Deep Research Skill for Claude

/deep-research — Multi-source research with archive

Plans, executes, and synthesizes research from multiple sources. Archives the output so the corpus compounds.

Step 1 — Frame the question

Restate the research question in one tight sentence. If ambiguous, ask the user:

• What's the decision this research will inform?
• What's the minimum useful answer? (Saves over-researching.)
• Any sources to prioritize or avoid?

Output: Research question: <one sentence>

Step 2 — Plan the sources

Pick from this menu based on the question type. Note which sources you'll hit and why.

| Source | When to use | Tool |

|---|---|---|

| Web search (Google) | Authoritative articles, docs, official statements | WebSearch |

| /last30days | What people are *actually saying* right now — Reddit, X, YouTube, HN, web recency | Skill({skill: "last30days", args: "<topic>"}) |

| Specific URLs | When the user hands over starting URLs | WebFetch |

| Browsable pages (auth-walled, JS-heavy) | Pricing pages, product tours, profiles | agent-browser via the compound-engineering:agent-browser skill |

| Memory | Prior research / decisions / context the user already captured | grep ~/.claude/memory/ |

| Notion | If the topic touches a known Notion workspace | Direct Notion API (key in $NOTION_API_KEY, see reference_notion_api.md) |

| Research archive | Prior /deep-research runs that touched this topic | grep ${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/deep-research/archive/ |

Run discovery passes in parallel where possible. Sequential only when one source needs another's output (e.g., agent-browser a URL discovered by WebSearch).

Step 3 — Execute discovery

Run each chosen source. For each result, capture:

• The source (URL or system)
• 1–3 sentence summary of what was said
• Date / recency
• Confidence in the source (high/medium/low)

Don't synthesize yet — just collect.

Step 4 — Synthesize

• Group findings by theme or sub-question
• Contradiction check — flag anywhere sources disagree. Don't average them; surface the disagreement.
• Confidence: high (multiple independent sources agree), medium (one strong source or several weak), low (single anecdote or speculation)
• Gaps: what would change the answer? What's NOT in the corpus?

Step 5 — Output the brief

Use this template:

# Research: <question>

**Date:** <YYYY-MM-DD>
**Decision this informs:** <one line>
**Confidence overall:** high / medium / low

## TL;DR
<2–4 sentences with the answer>

## Key findings

### 1. <Finding>
<2–4 sentences>. Sources: [1], [3], [5]

### 2. <Finding>
...

## Contradictions / uncertainty
- <where sources disagree, with each side cited>

## Gaps
- <what's missing from the corpus>
- <what to research next to close the gap>

## Recommended next steps
1. <action>
2. <action>

## Sources
[1] <Title> — <URL or system> (<date>) — <confidence>
[2] ...

Step 6 — Archive

Archives live in ${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/deep-research/archive/ (create the directory if missing). Never write archives inside the skill's own folder — skill installs and upgrades re-sync from source and wipe anything saved there. Migration: if this skill's folder contains an old references/research-archive/ with user entries, move those files into the archive directory first.

Write the brief to <archive dir>/<YYYY-MM-DD>-<slug>.md so it's grep-able forever. Slug = kebab-case of the topic.

Also append a one-line entry to <archive dir>/INDEX.md (create if missing):

- 2026-06-15 — [<topic>](./<filename>.md) — <one-line TL;DR>

Step 7 — Surface

After archiving:

• Show the full brief in chat
• Tell the user the archive path
• Offer: *"Push to Notion or save to a project's docs?"*

Composes with

• business-brainstorm — calls this skill during the market validation step
• /domain — when research includes "is the .com available"
• /last30days — one of the data sources

Notes on quality

• Always cite. Every claim in the brief needs a source pointer.
• Recency matters — note dates on each source. For fast-moving topics (AI, startups), de-weight sources >12 months old.
• Don't trust a single source for high-stakes claims. Re-search until you have at least 2 independent corroborations or surface the uncertainty.
• No padding. If the answer is one paragraph, return one paragraph. The template is a maximum, not a minimum.