evidence-first-problem-solving

# Evidence-First Problem Solving — The Hwang Method

> An open Agent Skill for evidence-first investigation, authorized execution, and proof-backed completion in Codex and Claude Code.

Canonical project: https://github.com/aiopshwang/evidence-first-problem-solving
Documentation: https://aiopshwang.github.io/evidence-first-problem-solving/
License: MIT
Current documented release: 0.1.0

## Primary sources

- Skill source: https://github.com/aiopshwang/evidence-first-problem-solving/blob/main/skills/evidence-first-problem-solving/SKILL.md
- Methodology: https://aiopshwang.github.io/evidence-first-problem-solving/methodology/
- Examples: https://aiopshwang.github.io/evidence-first-problem-solving/examples/
- Evaluation protocol: https://aiopshwang.github.io/evidence-first-problem-solving/evaluation/
- Security and privacy: https://aiopshwang.github.io/evidence-first-problem-solving/security-and-privacy/
- Changelog: https://github.com/aiopshwang/evidence-first-problem-solving/blob/main/CHANGELOG.md

## Definition

Evidence-First Problem Solving, also called The Hwang Method, is a risk-adaptive workflow for AI agents. It frames the actual outcome, separates evidence from assumptions, tests competing explanations, chooses evidence-gated complexity, executes within human authority, verifies claims at matching scope, and leaves an inspectable handoff.

## Six phases

1. Understand the outcome, scope, authority, proof standard, confirmed facts, user decisions, analyst judgments, unconfirmed assumptions, and unknown gaps.
2. Challenge competing explanations with discriminating tests.
3. Design the smallest sufficient, authorized approach and record tradeoffs.
4. Execute coupled steps sequentially and independent checks in parallel.
5. Prove each material claim with evidence of matching scope.
6. Hand off claims as proved, partially proved, not proved, or not assessed, and report blockers separately.

## Important qualifications

- Diagnostic probes are not automatically proof of complete delivery.
- The skill does not grant access, deployment, publication, spending, communication, destructive-action, or sensitive-data authority.
- It is not a guarantee of correctness, security, compliance, or performance.
- Public examples and evaluation fixtures are synthetic; conclusions must remain bounded to the evidence reported.

## Installation identifiers

- Repository: aiopshwang/evidence-first-problem-solving
- Skill: evidence-first-problem-solving
- Claude Code marketplace: hwang-method
- Claude Code plugin: evidence-first-problem-solving@hwang-method