7-Day Conference Poster Rescue: the multi-AI workflow that saved my accepted abstract.

Each tool has one job. Perplexity maps the landscape. Google Scholar turns those names into a folder of full-text PDFs. A PDF → Markdown converter strips the formatting noise that confuses NotebookLM. NotebookLM synthesizes the corpus, generates a deep research report, and reads the papers to me as a podcast. Claude Design turns my hand-sketched prototype into an editable SVG figure.

The handoffs are the point. The poster itself I assemble manually, section by section, in IMRaD order. The AI tools build the inputs; I build the artifact.

Each tool has one job. Using the wrong tool for that job is where most researchers waste time. If you're new to NotebookLM, start with the Quick Start guide first.

Because sketching forces commitment to spatial decisions that text leaves ambiguous. "Show the validation panel in parallel" is interpreted differently by different AI tools. A sketch with the panel drawn below the main flow and an arrow showing the feedback loop is unambiguous.

What goes in the sketch: box positions, arrows (including feedback loops), labels, color regions, and a one-sentence caption.

This is the step I missed the first time. Ten minutes with a pen and a notebook page. The sketch doesn't have to be pretty — it has to be specific. The prototype is also a quality control checkpoint. If I can't sketch it in ten minutes, I don't understand the data well enough to put it on a poster — back to the NotebookLM report.

This is where I get realistic. The final poster has two tables and one figure — not three or four figures. Posters that try to show too much visual content end up with everything at 16pt because there's no room left.

The figure is the methodology figure (Figure 2 from Prompt #2) in the Methods column. It does most of the work because methodology is what reviewers want to see clearly.

The two tables are in the Results column. Table 1 is the comparison table; Table 2 is the outcomes table. Both generated from prompts in the Deep Research Protocol, then edited manually in PowerPoint with real values from my data.

Branding is the step that AI tools genuinely cannot do for me. It is manual work, about 45 minutes, and it is the difference between a draft and a poster that looks like it came from a lab.

• Institution logo — top corners of the title bar. Always vector (SVG or EPS), never a screenshot.
• Lab or research group logo — opposite corner, smaller than the institution mark.
• Color palette match — every figure and table recolored to match the lab's official colors.
• Conference-specific elements — some conferences require their logo or session number.
• Author affiliations — institutional affiliations under the author names.
• Funder acknowledgments — small strip at the bottom with grant numbers and funder logos.

I do not let any AI tool generate institutional logos or attempt to match brand colors from a screenshot. Those decisions belong to my institution's brand guidelines. Manual, every time.

Manually, section by section. IMRaD — Introduction, Methods, Results, Discussion — maps cleanly onto a three-column poster layout.

1. Introduction (top-left column). 100-150 words. Background, gap, research question.
2. Methods (left column, below Introduction). 80-120 words. The methodology figure goes here.
3. Results (middle and right columns, top). Bullet points, not paragraphs. The two tables go here.
4. Discussion (right column, bottom). 100-150 words. Implications, limitations, next steps.
5. References (bottom strip). 8-12 most important. 16-18pt is fine.

Layout standards: landscape orientation, three columns ~15 inches wide, title bar 4-5 inches tall, body font 24pt minimum, headings 32pt, margins 1.5 inches, bleed 0.125 inches.

This is the step that takes 1.5 hours and that I never skip. NotebookLM is source-grounded but "source-grounded" does not mean "always correct." Three things go wrong:

1. It picks the right paper but the wrong sentence. The citation is real, the claim is in the paper, but the sentence pulled isn't the one supporting the claim.

2. It compresses two papers into one synthesized claim that neither makes. Each citation checks out; the combined claim is mine.

3. It silently drops nuance. "Some studies suggest X" becomes "Studies show X." The qualifier matters.

The protocol: for every claim that ends up on the poster, open the cited PDF and find the original sentence. Read at least one paragraph around it. For claims from multiple sources, check both. And read the most central 3-4 papers in full.

The honest accounting: this step is why the workflow takes 10 hours, not 6. Skipping it is how researchers arrive at the poster session with a claim they can't defend.