# Ctx Research

> Deep research grounded in the project's durable context system. Use for /ctx-research, "ctx research", "research X starting from our contexts". Reads existing contexts first, sharpens the question to the real knowledge gap, fans out multi-agent research, verifies key claims, and files results back.

- Skill: `keli-wen/ctx-research` (Agent Skill)
- Install (CLI): `npx skillmds@latest add keli-wen/ctx-research`
- Raw SKILL.md: https://api.skillmd.com/api/skills/keli-wen/ctx-research/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: keli-wen (https://skillmd.com/u/keli-wen)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/keli-wen/ctx-research

---


# ctx-research

Research that starts from what the project already knows and ends with the project knowing more. It differs from generic deep research in two ways: the question is sharpened against existing contexts before any searching, and verified findings are written back into the context system.

**Language**: mirror the user's language.

## Steps

1. **Ground**: locate the project's durable context layer (e.g. a `contexts/` tree with a map file, or a docs/ADR system). Read the entries relevant to the question. Sort what you find into: already answered (do not re-research), assumed but unverified, genuinely open.
2. **Refine**: restate the research question as the delta between what the user needs and what contexts already hold. Confirm scope with the user only if the delta is ambiguous; otherwise proceed.
3. **Fan out** (using `/fan-out` patterns): parallel researchers with distinct angles — official docs, community and prior art, papers, code in the wild. Each returns raw, citable findings with source name and URL or path.
4. **Verify**: every claim that will drive a decision gets an adversarial pass — a separate check that tries to refute it or show it is stale. Mark key claims confirmed / plausible / refuted.
5. **Synthesize**: one report — answer first, then evidence, then open questions. Keep cited facts and your own judgment visibly separate.
6. **File back**: hand the durable part to `/to-ctx`, which places it in the project's context system per its rules. The chat gets the summary plus the file path.

## Rules

- Never re-research what contexts already answer; cite the context file instead.
- Verification effort scales with decision impact, not with how interesting a claim is.
- Dependency direction: may use `/fan-out`, `/to-ctx`, and `/handoff`; never invokes user-facing skills.