SKILL.md

---
name: human-context
description: Collect missing human context through Roux and return synthesized findings to the work in progress. Use when a user asks to launch, create, share, check, or synthesize a Roux; asks teammates, customers, or stakeholders for input; needs qualitative evidence for a decision; or when Codex identifies a consequential information gap that only people can answer.
---

# Human Context

Use Roux to close human information gaps without inventing an answer or exposing
respondent identities and raw transcripts.

## Decide whether to use Roux

- Use Roux when the missing information is held by people: lived experience,
  objections, priorities, constraints, decision rationale, or local knowledge.
- Continue directly when the answer is already available in files, code, connected
  systems, or reliable public sources.
- Suggest a Roux when it would help, but only call `create_roux` after the user asks
  to launch it or approves the proposed goal and questions.

## Launch a Roux

1. State the information gap in one sentence.
2. Draft a specific outcome-oriented goal.
3. Write two to five neutral questions that a respondent can answer without extra
   context. Avoid compound, leading, or duplicative questions.
4. Use `list_roux_workspaces` when you need to confirm which workspace this
   OAuth connection can access. It returns the single workspace selected during
   authorization. If the user needs a different workspace, ask them to reconnect
   Roux and select it during authorization; never try to escape the grant.
5. Call `create_roux` with a concise title, the goal, relevant work context, the
   questions, a realistic response target, and a short duration.
6. Return the respondent link and summarize exactly who should receive it and what
   the answers will unblock.

Default to three target responses and eight minutes unless the user provides a
better constraint.

## Find and check Roux

- Use `search_roux` for a known title, phrase, description, or goal.
- Use `list_roux` for recent Roux or when the user wants to browse. Follow
  `pagination.next_cursor` when more results are needed.
- Call `get_roux_status` with the `roux_id` returned by discovery or creation.
- Report completed versus target responses and the current state.
- Do not infer who responded. Roux intentionally returns counts without respondent
  PII or raw response content.
- Do not poll repeatedly unless the user asks to monitor progress.

## Synthesize and continue the work

1. Generate synthesis only after at least one completed response and when the user
   asks for the findings or wants the original work completed.
2. Call `generate_roux_synthesis`, then use `get_roux_synthesis` once processing is
   ready.
3. Distinguish synthesized findings from the agent's own inference.
4. Apply the goal answer, themes, and open questions directly to the original
   decision or artifact. Preserve uncertainty and call out unresolved questions.

## Handle failures

- On an authorization error, ask the user to reconnect Roux; never request or expose
  access tokens.
- On insufficient credits, report the exact Roux error and let the user choose how
  to proceed.
- When there are no completed responses, return to collection instead of generating
  a speculative synthesis.