JRC C4 — practical AI usage, illustrated with a forecasting experiment

A three-level prompt-engineering demo for a JRC C4 talk on practical AI for research code. The same forecasting task — Germany hourly electricity load for the first week of January 2020 — is run through Claude Code (Opus 4.7) at three levels of user sophistication: prompt and workspace context scale together to model a beginner (L1), an average user (L2), and a pro (L3). Result: MAPE 10.76 → 5.52 → 3.43, same agent, same data.

Layout

• prompts/{L1,L2,L3}.md — the three frozen prompts. Inputs to the experiment.
• data/opsd_de_load.csv — canonical copy of the German hourly load data (Open Power System Data, 2015–2020, ~50k rows).
• scripts/fetch_data.py — re-runnable producer of data/opsd_de_load.csv.
• scripts/build_slide_figures.py — generates the bespoke deck figures (prompt_lengths.png, 07_feature_importance_normalised.png).
• scripts/verify_slide_numbers.py — cross-checks every quoted MAPE, coverage, word count, and figure path in the deck against the source metrics. Runs in CI-style: prints 21 OK/FAIL lines.
• runs/L1/ — beginner workspace: just the data file. No AGENTS.md.
• runs/L2/ — average-user workspace: data + 7-line AGENTS.md.
• runs/L3/ — pro workspace: data + 113-line AGENTS.md with output schema, do-nots, style. Includes the bake-off code (runs/L3/code/) and full results (runs/L3/figures/, metrics.json, transcript.md).
• slides/ — see next section.
• RUNBOOK.md — step-by-step for the human operator running Phase 1.

Slide deck

File	Purpose
slides/talk_v2.md	Marp source of the 24-slide standalone deck (the full version).
slides/talk_v2.pdf	Rendered standalone deck. View directly on GitHub.
slides/talk_v2.pptx	PowerPoint export of the same. Editable.
slides/jrc_section_v2_beamer.tex	XeLaTeX/Beamer source of the 8-slide drop-in for the JRC C4 parent deck. JRC blue + EC emblem footer.
slides/jrc_section_v2_beamer.pdf	Rendered Beamer drop-in.
slides/assets/	Bespoke figures used by the decks (prompt-length bar, normalised feature-importance, extracted EC emblem).

Regenerate decks

# Marp standalone (PDF + PPTX)
CHROME_PATH="/Applications/Brave Browser.app/Contents/MacOS/Brave Browser" \
    marp slides/talk_v2.md -o slides/talk_v2.pdf --allow-local-files
CHROME_PATH="..." marp slides/talk_v2.md -o slides/talk_v2.pptx --allow-local-files

# Beamer drop-in (PDF only; run twice for refs)
cd slides && xelatex jrc_section_v2_beamer.tex && xelatex jrc_section_v2_beamer.tex

Per-slide PNG export (drag-and-drop into PowerPoint)

mkdir -p slides/jrc_png && rm -f slides/jrc_png/*.png
pdftoppm -r 900 -png slides/jrc_section_v2_beamer.pdf slides/jrc_png/slide

The slides/jrc_png/ folder is gitignored — regenerate locally when needed. 900 dpi gives ~5670×3189 px per slide (~1 MB each). 600 dpi is the conservative middle ground; 400 dpi is too soft for projector use.

PowerPoint Online (Teams) note

PowerPoint Online (the web version, used inside Teams) does not expose the desktop app's "Do not compress images" setting. Whatever PNG you drag in is what's stored; on display the browser handles the scaling. If inserted images look soft, the cause is almost always one of:

1. You inserted lower-DPI PNGs earlier and never replaced them.
2. The deck is being previewed in the small edit pane, not the full-screen slideshow view.
3. Browser zoom level / monitor pixel density.

If onscreen quality during projection is unsatisfactory, share the PDF directly via Teams screen-share instead of routing through PowerPoint at all — it avoids both the PowerPoint image pipeline and the browser scaler.

Quickstart (reproduce the experiment)

uv sync                                    # exact versions from uv.lock
uv run python scripts/fetch_data.py        # produces data/opsd_de_load.csv

Then follow RUNBOOK.md for the Phase-1 manual runs (L1 → L2 → L3). Each run lands in runs/Lx/ with code, figures, metrics, and a methods write-up.

The deliberate point of the three personae: workspace context scales with the prompt. A real beginner doesn't have AGENTS.md set up; a pro does. The three rows of the headline aren't just three different prompts, they're three different user personae.

Tools used

The forecasting experiment, the analytical write-ups in runs/L*/transcript.md, the deck source (slides/talk_v2.md, slides/jrc_section_v2_beamer.tex), and the supporting scripts were all produced in collaboration with Claude Code (Opus 4.7) as the agent under test and as the build assistant. Where the experiment runs the agent at three increasingly specific prompt levels, the deck and code around it were built with the L3 workflow throughout: detailed prompts, a project-level AGENTS.md, and the /review-loop skill for plan review and code review.