Collection of Agent Skills for our internal teams.
npx skills add inkeep/team-skills/plugins/shared -y
Use by typing e.g. /skill-name in Claude Code or Cursor, e.g. /research.
Shared skills include:
| Example | Description |
|---|---|
/research <topic> |
Evidence-driven research with formal reports and evidence files collected from high authority web sources, inspecting open source code repos, docs, research articles, etc. |
/write-skill <goal> |
Create or refine/revise Agent Skills (SKILL.md + references/scripts/assets) from existing knowledge or giving it iterative feedback. |
/write-agent <goal> |
(advanced) Design and write Claude Code agents and prompts. |
Each team plugin (eng, gtm) is a proper Claude Code plugin with its own .claude-plugin/plugin.json which contains all the shared skills in addition to team-specific skills.
Two install methods. Choose based on your setup:
| Claude Code Plugin (recommended) | Skills CLI | |
|---|---|---|
| Auto-update | Yes — skills update every session | No — manual npx skills update |
| Agent support | Claude Code only | Claude Code, Cursor, Cline, Codex, etc. |
| Best for | Claude Code-only users who want zero-maintenance updates | Multi-agent users (e.g., Cursor + Claude Code) |
Auto-updates on every session. Skills stay current without manual intervention.
# Add the marketplace with auto-update (one-time)
claude plugin marketplace add https://github.com/inkeep/team-skills.git && node -e "const f=require('os').homedir()+'/.claude/plugins/known_marketplaces.json',d=require(f);d['inkeep-team-skills'].autoUpdate=true;require('fs').writeFileSync(f,JSON.stringify(d,null,2));console.log('Auto-update enabled for inkeep-team-skills')"
# Install your team's plugin
claude plugin install eng@inkeep-team-skills # engineering
claude plugin install gtm@inkeep-team-skills # GTMWorks with any agent that supports the Agent Skills standard. Updates are manual.
# Install a team's skills
npx skills add inkeep/team-skills/plugins/eng -y # engineering
npx skills add inkeep/team-skills/plugins/gtm -y # GTM
# Or install shared skills only
npx skills add inkeep/team-skills/plugins/shared -yPlugin users: Skills auto-update on every Claude Code session — no action needed.
Skills CLI users: Run manually when you want the latest:
npx skills updateThe two most powerful shared skills.
YOU
|
┌──────────────┴──────────────┐
| |
"How does X work?" "I need agents to be
"How do others do Y?" great at doing X"
| |
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ /research │ │ /research │
│ │ │ │
│ Prior art & │ │ Gather the best │
│ deep dives │ │ human knowledge │
└────────┬─────────┘ └────────┬─────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ REPORT │ │ REPORT │
│ ├ REPORT.md │ │ ├ REPORT.md │
│ └ evidence/*.md │ │ └ evidence/*.md │
└────────┬─────────┘ └────────┬─────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ Read the report │ │ /write-skill │
│ or ask Claude │ │ │
│ questions │ │ Distill into a │
│ about it │ │ SKILL.md agents │
└────────┬─────────┘ │ can execute │
│ └────────┬─────────┘
▼ │
┌─────────────────┐ ▼
│ Want to go │ ┌─────────────────┐
│ deeper? │ │ SKILL │
│ │ │ ├ SKILL.md │
│ /research ─────┐│ │ ├ references/ │
│ with follow-up ││ │ ├ scripts/ │
└─────────────────┘│ │ └ templates/ │
▲ │ └────────┬─────────┘
└──────────┘ │
report is refined ▼
& expanded ┌─────────────────┐
│ Use the skill │
│ in a real │
│ session │
└────────┬─────────┘
│
▼
┌─────────────────┐
│ Notice a gap? │
│ │
│ /write-skill ──┐│
│ with feedback ││
└─────────────────┘│
▲ │
└──────────┘
refine until sharp
Use case 1: Prior art research. How does Stripe handle webhooks? How does Linear model project hierarchies? What retry strategies does our queue library actually support? /research digs through docs, OSS code repos, and articles so you get a sourced report instead of guessing or tab-hopping.
Use case 2: Skill generation. Same /research step as above, but then /write-skill distills the findings into a SKILL.md that agents can execute. The report is raw analytical knowledge; the skill is the operationalized workflow from that knowledge.
Invoke by typing /research <topic> within Claude Code or Cursor.
# New research (defaults to formal report)
/research How does Temporal handle workflow versioning?
# Quick answer (no report)
/research Just tell me — does Better Auth support SCIM?
# Extend an existing report
/research Update the claude-plugins-architecture report with Cowork CLI parity findings
# Compare systems
/research Compare pg-boss vs BullMQ for job queues — focus on persistence, retry, and observability
Conducts evidence-driven research and produces one of three outputs:
- Formal report — persistent artifact in
~/.claude/reports/<name>/with evidence files. Default for non-trivial research. - Report update — refinement or additions to an existing report. Triggered when you're iterating on an existing report or mention it e.g.
update X report with new research on XYZ. Automatically figures out best way to update the report. - Direct answer — findings delivered in conversation. Used when you say "just tell me" or it's a quick question.
Reports live in ~/.claude/reports/<name>/ with:
REPORT.md— synthesized findings with executive summary, rubric, detailed sectionsevidence/*.md— primary-source proof files (one per dimension)meta/_changelog.md— append-only history of updates
To open the directory of reports in Cursor:
cursor ~/.claude/reports or simply navigate to it.
In MacOS, while in a Home directory (e.g. nickgomez/, click on cmd + shift + . to see hidden .claude folder.)
- Checks existing reports first. Before starting new research, it scans
~/.claude/reports/for overlap. If prior research covers a topic, it surfaces those findings and asks whether to reuse, extend, or start fresh. - Scopes before researching. It proposes a research rubric (dimensions, depth, stance) and waits for a confirmation before diving in. You can adjust scope, add/remove dimensions, or change the output format.
- Evidence-backed. Every finding links to an evidence file with primary sources (code snippets for OSS repos, doc quotes, research studies, etc.). Claims are labeled CONFIRMED / INFERRED / UNCERTAIN / NOT FOUND. Auto-prioritizes by time and authority.
- Recaps and suggests follow-ups. After delivering findings, it summarizes key results and offers 2-4 natural next directions.
# Create a new skill from scratch
/write-skill Create a skill for writing cold outbound emails called 'write-email'
# Create a new skill from a report
/write-skill Create a skill for writing cold outbound emails by looking at the 'b2b-outbound-email' report and evidence. Help me turn that knowledge into a skill 'write-email'.
Invoke with /write-skill <goal> when you want to create or modify an Agent Skill.
Guides you through authoring a SKILL.md (+ optional references/, scripts/, templates/) that works across Claude Code, Cursor, and other agents supporting the Agent Skills standard.
- Asks clarifying questions first. Captures intent, audience, constraints, and success criteria before drafting. For ambiguous requests, it offers 2-4 targeted questions with recommended defaults.
- Outputs the full skill. Delivers folder tree + complete file contents. No partial stubs.
- Use to update a skill as well! If you give it feedback on how a skill could behave better (e.g. after trying out a real skill elsewhere), it'll procedurally update the skill in a conservative way to align it with what you describe as the correct behavior.
skill-name/
├── SKILL.md # Main skill (frontmatter + workflow)
├── references/ # Deep-dive docs loaded on demand
├── scripts/ # Executable utilities
└── templates/ # Reusable output templates
# Update an existing skill (preserves intent)
/write-skill The emails being produced by 'write-email' sound too abstract for the 'b2b-, help me update the skill so that the emails don't use abstract concepts.
Note: after using a skill in an interactive session, you can invoke /write-skill at the end of the session and give it feedback how that session could have gone better.
Claude-code specific skill that creates or updates Claude Code agents (.claude/agents/*.md). Handles single-purpose subagents and multi-phase workflow orchestrators.
Useful pattern:
/write-agent can you take skills /research-prospect and /write-email and help me write an agent that researches prospects and writes a sequenece of emails?