Composing Data Insights

Composing Data Insights

Overview

Turn user-supplied data into a polished deliverable for executive review. Three caller-selectable rendering modes:

• html — a single self-contained .html file (inline CSS / inline SVG, no CDN by default)
• image — one or more PNG executive infographic images, produced by invoking the bundled scripts/generate_image.py
• image+html — HTML plus one or more PNG images, written to the same directory and sharing a stem; the HTML references the PNG files via relative <img> tags (never base64 inline)

The four-bucket analysis (facts / interpretation / risks / recommendations) is computed in every mode — it is upstream of both the HTML body and the image prompt, never a standalone deliverable.

Operating constraints (always)

• No data fetching. Skill never queries databases, APIs, or scrapes — the user supplies the bytes.
• No re-implementing image I/O. Image generation always goes through the bundled scripts/generate_image.py; never emit curl, never base64-decode in chat, never hardcode a gateway URL.
• Language-neutral. Output language follows the data and the user's chat — nothing is hardcoded to any specific language.
• PII / secrets aware. Sensitive fields are detected, confirmed with the user, and aggregated / masked / excluded before they reach the HTML body or the image prompt. See references/pii-and-secrets.md.

When to Use

Use this skill when the user:

• Pastes / attaches data (CSV, Excel, JSON, table, metric snippet) and wants a deliverable for review
• Provides, or another skill hands off, a normalized report model that needs facts / interpretation / risks / recommendations and polished Markdown / HTML / image output
• Asks for a "report", "dashboard", "infographic", "executive summary", or similar polished artifact
• Wants a written analysis with insights, hidden risks, and recommendations

Do not use this skill when:

• The user is asking for a live SQL pull, API fetch, or scrape — that's data-acquisition, out of scope here
• The user wants "just an image" of something unrelated to data (e.g. "draw a cyberpunk city") — use a separate general image capability outside this skill if the host has one
• The user asks for a single number to format / a quick arithmetic result — too small, just answer in chat
• The user wants a multi-page paginated report or interactive dashboard — out of scope

Mode Contract

The caller picks one of image / html / image+html. The caller may also specify image_count from 1 to 3. The skill must not silently choose mode. If mode is unspecified, ask once before rendering:

Which mode would you like — image (PNG infographic image[s]), html (self-contained report), or image+html (HTML plus PNG image[s] in the same directory)?

For image+html:

• Files share a stem from the user-provided output path (e.g. /tmp/q1-review → q1-review.html + q1-review.png, or q1-review-1.png ... q1-review-3.png when image_count > 1)
• Files land in the same directory
• The HTML references PNG files via relative <img> tags — never base64 inline
• Render the HTML first; invoke the bundled scripts/generate_image.py second; if the image step fails, the HTML is still on disk and the missing PNG is flagged at the top of the report

Default Workflow

For every mode, follow these steps:

1. Understand the data. Identify metrics, dimensions, time periods, units, segments, and business context. If the input is an upstream report model, preserve its scope, period, caveats, and evidence links instead of flattening them away. State what's clear and what's ambiguous.
2. Sensitivity check. Scan for PII / secrets — emails, phone numbers, person names, customer / company names, API tokens, passwords, secret-like strings, internal IDs that look like credentials. If any are present, confirm with the user before proceeding and aggregate / mask / exclude before any field enters the HTML body or the image prompt. Use references/pii-and-secrets.md for the detection list and the masking rules.
3. Quality-check. Use references/risk-and-pattern-checks.md to flag missing values, duplicates, inconsistent labels, suspicious zeros, outliers, small sample sizes, unclear definitions, mixed units. Mention quality issues explicitly — never silently clean or ignore.
4. Extract insights. Use references/analysis-framework.md to organize facts, interpretation, risks, and recommendations. Find trends, drivers, segment differences, conversion bottlenecks, anomalies, non-obvious patterns.
5. Surface risks. Include risks even when the headline trend is positive. Each risk gets: name, severity (High / Medium / Low), evidence from the data, why it matters, suggested action.
6. Render per mode:
   • html — write self-contained .html using references/html-report-template.md as a starting point. Adapt to data; do not ship placeholder text verbatim.
   • image — finalize prompt using references/image-prompt-template.md, then invoke the bundled scripts/generate_image.py (see Hand-off below). Do not emit curl snippets.
   • image+html — produce the HTML first with the relative <img> reference; then invoke scripts/generate_image.py to materialize the PNG to the same directory and stem.

Hand-off to scripts/generate_image.py

When the mode includes an image, the rendering goes through the bundled scripts/generate_image.py. This skill never re-implements the HTTP call.

Invocation pattern (run by the agent, not embedded in the deliverable):

# Use mktemp so concurrent runs don't collide and to avoid /tmp symlink races.
prompt_file=$(mktemp -t composing-prompt.XXXXXX)
trap 'rm -f "$prompt_file"' EXIT INT TERM   # cleanup on success, error, or kill

cat > "$prompt_file" <<'EOF'
<finalized prompt text — see references/image-prompt-template.md>
EOF

python3 "<skill-dir>/scripts/generate_image.py" \
  --prompt "$(cat "$prompt_file")" \
  --output "<user-path>.png"

When image_count > 1, pass --n <image_count> and let the bundled script expand the output paths to <stem>-1.png, <stem>-2.png, and so on. Keep image_count bounded to 3 for report / slide-style deliverables unless the caller explicitly justifies more.

<skill-dir> is the directory where this SKILL.md lives (typically .ai/skills/composing-data-insights in this repo, but the bundled script is co-located so the skill is portable as a single copy). <user-path>.png must be quoted — paths with spaces (common on macOS / Windows-mounted shares) otherwise break the --output argument.

Rules:

• Never emit curl to chat or to any deliverable
• Never read OPENAI_BASE_URL or OPENAI_API_KEY directly — generate_image.py does that
• Never base64-decode or pipe jq | base64 --decode — fragile, already handled inside the script
• Never use a fixed temp file path like /tmp/composing-prompt.txt — concurrent runs will collide; use mktemp
• Always pair mktemp with trap 'rm -f "$prompt_file"' EXIT INT TERM so the file is removed on success, error, or kill (pii-and-secrets.md requires cleanup on success or failure)
• Always quote --prompt and --output arguments

If the user's environment is missing OPENAI_API_KEY, generate_image.py exits with code 2 and prints a one-line reason to stderr — surface that verbatim and stop. Do not hardcode a fallback gateway URL.

Upstream Report Model Handoff

Other skills may pass a normalized report model instead of raw tabular data. Treat that model as the structured input and keep these fields intact when present:

• period
• scope
• metrics
• items
• people
• caveats
• source URLs or evidence links

For GitHub reports, github-activity-report owns acquisition and normalization. This skill owns the generic analysis discipline and presentation layer. Do not fetch GitHub data from this skill.

Common Mistakes

Smoke

End-to-end smoke runbook + toy fixture live at smoke/README.md and smoke/toy-saas.csv. Use them to verify all three modes after edits. Image and image+html modes hit the real gateway; html mode is offline.