Conclusion Labels

A tiny Agent Skill for separating verified facts, inferences, and recommendations in evidence-heavy responses.

Design philosophy

Long agent responses often place sourced facts, interpretations, and advice in the same stream. conclusion-labels adds a small visual contract:

• [V] marks a conclusion supported by current evidence.
• [I] marks a conclusion inferred from stated evidence.
• [R] marks a recommendation or value judgment.

The labels are deliberately compact. They apply only to material conclusions, leaving routine prose untouched. This keeps attention on evidence boundaries without turning every sentence into metadata.

The goal is confidence calibration, not artificial certainty. Requiring the agent to classify a conclusion can make evidence-backed conclusions more decisive, expose assumptions behind inferences, and keep recommendations distinct from facts. It does not verify claims by itself or guarantee factual accuracy.

Observed effect

In extended, evidence-heavy work, this convention has made important conclusions easier for the user to notice and has encouraged the agent to communicate the basis and confidence of its conclusions more explicitly. This is an observed workflow effect, not a controlled measurement of model accuracy.

Install as an Agent Skill

npx skills add ydxrobot/conclusion-labels-skill

The Skill is model-invoked for evidence-heavy analysis, audits, research, and consequential decisions. It has no scripts, tools, or external dependencies.

Compatibility

The standard SKILL.md is compatible with Codex and Claude. agents/openai.yaml provides Codex-specific interface metadata without changing the shared behavior.

Add directly to agent instructions

To make the convention part of an agent's standing behavior, copy the following line into AGENTS.md, CLAUDE.md, or a system prompt:

In evidence-heavy analysis, audits, research, and consequential decisions, prefix each material conclusion with exactly one label: `[V]` verified fact supported by current evidence, `[I]` inference drawn from stated evidence, or `[R]` recommendation. Keep routine prose unlabeled; state material assumptions and uncertainty explicitly.

The Skill body and the copyable instruction are identical. The Skill adds discovery and invocation metadata; it does not define a different behavior.

Example

[V] The test suite passes on Python 3.11.

[I] The reported failure is likely specific to Python 3.12.

[R] Add Python 3.12 to CI before the next release.

License

MIT. See LICENSE.