Llm resource usage

LLM Resource Usage

A skill that extracts a user's real LLM usage from local harness files, enriches it with environmental and datacenter data, and produces an infographic + narrative.

---

Step 1: Run the extraction + enrichment pipeline

python scripts/run.py [DAYS]

Default is 30 days. The script auto-detects available harnesses and tries tokscale first (covers 18+ tools), then falls back to direct file parsing for Claude Code, OpenCode, OpenClaw, and Goose.

Capture the JSON output into a variable. If there's an error or no data, tell the user what harnesses were detected and what's missing. If record_count is 0, suggest they install tokscale: cargo install tokscale or check https://github.com/junhoyeo/tokscale for other install methods.

---

Step 2: Parse the enriched JSON

The JSON has these top-level keys:

Key	Contents
meta	Days analyzed, extraction method, detected harnesses, warnings
totals	input_tokens, output_tokens, cache_read, total_tokens, cost_usd
environmental	energy_wh, co2_g, water_direct_ml, datacenter info, renewable claim
equivalencies	phone charges, car km, water glasses, tree days, Google searches
cache	hit_rate, tokens_saved, cost_saved_pct
efficiency	tokens_per_dollar, avg_daily, dominant model/provider
by_model	Per-model breakdown with env data and datacenter info
by_harness	Per-harness token counts
timeline	Daily usage array [{date, input_tokens, output_tokens, cost_usd}]

---

Step 3: Generate the HTML report

Run the renderer to produce a standalone HTML file that works in any browser:

python scripts/run.py [DAYS] | python scripts/render.py > report.html

Then tell the user the path to report.html so they can open it in a browser. If the present_files tool is available (claude.ai only), also use it to attach the file for download. The file includes:

• Stat cards (tokens, cost, energy, CO₂)
• Horizontal model breakdown bars with energy per model
• Datacenter location card + cache efficiency
• Equivalency comparisons (phone charges, Google searches, driving, water, tree days)
• Stacked daily timeline chart (Chart.js loaded from CDN)
• Light and dark mode via prefers-color-scheme, print-friendly styles

If running in claude.ai (the visualize:show_widget tool is available), also call it to render the summary inline — use the design spec below. Do both: the HTML file for the user to keep, the widget for immediate in-chat display.

Inline widget spec (claude.ai only)

Call visualize:show_widget twice — once for the summary infographic, once for the timeline.

Design system rules

• No emoji. Use inline SVG icons for phone, car, water, search, tree.
• No dark backgrounds. Use var(--color-background-secondary) for cards.
• Use var(--font-mono) for all numbers.
• Teal #1D9E75 for environmental metrics (energy, CO₂, cache hit). Blue #378ADD for tokens. Amber #EF9F27 for cost.
• Renewable badge: background #E1F5EE, text #085041.
• Card style: background: var(--color-background-secondary); border-radius: var(--border-radius-md); padding: 14px 16px;
• All labels: font-size: 11px; color: var(--color-text-tertiary); letter-spacing: .06em; text-transform: uppercase; margin: 0 0 3px; font-weight: 500;

Visualization A — Summary infographic

Structure (stacked vertically):

Row 1 — 4 metric cards in a grid-template-columns: repeat(4, minmax(0,1fr)); gap: 8px grid:

• Tokens: format large numbers as 1M, 2.4M, 847K etc. Sub-label: "Xk in · Yk out"
• Estimated cost: $X.XX — note "(estimated)" if cost_usd was 0 in raw data
• Energy: show in Wh if <1000, else kWh to 2dp. Sub-label "incl. PUE overhead"
• CO₂: show in g if <1000, else kg. Sub-label "location-based grid"

Row 2 — Model breakdown card (full width): For each model in by_model (sorted by token share descending), show:

• Color swatch (8×8px rounded) — teal=medium, blue=nano, purple=large, amber=small
• Model name, class label in tertiary, and right-aligned "XX% · YYY Wh"
• A CSS progress bar: grey background var(--color-border-tertiary), colored fill, height 7px, border-radius 3px
• Bar width = (model_tokens / total_tokens) * 100%

Row 3 — Two columns (grid-template-columns: 1fr 1fr; gap: 8px):

Left column (stacked):

• Datacenter card: provider/cloud/city, PUE and WUE on one line, renewable claim badge
• Cache card: hit_rate % in teal + cost_saved_pct % in amber, side by side with a thin divider. Add a small visual bar showing the hit rate.

Right column:

• Equivalencies card: for each equivalency, one row with SVG icon + number + label. Show: phone charges, Google searches, car meters (as "X m" or "X km"), water in mL or L, tree absorption days. Rows separated by border-bottom: 0.5px solid var(--color-border-tertiary).

Footer: font-size: 10px; color: var(--color-text-tertiary) — one line: "~Estimates. Sources: [provider] sustainability report (PUE/WUE), EPA eGRID 2023 (grid carbon), Epoch AI 2025 (model energy)."

Visualization B — Usage timeline

Use Chart.js 4.4.1 from cdnjs.cloudflare.com. Stacked bar chart. Canvas wrapper: position:relative; width:100%; height:220px.

// Color scheme — hardcoded hex (Canvas cannot resolve CSS variables)
// Input tokens: #B5D4F4 (blue-200), Output tokens: #378ADD (blue-400)
// isDark check: const isDark = matchMedia('(prefers-color-scheme: dark)').matches;
// gridColor: isDark ? 'rgba(255,255,255,0.07)' : 'rgba(0,0,0,0.07)'
// tickColor: isDark ? 'rgba(255,255,255,0.45)' : 'rgba(0,0,0,0.45)'

Config:

• type: 'bar', two datasets both with stack: 'tokens'
• Input dataset: backgroundColor: '#B5D4F4', border-radius on bottom corners
• Output dataset: backgroundColor: '#378ADD', border-radius on top corners
• Y-axis ticks: callback: v => v >= 1000000 ? (v/1000000).toFixed(1)+'M' : v >= 1000 ? (v/1000).toFixed(0)+'K' : v
• Tooltip: show both datasets in K or M format
• Custom HTML legend (not Chart.js built-in): two squares above the canvas
• scales.x.ticks: { autoSkip: false } — show all date labels
• Only render if timeline.length >= 2. If fewer than 2 data points, skip this visualization with a note.

Skip the cost line overlay if all cost_usd values in the timeline are 0.

---

Step 4: Write the narrative

After the visuals, write a 3–4 paragraph narrative covering:

1. What they used: Total numbers in plain English. Which harnesses and models. Most active days. How diverse their model use was.

2. What it cost the planet: Energy (with a memorable comparison — e.g. "enough to charge your phone X times"), CO₂ (with driving equivalent), water (with glass/ bottle equivalent). Note the datacenter location and renewable energy claim. Be accurate but put things in perspective — individual use is generally small.

3. Efficiency insights: Cache hit rate and what it saved them. Tokens-per-dollar efficiency. Whether they're using the right model sizes for the work. If they have a high proportion of Opus/GPT-4 usage, gently note that Haiku/Sonnet would be more energy-efficient for simpler tasks.

4. One interesting or surprising fact: Something specific to their data — e.g. "Your cache efficiency of 34% means you avoided X tokens of redundant processing", or "Your output-to-input ratio of 1:4 is unusually high, suggesting long generative tasks rather than Q&A", or mention the specific AWS facility their tokens likely ran through.

---

Formatting and tone

• Lead with the visuals, then the narrative
• Use specific numbers, not vague descriptions
• Be honest about uncertainty in environmental estimates — prefix with "~" and note at the end: "Environmental figures are estimates based on published research (Samsi et al 2023, Jehham et al 2025) and provider sustainability reports. Actual values vary with hardware generation, server utilization, and grid mix."
• Don't be preachy about environmental impact — present it as interesting context, not guilt
• If no data was found, clearly explain which harnesses were checked and how to get tokscale installed

---

Handling edge cases

No data found: Check meta.record_count == 0. Tell the user which paths were checked. Suggest installing tokscale or checking if the harness stores data in a non-default location.

Only one model: Skip the model breakdown chart; instead make the summary infographic larger and more detailed.

Very large numbers (>1B tokens): Format as "X.XB". Convert energy to kWh, CO₂ to kg, water to liters.

Very small numbers (<10k tokens): Still produce the output but note it's a small sample. The timeline chart won't be interesting — skip it and just show the summary infographic.

Warnings in meta.warnings: Surface them clearly above the visuals. The most common warning is "tokscale not installed" — display this with the install command.

Missing cost data: The enrich script estimates cost from public pricing tables. Note this with "~$X (estimated from public pricing)".

---

Reference data (don't load unless needed)

The references/data.json file has complete lookup tables for provider datacenter info, model energy classes, grid carbon intensity, and equivalency constants. The enrich script embeds these automatically — you don't need to read this file unless you want to add a custom calculation or override a value.

---

Quick test (no real harness needed)

To verify the pipeline works, run:

echo '{"extraction_method":"test","days_analyzed":30,"detected_harnesses":["claude_code"],
"record_count":150,"warnings":[],"aggregated":{"totals":{"input_tokens":800000,
"output_tokens":200000,"cache_read":240000,"cache_write":80000,"total_tokens":1000000,
"cost_usd":0},"by_date":{},"by_model":{"claude-sonnet-4-6":{"input_tokens":650000,
"output_tokens":160000,"cache_read":240000,"provider":"anthropic","cost_usd":0},
"claude-haiku-4-5":{"input_tokens":150000,"output_tokens":40000,"provider":"anthropic","cost_usd":0}},
"by_harness":{"claude_code":{"input_tokens":1000000}},"by_provider":{},"timeline":[]}}' \
| python scripts/enrich.py