Start Here

Step 0 — Load shared governance

Use the website’s single asset dialog and compact copy control, or use scripts/compose_assets.py. The dialog keeps the goal, “Use this when”, “Do not use this when”, and linked research outside the literal Markdown body. The copied output contains only the required inputs, specialist instructions or procedure or hard gates, evidence and failure handling, applicable shared requirements once, and shared operating rules. Pass --with-references only when a review artefact needs citations appended. Use RESEARCH-DNA.md when you need the source map, PROMPT-PATTERN-MATRIX.md for task selection, and HUMAN-AI-QUALITY-STANDARD.md for sociotechnical and trust-calibration review.

Use this page when you need the shortest safe route through the library.

For a new Codex task

1. Read AGENTS.md and docs/engineering/workflow.md, then use the complete composed agent body.
2. Choose the narrowest relevant prompt; use the master prompt only for broad cross-cutting work.
3. Add skills/01-scoping-packet-skill.md if the task is non-trivial.
4. Attach contracts/01-completion-honesty-contract.md so final claims can be accepted or rejected.
5. Resolve the selection with python3 -S scripts/compose_assets.py --asset <id> or the website copy action.
6. Run the relevant commands from requirements.toml.
7. End with one status value: verified, partially verified, not verified, or blocked.

For UI work

Use:

• prompts/21-ui-structure-audit-prompt.md
• prompts/22-visual-design-quality-gate-prompt.md
• prompts/26-keyboard-journey-audit-prompt.md
• contracts/10-ui-visual-quality-contract.md
• contracts/08-keyboard-interaction-contract.md

For release claims

Use:

• prompts/33-release-evidence-prompt.md
• skills/05-verification-reporting-skill.md
• contracts/19-release-evidence-contract-plus.md

For template-library maintenance

Use:

• prompts/36-github-pages-template-hub-prompt.md
• skills/24-template-catalogue-design-skill.md
• contracts/29-template-library-quality-contract.md

Evidence-oriented use order

Use the assets in this order when quality matters:

1. Start with the master agent prompt to set the operating frame.
2. Use least-to-most decomposition to break the task into small verifiable pieces.
3. Use ReAct discipline: observe files/UI/tests before acting, then observe results after acting.
4. Use branch evaluation when the first solution is not obviously safe.
5. Attach the relevant skill for the domain task.
6. Attach the relevant contract before claiming acceptance.
7. Fill the traceability matrix before final status.
8. End with verified, partially verified, not verified, or blocked.

Never ask for hidden chain-of-thought. Ask for a public decision record and evidence trail.