advisor-mode

advisor-mode

A skill that turns an AI coding assistant into a blunt advisor instead of an eager assistant. It challenges your assumptions, rates its own confidence, refuses flattery, and holds its position under pushback unless you give it genuinely new information.

This repo is a Claude Code plugin marketplace — install with /plugin, no cloning. It's also Codex-compatible: the same skill folder drops into ~/.agents/skills/.

---

Before / after

You: "I want to store session tokens in localStorage, it's easiest. Help me wire it up."

Plain assistant: "Great choice! localStorage is simple to use. Here's how to set it up…"

advisor-mode: "No — localStorage is readable by any JS, so one XSS becomes full account takeover. [Certain] Use httpOnly cookies. The risk in your approach: an injected script exfiltrates every active session and replays it after the user closes the tab."

---

Install

Claude Code (recommended — no clone)

/plugin marketplace add daksh7011/advisor-mode
/plugin install advisor-mode@advisor-mode

That's it. The skill auto-loads on relevant prompts, or invoke it by name with /advisor-mode:advisor-mode.

To update later: /plugin marketplace update advisor-mode.

Codex (or any agent that reads ~/.agents/skills/)

Plugins are Claude-Code-only, so Codex users copy the skill folder directly:

git clone https://github.com/daksh7011/advisor-mode.git /tmp/advisor-mode
cp -r /tmp/advisor-mode/plugins/advisor-mode/skills/advisor-mode ~/.agents/skills/advisor-mode

No edits to your CLAUDE.md / AGENTS.md / system prompt required either way.

Use

• Auto-load — fires when you ask for honest/blunt feedback, a devil's advocate, or are reviewing a decision, plan, or claim.
• Force it on — invoke by name (/advisor-mode:advisor-mode in Claude Code). Once active it applies to every reply for the rest of the session.

The eight rules it enforces

1. Never start with agreement — the first sentence challenges.
2. Rate confidence on every claim: [Certain] / [Likely] / [Guessing].
3. Kill flattery filler ("good question", "you're absolutely right", "absolutely", "definitely").
4. Disagree with structure: reason + alternative + specific risk.
5. Uncomfortable answer first — never buried in paragraph three.
6. No warm-up paragraphs.
7. Don't fold under pushback that contains no new information.
8. Don't assume — get the facts first; never present an assumption as fact.

Full text, rationalization counters, and red-flags checklist: see SKILL.md.

Verify it works

bash scripts/verify-flip.sh

Prints two prompts — one without the skill, one with — plus a checklist so you can watch the behavior flip. Run it on a machine whose global config does not already contain advisor-style rules, or the control will look identical to the treatment.

Limitation — read this

A skill is trigger-loaded, not a permanent system prompt. It cannot guarantee the behavior on literally every message regardless of topic. For always-on behavior, invoke it by name at the start of a session, or put the seven rules directly in your CLAUDE.md / AGENTS.md. This plugin is the portable, shareable version of those rules — one command, no system changes.

Repo layout

advisor-mode/                                 ← marketplace (repo root)
├── .claude-plugin/marketplace.json
├── plugins/
│   └── advisor-mode/                         ← the plugin
│       ├── .claude-plugin/plugin.json
│       └── skills/advisor-mode/SKILL.md      ← the skill (copyable for Codex)
├── scripts/verify-flip.sh
└── LICENSE

License

MIT