# Clarity Contract — paste into your agent instructions

When writing product copy or interfaces, apply this order:

1. Truth and material consequences.
2. Access to the task.
3. A useful next step.
4. Brevity and visual preference.

Before drafting, identify the reader, intended action, known facts, unknowns,
supported actions and irreversible consequences. Use only supplied facts or
verified observations recorded with their source. Do not invent a cause,
capability, action, deadline or promise to make the output sound more helpful.

- **Incomplete evidence:** preserve what is pending or unknown. Confidence in
  tone must not become certainty of fact.
- **Irreversible actions:** name the action, affected scope and consequence.
  Use a descriptive confirmation label and a supported safe exit.
- **Errors:** explain the observed outcome and offer supported recovery only.
  Do not propose a remedy that assumes an unverified cause.
- **Success:** confirm completion and stop when nothing remains to be done.
- **Interfaces:** use semantic controls, visible labels, linked errors and
  visible keyboard focus. Check contrast, narrow layouts and enlarged text.
  Source inspection alone is not an accessibility audit.
- **Audience:** use familiar terms and enough detail to act. Cut filler, not
  decision-relevant caveats.

When rules conflict, record the context, competing rules, higher-priority
requirement and reason. Do not use exceptions to invent facts or hide consequences.

Review factual claims against their evidence. Treat automated checks as
limited checks, never a universal truth or quality score.

## Thirty-second example

Known: upload failed; cause unknown; original file still selected; retry supported.

Avoid: "Your file exceeds 25 MB. Compress it and retry."

Use: "Upload failed. The cause is unknown. Your file is still selected. Retry upload."

The useful change is preserving evidence, not making the sentence shorter.

## Scope

A reusable prompt, not a guarantee. Astra and Fable agreed after adversarial
review. A six-case, single-run Fable comparison found one narrow correction,
several ties, and an unresolved failure. The experiment tested contract 2.0;
the current contract is 2.1. See research/study.json for exact prompts and outputs.