Lz Advisor

lz-advisor

Claude Code plugin implementing the advisor strategy -- pair an Opus advisor with your session model for near-Opus intelligence at Sonnet cost.

Overview

The advisor strategy pairs a stronger model (Opus 4.7, auto-selected via the opus alias) with a faster model (typically Sonnet 4.6) to get the best of both worlds. The faster model handles the bulk of token generation -- reading files, writing code, running tools -- while the stronger model provides strategic guidance at high-leverage moments: before committing to an approach, when stuck, and before declaring done.

Anthropic's internal benchmarks (measured on Opus 4.6) show that Sonnet 4.6 paired with an Opus advisor achieved +2.7 percentage points on SWE-bench Multilingual coding benchmarks at 11.9% lower cost compared to Sonnet solo. The advisor adds most value on the first call, before the approach crystallizes, and on the final check after work is complete.

This plugin uses Claude Code's native Agent tool to implement the pattern -- no API keys, no external dependencies, no additional setup beyond installation. The skills consult one of three Opus agents at strategic moments: advisor (used by /lz-plan and /lz-execute), reviewer (used by /lz-review), and security-reviewer (used by /lz-security-review).

Skills

Skill	Description
/lz-plan	Get an Opus-informed strategic plan before writing code. The executor orients on the codebase, consults the advisor for strategic direction, then produces an actionable plan.
/lz-execute	Execute tasks with strategic Opus consultation at high-leverage moments. The executor works through the task, consulting the advisor before substantive work, when stuck, and before declaring done.
/lz-review	Opus-powered code quality review of completed work. The executor scans the changes and packages context for the advisor to provide deep quality analysis.
/lz-security-review	Opus-powered security-focused threat analysis. The executor identifies attack surfaces and packages findings for the advisor to assess threats and recommend mitigations.

Installation

Add the marketplace and install the plugin:

/plugin marketplace add LayZeeDK/lz-advisor-claude-plugins
/plugin install lz-advisor@lz-advisor-claude-plugins

For local development or testing from a cloned repo:

claude --plugin-dir plugins/lz-advisor

After installation, verify the plugin loads correctly:

claude --debug

Look for lz-advisor in the loaded plugins output. If skills do not trigger, verify that lz-advisor appears in enabledPlugins in your Claude Code settings.

How it works

• Skills run on your session model (typically Sonnet 4.6) as the executor.
• At strategic moments, the executor consults the relevant Opus agent -- advisor for /lz-plan and /lz-execute, reviewer for /lz-review, security-reviewer for /lz-security-review.
• The advisor provides concise guidance -- under 100 words, enumerated steps focused on what to do.
• The executor continues with the task, informed by the advice.

The advisor is consulted at the moments where a stronger model adds the most value:

1. Before substantive work -- after orientation (reading files, understanding the codebase), the executor consults the advisor before writing any code or committing to an approach.
2. When stuck -- if errors recur, the approach is not converging, or results do not fit expectations.
3. Before declaring done -- a final check to verify the approach is sound and catch anything the executor may have missed.

Requirements

• Claude Code with access to Sonnet 4.6 or later and Opus 4.7 or later (the opus alias auto-resolves to the current Opus generation)
• No API keys or external dependencies required

Opus 4.7 (released 2026-04-16) is auto-selected via the opus alias; no user action is required to adopt it.

References

• references/advisor-timing.md -- Anthropic's suggested timing patterns for advisor consultation: when to consult, how much weight to give the advice, and how to package context for a fresh-context advisor call.
• references/context-packaging.md -- Shared packaging contract and templates (Proposal, Verification) used by all four skills. Defines the Common Contract (4 packaging rules), the Proposal template for strategic direction calls, the Verification template for review calls, a short-form Proposal variant for mid-execute re-consultations, and a decision table mapping skill + call site to template.

What's New

2.0.0

Breaking release. The four skills are renamed with an lz- prefix -- /lz-plan, /lz-execute, /lz-review, /lz-security-review (qualified lz-advisor:lz-<skill>) -- so they no longer shadow Claude Code's built-in /plan, /review, and /security-review. The /lz-security-review report also adopts the canonical 5-tier security severity taxonomy (Critical / High / Medium / Low / Informational, plus Open Questions). See CHANGELOG.md for the full old-to-new migration table.

License

MIT