the-polyglot-protocol - SKILL.md Agent Skill

name: the-polyglot-protocol description: Use when generating, reviewing, refactoring, or planning code in polyglot repositories and you need senior-engineer guidance for language choice, architecture, testing, security, performance, dependencies, concurrency, CPU/GPU use, release safety, and LLM coding guardrails.

The Polyglot Protocol

Use this skill before code generation, code review, dependency changes, architecture changes, or workflow changes in a repository that may involve one or more programming languages.

Required Workflow

Read docs/workflow/dev-workflow.md for the operating process.
Read docs/language-guidelines.md for language-selection rules.
Read docs/languages/pre-codegen-checklist.md.
Read docs/languages/do-not-generate-policy.md.
Read the relevant file under docs/languages/ before generating language-specific code.
For Python scripts, also read docs/languages/python/python.md.
For TypeScript or JavaScript, also read docs/languages/typescript/typescript.md.
After code generation, audit the entire generated codebase against this protocol and docs/languages/scoring-rubric.md. Report every criterion with its score, remediate all gaps until every criterion reaches 10/10, and report the final total score. If a criterion cannot reach 10/10, document the blocker, evidence, and follow-up requirement before completion.
Run validation when this skill folder is available:

python scripts/validate-workspace.py

Core Rules

Default operational scripts to Python unless the repository already has a better native script runner.
Preserve existing project language, framework, package manager, formatter, test runner, and layout.
Use platform-native code for product surfaces: PHP for WordPress, Shopify Liquid for Shopify themes, Kotlin for Android, Swift for Apple apps, SQL for database work, and HTML/CSS/TypeScript for web surfaces.
Prefer the shortest reliable supported path before adding infrastructure.
Verify current tool, dependency, and runtime versions from official or primary sources before pinning or upgrading.
Do not invent APIs, configuration keys, package names, services, flags, or framework behavior.
Generated code is not complete until a whole-codebase protocol audit reaches 100/100 across the rubric, or a documented blocker explains why that target is impossible in the current scope.
Treat unsupported Docker, Lighthouse, browser, deployment, Git, or CI steps as explicit N/A items with evidence.

Reference Map

Language choice: docs/languages/decision-matrix.md
Hard stops: docs/languages/do-not-generate-policy.md
Required preflight: docs/languages/pre-codegen-checklist.md
Examples: docs/languages/examples.md
Version checks: docs/languages/install-version-commands.md
Scoring: docs/languages/scoring-rubric.md
LLM failure modes: docs/languages/top-llm-coding-nuances.md
Current score: docs/languages/score-report.md

Adapter Files

Use the adapter files when another agent expects a specific guidance filename:

Codex: adapters/codex/AGENTS.md
Claude Code: adapters/claude-code/CLAUDE.md
OpenCode: adapters/opencode/AGENTS.md