Skip to content

CARBON_SKILL.md

Latest commit

 

History

History
225 lines (175 loc) · 6.4 KB

CARBON_SKILL.md

File metadata and controls

225 lines (175 loc) · 6.4 KB

FLUENTLY SKILL: carbon_tracker

============================================================

Skill type: transversal (applies across all frameworks)

Trigger: carbon | CO₂ | emissions | token cost | sustainability

Requires: CARBON_KNOWLEDGE.md in context (or as attached knowledge)

Code needed: NO — pure reasoning + arithmetic

============================================================

SKILL IDENTITY

name: carbon_tracker version: 0.1.0 type: transversal category: observability description: > Track, calculate, and report CO₂ emissions from LLM token consumption across any conversation, session, or project. Compare emissions across Fluently framework approaches. No code or library required.

WHEN THIS SKILL ACTIVATES

Activate when the user or calling agent:

  • Asks about carbon, CO₂, emissions, environmental impact, or energy use
  • Asks to compare frameworks by efficiency (not just output quality)
  • Asks "how much has this cost in CO₂?"
  • Asks "which framework is greener?"
  • Enables carbon tracking at session or project start
  • Requests a summary at end of task

INPUTS

inputs:
  required:
    - model:          string    # e.g. "claude-sonnet-4-6"
    - input_tokens:   integer
    - output_tokens:  integer

  optional:
    - framework:      string    # "4D" | "linear" | "cyclic" | null
    - domain:         string    # e.g. "rfp_drafting"
    - difficulty:     enum      # "low" | "medium" | "high"
    - scope:          enum      # "call" | "session" | "project"
    - compare_runs:   list      # list of {label, framework, input_tokens, output_tokens}
    - show_analogy:   boolean   # default true
    - show_comparison: boolean  # default false

SKILL STEPS

STEP 1 — Identify model key

Match model string to reference table key using substring rules defined in CARBON_KNOWLEDGE.md.

Output: model_key, input_rate, output_rate, confidence

STEP 2 — Calculate CO₂

input_co2  = input_tokens  × input_rate
output_co2 = output_tokens × output_rate
total_co2  = input_co2 + output_co2         (in gCO₂eq)
total_tokens = input_tokens + output_tokens
per_100k   = (total_co2 / total_tokens) × 100,000
range_min  = total_co2 × 0.60
range_max  = total_co2 × 1.40

STEP 3 — Accumulate (if scope = session or project)

Add to running totals maintained in agent memory or conversation state:

session.total_co2     += total_co2
session.input_tokens  += input_tokens
session.output_tokens += output_tokens

STEP 4 — Select analogy

From the analogies table in CARBON_KNOWLEDGE.md, find the closest match to total_co2 and format as a plain-language equivalent.

STEP 5 — Render output

Use the output format defined in CARBON_KNOWLEDGE.md.

If scope = call: use the full carbon block format. If scope = session or project: prepend cumulative totals. If show_comparison = true and compare_runs provided: use comparison format.

Always append the honest caveats section.

OUTPUTS

outputs:
  total_co2_gCO2eq:     float    # e.g. 0.034
  range_min_gCO2eq:     float
  range_max_gCO2eq:     float
  per_100k_gCO2eq:      float    # normalised rate
  total_tokens:         integer
  model_key:            string
  confidence:           string   # "high" | "medium" | "low"
  analogy:              string   # human-readable equivalent
  formatted_block:      string   # ready-to-display text block
  comparison:           object   # only if compare_runs provided
    most_efficient:     string
    baseline:           string
    runs:               list

COMPARISON SUB-SKILL

When compare_runs is provided, execute comparison:

For each run in compare_runs:

  1. Calculate CO₂ using STEP 1–2
  2. Calculate per_100k

Then:

baseline_run   = run where framework is null, or highest per_100k if none
for each run:
  saving_pct = (1 - run.per_100k / baseline_run.per_100k) × 100

Rank runs by per_100k ascending (most efficient first).

Format using the comparison block from CARBON_KNOWLEDGE.md.

EXAMPLE INVOCATIONS

Single call tracking

Skill: carbon_tracker
Inputs:
  model: "claude-sonnet-4-6"
  input_tokens: 843
  output_tokens: 400
  framework: "4D"
  domain: "rfp_drafting"
  scope: "call"

Expected output:

┌── Carbon Estimate ──────────────────────────────────┐
  Model          claude-sonnet-4-6
  Framework      4D
  Domain         rfp_drafting
  Tokens         1,243  (in: 843 / out: 400)
  CO₂ estimate   0.034 gCO₂eq  [0.020 – 0.048]
  Per 100k       2.7 g / 100k tokens
  Equivalent     ~1% of a smartphone charge
  Confidence     medium
  Reference      Fluently Carbon · RATE_TABLE 2026-03
└────────────────────────────────────────────────────┘
⚠   Anthropic does not disclose inference energy data.
   Values are modelled estimates. Source: EcoLogits (ecologits.ai).

Framework comparison

Skill: carbon_tracker
Inputs:
  domain: "rfp_drafting"
  difficulty: "medium"
  show_comparison: true
  compare_runs:
    - label: "no_framework"  framework: null  input_tokens: 1100  output_tokens: 620
    - label: "fluently_4D"   framework: "4D"  input_tokens: 843   output_tokens: 400
    - label: "fluently_linear" framework: "linear" input_tokens: 920 output_tokens: 480

SKILL CONSTRAINTS

  • Never fabricate CO₂ numbers. If model is unknown, use DEFAULT key and mark confidence "low".
  • Never claim figures are measured — always say "estimate" or "modelled estimate".
  • Never omit the honest caveats when reporting to a user.
  • Reasoning / extended thinking: always note the ×5–10 multiplier if active.
  • If input_tokens and output_tokens are both 0: return null with message "No token data available — carbon estimate requires token counts."

SKILL METADATA

metadata:
  depends_on:
    knowledge: CARBON_KNOWLEDGE.md
    code: optional (ecologits via fluently[carbon])
  compatible_with:
    frameworks: [4D, linear, cyclic, none]
    providers: [anthropic, openai, google, mistral, cohere, huggingface]
  output_formats: [block, compact, json, markdown_table]
  tags: [carbon, co2, sustainability, observability, transversal]
  reference_url: https://ecologits.ai/latest/
  license: MIT (EcoLogits methodology)