# Collab

> Work with the human to answer, investigate, or deliver an authorized task; clarify material unresolved choices, keep progress visible, and carry implementation through verification and delivery. Use when explicitly asked for the collab workflow.

- Skill: `possibilities/collab` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add possibilities/collab`
- Raw SKILL.md: https://api.skillmd.com/api/skills/possibilities/collab/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: possibilities (https://skillmd.com/u/possibilities)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/possibilities/collab

---


# Collab

Work with the human to carry their intended task through to a verified result.
Adapt the conversation and amount of planning to the work.

## Read the request

<!-- fragment: read-the-request.md -->

When context resolves an ambiguity, state a useful assumption and proceed.
Ask a focused question when different readings would materially change the
result and the available evidence cannot choose between them. Continue
independent work while the answer is pending.

## Act within the scope

<!-- fragment: working-scope.md -->

For a question or synthesis, give the answer in the requested form. For a
small change, state the change briefly and do it. For substantial work, share
a concise direction and proceed within authorization. Do not turn a clear
directive into a mandatory sketch approval round.

Read [Building and delivery](references/building-and-delivery.md) when the
request involves software implementation, integration, or installation. It
contains the sketch format, shared-checkout rules, domain conventions, and
operator preferences. Use the appropriate specialist skill for the actual
work rather than loading every available runbook.

## Stay in conversation

Read follow-ups for decision content: apply corrections, carry prior
approval, and keep the active objective accounted for. A short answer can
resolve an open choice without a special approval keyword.

Keep progress updates useful and brief: findings, decisions, and what the
next step will resolve. Use the native question or handoff mechanism when a
human decision is needed. Bundle related missing facts when that makes the
question easier to answer.

End with a self-contained account of the result, verification, and anything
still pending. Match claims to what actually completed. When no work remains,
say so; do not declare completion while an authorized delivery step remains.

