results-to-slides

Results to Slides

Discover experiments, collect images/metrics, organize into slides. You provide the story — this skill handles the grunt work. Output: slide markdown + editable PPTX.

Core Rules

1. State what was done and what the result was. Do not editorialize. Never use: "breakthrough", "key insight", "importantly". Just experiment + number.
2. Every slide MUST have a visual. No text-only slides (except pure tables). If a concept has no experiment images, create a diagram. If no diagram makes sense, show dataset examples.
3. Images first, text second. Plan images for each slide BEFORE writing bullets. If you can't find an image for a slide, restructure or merge it.

---

Parallel Execution

Maximize use of the Agent tool. Whenever you have 2+ independent tasks, launch parallel agents.

• Phase 1: Launch agents in parallel to read CLAUDE.md, README.md, memory files
• Phase 2 (biggest win): Launch separate agents for git log, output folders, scripts, media
• Phase 3: Launch agents in parallel to read different experiment scripts and output folders
• Phase 5+6: Sequential (markdown must be written before converting)

---

Important Paths

• Skill directory: ${CLAUDE_SKILL_DIR}
• Converter: ${CLAUDE_SKILL_DIR}/md_to_pptx.py
• Backgrounds: ${CLAUDE_SKILL_DIR}/backgrounds/
• Theme CSS: ${CLAUDE_SKILL_DIR}/theme.css
• Slide element reference: [slide_reference.md](slide_reference.md)

---

Phase 0: Parse Arguments & Setup

The user provides: /results-to-slides START_DATE END_DATE

Arguments come as $ARGUMENTS containing two MMDD date strings (e.g., 0301 0308).

Parse the dates:

• $0 = start date (MMDD), $1 = end date (MMDD)
• Infer year from system date. End date is INCLUSIVE (use end_date + 1 for find bounds)

If no arguments provided, use AskUserQuestion:

• header: "Date range"
• question: "What date range? Use MMDD MMDD format (e.g., 0301 0308)."
• options: ["Last week", "Last 2 weeks", "Last month"]

Output directory: presentation/YYYY_MM_DD/ using the end date.

Ask Presentation Preferences

Ask all preferences upfront before discovery, in a SINGLE AskUserQuestion call:

1. Background theme: "Light (Recommended)", "Warm", "Dark"
2. Emphasis areas: "What should the presentation emphasize? (e.g., dataset, model architecture, results comparison, failed approaches)" — free text, helps prioritize slide count allocation.

| Choice | Background file | --theme flag | |--------|----------------|----------------| | Light | slide_bg_light.png | light | | Warm | slide_bg_warm.png | light | | Dark | slide_bg_dark.png | dark |

Do NOT ask for slide count. Generate as many slides as the content needs — images drive the count, not a preset number.

---

Phase 1: Research Context Discovery

Scan project docs (CLAUDE.md, README.md, memory files, docs/) to understand:

• Research question and direction
• Key metrics and what "good" vs "bad" looks like
• Terminology, model names, dataset names, output folder conventions
• Prior results and baselines
• Previous presentations (check presentation/ directory for prior slide decks to recap)

If no research goal found, use AskUserQuestion:

• header: "Research context"
• question: "Brief description of what this project does?"
• options: ["Let me describe it", "Use README description"]

Use this context internally for smart organization (grouping, filtering, prioritization). Do NOT let it leak into slide text.

---

Phase 2: Experiment Discovery

Use ALL methods — each catches things the others miss.

Git Log

git log --after="YYYY-MM-DD_START" --before="YYYY-MM-DD_END+1" --oneline --stat

Extract: commit messages (often contain results), files changed, dates.

Output Folders

find . -maxdepth 2 -type d -newermt "YYYY-MM-DD_START" ! -newermt "YYYY-MM-DD_END+1" 2>/dev/null | sort

Also check: outputs/, results/, experiments/, runs/, logs/, checkpoints/. Do NOT assume MMDD_ naming — use modification time as primary signal.

For each folder: list contents, look for metrics files (metrics.json, *.pkl, scores.txt, *.log) and images (*.png, *.jpg, *.gif).

Scripts

find . -name "*.py" -newermt "YYYY-MM-DD_START" ! -newermt "YYYY-MM-DD_END+1" 2>/dev/null | grep -v __pycache__ | sort

Also check .sh scripts. Read the code to understand what each experiment does — folder names are opaque, scripts tell you everything.

Media Selection — AGGRESSIVE

Collect FAR more images than you think you need. This is the most important discovery step.

1. Read experiment scripts for plt.savefig(...), Image.save(...), cv2.imwrite(...) calls
2. Look for naming patterns: baseline.*, best_*.*, comparison.*, grid.*, eval*.*
3. Also look for videos: .mp4, .avi, .mov — the converter embeds them as playable media
4. For comparison experiments: collect ALL images across ALL prompts/seeds
5. Grid sizing: 1 image → cols-2, 2 → cols-2, 3 → cols-3, 4 → cols-4

Build Timeline

Cross-reference git commits ↔ output folders ↔ scripts. Build:

DATE | EXPERIMENT_NAME | SCRIPT | OUTPUT_FOLDER | KEY_RESULT | IMAGES

---

Phase 2.5: Diagram Planning

For slides explaining concepts, pipelines, or architectures — create diagrams.

• Write as clean HTML/CSS, render to PNG with the diagram renderer: ```bash # One-time setup: cd /tmp && npm install puppeteer node ${CLAUDE_SKILL_DIR}/render_diagram.js input.html output.png [--width 1200] [--scale 2] ```
• Read the rendered PNG to verify readability before embedding.
• Do NOT use matplotlib for diagrams — only for actual data plots. HTML/CSS diagrams are far more readable in presentations.
• Save HTML files in the presentation directory for reuse.

---

Phase 3: Organize Experiments into Slides

Order slides chronologically — the audience follows the story of discovery.

Visual Budget Rule

Every slide in the outline must have one of: [IMG], [DIAGRAM], [TABLE], [DATASET]. If a slide has none, merge it into an adjacent slide or add a visual.

Key Rules

• Group variations of the same idea into one slide with a comparison table
• Filter debug scripts, typos, one-off tests
• Success/failure galleries: minimum 5 success + 5 failure slides showing cols-4 grids of baseline | method variants. Pick diverse examples — don't repeat the same prompt. Annotate failures: "(no change)", "(worse)", "(baseline better)".

Per Slide

Each experiment slide contains:

1. What was done (factual description from script/commit)
2. What the result was (metrics from logs/files)
3. Visual evidence (images from output folder)

---

Phase 4: Script Generation & User Review

Generate a structured outline:

SLIDE SCRIPT
============

Slide 1: Title
  Type: Title (lead)
  Heading: [Project Name — Weekly Update]
  Subtitle: [Date range, N experiments]
  Chips: [N slides], [date range], [key topics]
  Visual: [HERO_IMG]

Slide 2: [Experiment Name]
  Type: Content
  Heading: [What was done: key metric]
  Bullets:
    - [Configuration detail]
    - [Result metric]
  Visual: [IMG] path/to/image.png | [DIAGRAM] pipeline | [TABLE] | [DATASET]
  Grid: cols-[2/3/4]

Every slide MUST have a Visual field. Flag any slide without one.

Review Mode

Use AskUserQuestion:

• header: "Review mode"
• question: "How would you like to review?"
• options: "Show each slide" (approve one at a time), "Show all slides" (feedback at once), "Skip review"

For per-slide review, use AskUserQuestion per slide with options: "Approve", "Edit", "Remove".

---

Phase 5: Generate Slide Markdown

Setup

mkdir -p presentation/YYYY_MM_DD
cp ${CLAUDE_SKILL_DIR}/theme.css presentation/YYYY_MM_DD/
cp ${CLAUDE_SKILL_DIR}/backgrounds/CHOSEN_BG.png presentation/YYYY_MM_DD/slide_bg.png

Write Markdown

Write to presentation/YYYY_MM_DD/slides.md.

Use ONLY elements from [slide_reference.md](slide_reference.md). The PPTX converter only understands those specific HTML elements.

File header:

---
theme: notebook-status
paginate: true
size: 16:9
html: true
---

Slide separator: --- on its own line.

Style Rules

Element order (strict, top to bottom): heading → bullets (max 3, prefer 2) → table → images (always last).

Banned elements on content slides: chips, split layout, arch-box, stat cards, eyebrow, decision card, timeline. Chips are ONLY allowed on the title (lead) slide.

Do NOT generate: next steps slides, takeaway slides, or any slide requiring editorial interpretation.

Headings: factual label of experiment + result. "ResNet-50 on ImageNet: 76.1% Top-1" not "ResNet-50 Is the Clear Winner".

Bullets: facts only. Max 3 per slide (prefer 2 — more space for images). Keep short.

Table slides: NO bullets. One context line above the table instead. Max 5 rows.

Formatting: **bold** for key numbers and method names. ` code ` for technical identifiers.

Image captions: terse identifiers, not sentences. Use <strong> for scores.

Image paths: relative to project root. Converter resolves via --base-dir.

Title slide hero image: The title slide can include a cols-2 image grid below the chips.

One visual concept per slide — never mix grid types or have multiple unrelated sections.

Image-to-claim consistency: If a slide mentions a specific example, the images on that slide MUST show that exact example. Verify image filenames match the described content.

When uncertain about an experiment's purpose, present observable facts and let the user interpret. Do NOT guess intent.

---

Phase 6: Convert to PPTX

python ${CLAUDE_SKILL_DIR}/md_to_pptx.py \
  --input presentation/YYYY_MM_DD/slides.md \
  --output presentation/YYYY_MM_DD/slides.pptx \
  --bg-image presentation/YYYY_MM_DD/slide_bg.png \
  --theme light \
  --base-dir .

Use --theme dark for dark background. Use --theme light for light and warm.

If python-pptx is missing: check conda/mamba environments, CLAUDE.md, or ask the user.

Report

Presentation generated!

  Markdown: presentation/YYYY_MM_DD/slides.md
  PPTX:     presentation/YYYY_MM_DD/slides.pptx
  Slides:   [count]

Open in PowerPoint or LibreOffice Impress to edit further.

---

Error Handling

• No experiments found: Suggest expanding date range
• No images: Generate text-only slides, note which could benefit from images
• Missing python-pptx: Report error, still deliver markdown
• Background image missing: Warn but continue — slides will have white background
• Git not available: Fall back to file modification times