vibe-suite

Design advisor agents — file format, value-driven system prompts, model tier selection, tool and working-directory scoping, budget/turn caps, and choosing between a consultative advisor and a batch subagent. Use when creating, editing, or reviewing an advisor definition.

The judgment criteria /vibe-suite:nl-audit audits by — seven dimensions for each of five NL-artifact types (skill, command, agent, rules, plugin) with their mini/full depth membership, plus fifteen category check sets for whole-repository audits. Load this when auditing NL artifacts for quality, when deciding which dimension a defect belongs to, or when a mini run must be distinguished from a full one. Judgment criteria only — the deterministic penalty tables are the scoring skill's.

Tool-agnostic conventions floor for natural-language artifacts — SKILL.md open spec, AGENTS.md canonical memory, prompt layering, vague quantifiers, naming, rule overrides, and MCP tool naming; per-tool schemas live in overlay skills.

Overlay of Antigravity (plus legacy Gemini CLI) conventions — workspace skills under .agent/, the .gemini/ paths, gemini-extension.json, GEMINI.md imports, slash commands in TOML, and the Gemini-lineage hook events; the spec has not settled since Antigravity 2.0, so most tool-specific checks stay advisory.

Claude Code overlay on the conventions floor — schemas and conventions for plugin.json, commands, skills, agents, rules, hooks, .mcp.json, marketplace.json, CLAUDE.md, memory, and settings at .claude/ canonical paths.

Overlay of Codex CLI conventions — the config.toml grammar, the .codex-plugin/plugin.json manifest, the .agents/skills/ layout, the agents/openai.yaml sidecar, hook events, the AGENTS.md hierarchy, and marketplace.json.

Drive a tracked work item to a reviewed pull request through a nine-step, three-phase pipeline: analyze, plan, execute, each phase a worker pass followed by an independent review and a bounded update-and-verify loop. Every project-bound fact comes from a profile, so the machinery is the same whichever repository it points at. Use when an issue should become a reviewed PR with its reasoning recorded, not when a quick edit will do.

Workflow patterns for coordinating multiple agents in Claude Code — sequential pipelines, parallel dispatch, retry loops with QC gates, plus shared partials. Use when the system being designed spans several agents, commands, or processing stages.

Use when authoring or reviewing NL programming artifacts (agents, skills, commands, rules, hooks) — applies proven patterns and flags anti-patterns such as monolithic prompts, oversized skills, vague quantifiers, Write/Edit on read-only agents, prohibitions without alternatives, and linter-duplicating rules.

Drive a written proposal through bounded rounds of independent review until it converges: freeze the input, produce a baseline, then review and revise on a shared rubric until the verdict is clean or the cap stops it. Use when a plan needs adversarial review rather than editing, or when it should be delivered bilingually. Not for reviewing code, and not for writing a first draft.

The criteria /vibe-suite:roast reviews code by — the nine cc-suite audit dimensions for the cross-model lanes, the separate five-dimension mini set, the six report styles and eight add-ons, and the reconciliation labels for --engine both. Load when running a roast, when deciding which dimension a code defect belongs to, or when a mini pass must be distinguished from a full one. Code review, not NL-artifact review: that is the auditing skill's.

The 51 rules of natural language programming — the style guide NL code quality is judged against. Apply when you write, review, or improve an NL artifact of any kind: a skill, agent, command, rule file, hook, prompt, plugin, or project memory file (whether CLAUDE.md, AGENTS.md, or GEMINI.md).

Generate time-bucketed static HTML statistics dashboards over the issue2pr runs/ tree: all-time plus day/week/month reports, index.html, and history.json under runs/_reports/, with vendored Chart.js inlined so charts render offline from file://. Past buckets are frozen snapshots refreshed once when their data changed (--force-regenerate/--period rebuild surgically); buckets follow --tz (default Asia/Shanghai). Requires --id-pattern from the resolved issue2pr profile and refuses without it; reviewer token panels are labeled from run metadata, never a hardcoded model name. Ad-hoc filters write one isolated report, never touching canonical history. Use when asked to summarize, visualize, or track issue2pr runs, AI-review activity, or token usage over time.

Score NL artifact quality on the 100-point rubric — deterministic per-type penalty tables, score bands, and multi-tool tier classification. Use when scoring artifacts, applying penalties, or calibrating lint judgment; borderline cases can pull the four worked examples (Weak Rule, Excellent Rule, Rewrite Agent, Excellent Agent) from references/calibration-examples.md on demand.

Shared security pattern database for plugin scanning — severity-rated signatures, context-aware capping rules, and PASS/REVIEW/BLOCK gate semantics.

Use when authoring test specs for NL artifacts (agents, skills, commands, rules, hooks, prompts), when running /vibe-suite:test, or when adopting TDD workflows for natural language programming.

The finding contract every vibe-suite reviewing artifact shares — severity scale and definitions, effort classes, the six-field finding format, output-header contract, zero-findings rule, anti-padding rules, and the untrusted-input and secret-handling rules. Load this before producing or consuming any audit finding.

Vocabulary discipline for suite artifacts — canonical/deprecated term semantics, scope and homonym rules, drift detection, and the registry file format.

Authoring guide for Claude Code agents — making them trigger reliably, run on the right model tier, and emit consistent output. Use when an agent definition is being created, improved, or reviewed.

Guide to writing hooks for Claude Code — picking the right event, the three hook types, matcher patterns, blocking versus advisory decisions, and portable paths. Use when a quality gate, automation step, or policy-enforcement rule calls for a hook.

How to plan, build, and review a plugin — deciding architecture, selecting components, laying out files, configuring the plugin.json manifest, versioning, and publishing to a marketplace, for Claude Code with mappings to Codex CLI and Antigravity.

How to engineer prompts, system instructions, and AI configuration for ANY LLM — universal guidance on role clarity, structured output, few-shot examples, and injection resistance.

How to author, improve, and review project rules — .claude/rules/ files that the model actually follows, covering the enforceability test, token budgets, positive framing, and path scoping.

Use when creating, improving, or reviewing SKILL.md files for any tool — descriptions that trigger reliably, body structure that teaches effectively, and progressive disclosure via references/.