Claude Skills

A collection of Claude Code skills. Each skill is a markdown file you drop into ~/.claude/skills/ and invoke with a slash command. No dependencies, no installs.

Skills

Skill	Command	What it does
deep-dive	/deep-dive	DAG-based deep research. Breaks questions into dependency-ordered sub-questions, runs parallel subagents, finds gaps, writes sourced reports.
news-from-discussions	/news-from-discussions	Morning briefing tailored from your past conversations and memory — checklists across your projects + a "For You" news digest based on what you've actually been thinking about, not topic tags.

Install

Install a single skill:

mkdir -p ~/.claude/skills/deep-dive
cp deep-dive/SKILL.md ~/.claude/skills/deep-dive/SKILL.md

Install all skills:

git clone https://github.com/kimsb2429/claude-skills.git
for skill in claude-skills/*/; do
  name=$(basename "$skill")
  mkdir -p ~/.claude/skills/"$name"
  cp "$skill"SKILL.md ~/.claude/skills/"$name"/SKILL.md
done

Skills in detail

deep-dive

A deep research skill based on how Google's Deep Research works. It builds a dependency graph (DAG) of sub-questions, researches them in parallel using subagents, identifies gaps, follows up, and writes a final report.

Most research skills search, then summarize. Deep Dive builds a dependency graph first:

• Questions that depend on other answers wait for those answers
• Independent questions run at the same time (3-6 subagents in parallel)
• Context flows forward. Later questions get findings from earlier ones
• The research plan adapts based on what's actually found

Out of 7+ community alternatives surveyed, this is the only Claude Code skill that uses true DAG-based planning.

Usage:

/deep-dive What are the best open-source alternatives to Datadog for monitoring microservices?

What you get:

• DAG plan printed as a table before anything runs
• Real-time progress via task checklist
• Gap iteration that catches what the first pass missed
• Final report saved to docs/deep-dive/ with executive summary, themed findings, open questions, and sources

A typical run produces a 500-2000 line report covering 4-8 sub-questions, citing 15-40 sources, in 5-15 minutes.

Feature	Deep Dive	Typical research skills
Query planning	DAG with dependencies	Linear pipeline
Parallelism	Wave-based (3-6 concurrent)	Sequential or basic parallel
Gap iteration	Yes (1 round)	Rarely
Progress tracking	Task checklist	None
Output	Persistent markdown report	Chat response
External deps	None	Often requires MCP servers or APIs

news-from-discussions

A personal morning briefing that derives "For You" picks from your own files — memory, project TODOs, anything in ~/.claude/projects/*/memory/. The hosted alternatives (ChatGPT Pulse, etc.) personalize from chat history and inboxes; this one personalizes from the structured notes you've already been writing about who you are and what you're building.

Two phases:

• Checklists — auto-discovers any direct subdir of ~/ containing a TODO.md, prints pending-first checklists per project
• News digest — 6 parallel WebSearch queries: 2 personalized "For You", general news, AI & tech, your dominant niche (inferred from memory), and community buzz (HN/Reddit/X). Suggested actions get appended to project TODOs after approval.

Usage:

/news-from-discussions

Pairs well with Claude Code Routines for scheduled morning delivery. Add --out ~/brief.md (or set NEWS_FROM_DISCUSSIONS_OUT) to write the briefing to a file, then wire delivery however you want:

# Send via local mail after Routines fires
mail -s "Morning brief" you@example.com < ~/brief.md

# Or push to a Slack webhook
curl -X POST -H 'Content-type: application/json' \
  --data "{\"text\": \"$(cat ~/brief.md)\"}" \
  $SLACK_WEBHOOK_URL

# Or commit to a private repo for cross-device read
cd ~/briefs && git add . && git commit -m "$(date +%F)" && git push

Differentiator vs. existing morning-brief tools: cross-project TODO aggregation as the spine, and personalization derived from already-written memory rather than asking you to configure topic tags.

Requirements

• Claude Code CLI (any version with subagent support)
• No external APIs, MCP servers, or dependencies
• Works with any Claude model (Opus, Sonnet, Haiku)

License

MIT