advisor

Advisor

You are being asked for an opinion grounded in accumulated experience, not generic best practice. Treat the prompt as a request to reflect, not to act.

When this skill applies

Triggers (case-insensitive substring match; the frozen list lives in src/advisor/detect.js):

• "would we change"
• "what would you do differently", "what would we do differently"
• "how can we improve", "how would you improve"
• "what are we missing"
• "is there a better way"
• "anything you'd change", "anything you would change"
• "review this approach", "second opinion"
• "how would you approach"
• "what would you change"
• "do anything different", "do differently"

Confirm before answering if it's ambiguous. If the user is asking for an action (e.g. "rename X") and not a reflection, answer that and skip this skill.

Inputs to assemble first

Before drafting any recommendation, gather this evidence — call these tools in this order:

1. Recall memories — memory_recall(scope: "all", query: <synthesised from the prompt>). Pull both the global and project DBs in one call so the merge is correct. Note each id.
2. Get the current focus — working_memory_get(slot: "current_focus") and working_memory_get(slot: "active_task"). If they are empty, the advisor context is genuinely fresh — note that explicitly.
3. Search recent sessions for the same theme — conversation_search(query: <theme>). Skip if the conversation is the first turn.
4. Read the live project state — at minimum:
   • The active project's top-level layout (one ls of cwd).
   • For code repos: the package.json / pyproject.toml / Cargo.toml and a single representative source file.
   • For config projects (like this one): the root config.toml, tui.toml, mcp.json, and the plugins/ directory.

The procedure requires all four inputs. If any source is empty, treat that as a data point — write it down — not a skip.

Reasoning procedure

The full step-by-step is in references/procedure.md. Short form: