's daily drops produce hundreds of bilingual nano-template inspirations a week this way — gpt-4o-mini handles structured fields + aliases, the template-prompt pipeline handles the visuals, a human reviews the batch before sync. See the cross-border ecommerce SEO case for the structured-content parallel.

The mistake all three tools made on this brief: treating *bilingual event print* as if it were *bilingual marketing image*. Share words. Don't share the fidelity bar.