engineering-flow-skills

Imported from yyqqCoding/engineering-flow-skills (AGENTS.md). Install upstream with npx skills add yyqqCoding/engineering-flow-skills. Copyright stays with the author.

AGENTS.md

This repository contains a portable development workflow for Codex CLI and Claude Code.

Working rules

• Treat docs/product-design.md and docs/behavior-spec.md as the product source of truth.
• Add instructions only when they correct a demonstrated model failure; remove no-op guidance.
• Keep the always-on core short. Full workflows must not be injected into every conversation.
• Keep Codex and Claude invocation metadata synchronized.
• User-invoked workflows must stay user-invoked on both platforms.
• Keep full skills user-invoked unless isolated positive, negative, and overlap benchmarks justify reopening one.
• Treat an explicitly invoked workflow as owning the same task across answers, approval, correction, resume, and compaction. End inheritance only on cancellation, an explicit workflow switch, or unrelated work.
• develop has one approval-gated entry point. Clarification answers are not implementation approval; omitted accepted behavior resumes implementation, while added scope receives an incremental checkpoint.
• Keep substantial requirement records in the project's authoritative convention or docs/requirements/<feature-slug>.md, with Draft, Accepted, and Implemented reflecting actual progress.
• Do not make commits, publish packages, or change user-global configuration unless explicitly requested.
• Use deterministic tests before model-judged tests. Behavioral benchmarks must isolate global plugins and skills.
• Never combine behavioral results across changed fixtures, scorers, plugin fingerprints, providers, models, or reasoning levels. Manually review model failures before changing instructions or scorers.

Repository structure

• docs/: product, behavior, invocation, and testing design.
• hooks/: minimal lifecycle hooks that emit the compact Core and route explicitly named workflows.
• skills/: shared Agent Skills packages.
• fixtures/: isolated repositories used by behavioral tests.
• tests/: static and behavioral test harnesses.
• scripts/: isolated benchmark runners, cohort filling, summaries, and deterministic test orchestration.

Completion

Before claiming a change complete:

• Run the smallest relevant static tests.
• Confirm plugin manifests reference every released skill.
• Confirm explicit/implicit invocation policies agree across Codex and Claude.
• Record behavioral evidence when changing skill wording or trigger descriptions.
• For a release-level stochastic claim, require at least three completed, uncontaminated samples per arm and scenario under matching benchmark and environment fingerprints.