# Discuss

> Ask clarifying questions and discuss direction before starting work. Use when the user invokes $discuss or says "discuss-first", wants to discuss requirements, clarify scope, or asks Codex to ask questions before proceeding with a task. Natural phrasings include "let's discuss before proceeding", "ask me questions for clarification or direction", and "let's discuss and make a plan first". Do not use for general requests to explain or discuss a topic when no pending task is being held back.

- Skill: `maaarcooo/discuss` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add maaarcooo/discuss`
- Raw SKILL.md: https://api.skillmd.com/api/skills/maaarcooo/discuss/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: maaarcooo (https://skillmd.com/u/maaarcooo)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/maaarcooo/discuss

---


When invoked, treat this as the start of a genuine discussion about the
task, not a clarification checklist.

Before doing any work:

1. **Ground yourself first, if the task touches existing code.** When
   the task references existing code, files, or project structure,
   briefly inspect the relevant parts before asking anything, so your
   questions cite real specifics ("you currently do X in `foo.ts`, do
   you want this to change?") rather than abstractions. Keep the scan
   light — enough to ask informed questions, not a full audit. Skip
   this for greenfield or non-code tasks.
2. **Share your initial read.** State what you understand the goal to
   be, the assumptions you are making, and anything that looks like a
   key decision or trade-off. Engage with the specifics of this task.
3. **Ask focused questions that target real forks in direction.**
   Prefer "do you want X or Y, because it changes Z" over generic
   intake questions. Skip anything already answered in the
   conversation or visible in the codebase.
4. **Attach a tentative opinion or recommendation where you have one**,
   so I can react to it rather than answer cold.

Delivery style is your judgement. Default to one small batch of 3-5
questions. Go one or two at a time only when later questions genuinely
depend on earlier answers.

Do not start the task until I have responded. After my answers:

- For complex or multi-step tasks, propose a concise plan and wait for
  confirmation.
- For simple tasks, proceed directly once clarified.
- If the user includes "plan" with the invocation (`$discuss plan`), always
  propose a plan after gathering answers, regardless of task complexity.
- Either way, when you proceed, state any remaining assumptions in one
  line (e.g. "proceeding with X and Y assumed") so nothing slips
  through silently.

Tone: conversational, like a collaborator thinking through the problem
with me, not a form to fill.