Skip to content

Repository files navigation

Quantitative Grounding

QG-01 is a portable, model-agnostic behavioral skill that adds the minimum sufficient quantitative structure needed to understand scale, comparison, likelihood, economics, uncertainty, and decision relevance without numerical theater or false precision.

Quantify when numbers reveal reality. Do not quantify when numbers merely imitate certainty.

Why this exists

AI answers can be qualitatively correct yet quantitatively ungrounded. They may describe something as large, likely, expensive, dominant, fast-growing, or important without establishing magnitude, baseline, time horizon, uncertainty, or economic significance.

QG-01 turns the recurring question "compared with what?" into an explicit reasoning protocol.

Quick start

System or developer prompt

Copy the contents of SYSTEM_PROMPT.txt into the persistent instruction layer of your model or agent.

Agent skill

Provide SKILL.md to the agent as a behavioral skill and activate it with:

Activate QG-01 by default for substantive analytical questions.

For a single request:

Run QG-01 on this question.

Depth modes:

QG-01/Lite
QG-01/Standard
QG-01/Deep

Codex plugin

Install the Codex Skills marketplace, then install this immutable, validated package:

codex plugin marketplace add https://github.com/andydrewie/codex-skills
codex plugin add quantitative-grounding@andydrewie-codex-skills

The root plugin.json follows Agent Plugins 1.0. The generated .codex-plugin/plugin.json keeps Codex 0.145 compatible while the root manifest is used by Codex 0.146 and newer.

Governing rule

A number should be included only when it is:

  1. Relevant: it materially changes interpretation.
  2. Decision-useful: it improves comparison, prioritization, forecasting, or action.
  3. Epistemically defensible: it comes from reliable evidence, transparent calculation, or clearly stated assumptions.

When reliable quantification is unavailable, remaining qualitative is the correct behavior.

Repository map

File Purpose
SKILL.md Normative behavioral specification and source of truth
agents/openai.yaml Codex UI metadata and default invocation prompt
plugin.json Portable Agent Plugins 1.0 manifest
.codex-plugin/plugin.json Generated Codex 0.145 compatibility manifest
skills/quantitative-grounding/ Generated fixed-location plugin skill mirror
SYSTEM_PROMPT.txt Compact deployment prompt
AGENTS.md Agent-facing integration and precedence rules
skill.json Machine-readable manifest
EVALS.md Behavioral tests and pass/fail criteria
schemas/quantitative-frame.schema.json Optional structured-output schema
examples/ Human and machine-readable usage examples
scripts/validate.py Dependency-free repository validator
CANONICAL_SUMMARY.md Human-readable formalization
CHANGELOG.md Version history

If files conflict, SKILL.md controls.

Agent integration contract

An agent implementing QG-01 should:

  • answer the actual question first;
  • add one to five high-value quantitative anchors when applicable;
  • state the baseline, denominator, period, or reference class when material;
  • distinguish fact, reported data, calculation, estimate, scenario, and speculation when provenance matters;
  • distinguish event probability from confidence in the estimate;
  • prefer ranges and scenarios when inputs do not justify a point estimate;
  • omit quantitative framing when it would create false precision.

See AGENTS.md for the complete agent-facing contract.

Validation

Run locally:

python scripts/validate.py

Check Codex skill-structure compatibility with the installed official validator:

python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py skills/quantitative-grounding

The GitHub Actions workflow runs the same validation on every push and pull request.

Version

Canonical skill version: 1.0.3

This packaging-only release adds portable Agent Plugins discovery and the transitional Codex compatibility bridge without changing QG-01 behavior.

Author

Created by Andrew Fai (@andydrewie) through iterative human-AI specification design.

License

MIT. See LICENSE.

About

Give Codex a better instinct for scale—an explicitly invoked skill for honest numbers, clear uncertainty, and no false precision.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages