ARCHIVED 2026-03-23 — Superseded by ultimate-claude-code-skills-playbook.md

Solanasis MCP & Plugins Installation Plan

Date: March 15, 2026 Status: READY FOR REVIEW — No changes made yet Reviewer: Senior Review Agent (verified) Purpose: Complete, step-by-step plan for installing remaining MCP integrations, marketplace plugins, and custom skills

CRITICAL CONTEXT: COWORK SESSION VS LOCAL INSTALLATION

THIS IS THE MOST IMPORTANT DECISION: You are currently in a Cowork session (a VM). Many installations will NOT persist after the session ends.

What Persists vs What Resets

Component	Cowork Session (VM)	Local Machine	Decision
Marketplace plugins (.plugin files)	Resets after session	Persist (in ~/.claude/)	INSTALL LOCALLY via Claude Code CLI
MCP servers (.mcp.json config)	Resets after session	Persist	CONFIGURE LOCALLY via Claude Code CLI
CLAUDE.md files	Can work in session	Persist better locally	CREATE BOTH (user + project)
Subagent files (~/.claude/agents/)	Resets after session	Persist	CREATE LOCALLY
Custom skills (.claude/skills/)	Can work in session	Persist better locally	Create in session, copy to local
Hooks (settings.json)	Can work in session	Better locally	LOCAL ONLY

Architecture Implication

Cowork is useful for:

Testing and validating new skills/plugins
Building custom skills using skill-creator
Verifying MCP connections work
Planning and documentation

Cowork is NOT sufficient for:

Permanent plugin installation (use Claude Code CLI)
MCP server setup (use Claude Code CLI + environment)
System-wide configuration (use local machine)

Therefore, this plan has TWO execution paths:

Testing & validation (CAN happen in Cowork NOW)
Permanent installation (MUST happen LOCALLY on your machine)

SECTION 1: MASTER INSTALLATION SEQUENCE

Phase 1A: MCP Servers (LOCAL CLI — Must Do First)

These are the research tools that make Chrome unnecessary for basic lookups.

Step 1: Add DuckDuckGo MCP (Recommended Over Brave)

Why first: Unlimited free searches, zero API key, zero setup friction. Gives Claude a fast search alternative to Chrome.

Local machine setup:

# Option 1: npm install (recommended)
npm install -g duckduckgo-mcp
 
# Option 2: Docker (if npm not available)
docker run -d -p 3000:3000 nickclyde/duckduckgo-mcp

Verify installation:

# Test that the command is available
which duckduckgo-mcp
# Should return: /usr/local/bin/duckduckgo-mcp (or similar)

Add to Claude Code CLI config:

Edit ~/.claude/settings.json (create if doesn’t exist):

{
  "mcpServers": {
    "duckduckgo": {
      "command": "npx",
      "args": ["-y", "@duckduckgo/duckduckgo-mcp"]
    }
  }
}

Verify in Claude Code:

Run claude in terminal
Type /mcp → should see duckduckgo listed
Test with: “Search for ‘SOC 2 compliance requirements’” → should use DuckDuckGo, not Chrome

Cowork test: After local setup, open a Cowork session and try a research query. DuckDuckGo should appear in the tools list.

Step 2: Add Exa MCP (Free Remote Endpoint)

Why: Semantic search for business research, company research, people search. Higher quality than DuckDuckGo for technical queries. Free tier with no API key.

Local machine setup — Three options (pick one):

Option A: Add Exa’s free remote MCP endpoint (RECOMMENDED — zero local setup)

Edit ~/.claude/settings.json:

{
  "mcpServers": {
    "exa": {
      "command": "curl",
      "args": ["-sL", "https://mcp.exa.ai/mcp?tools=web_search_exa,web_search_advanced_exa,company_research_exa,people_search_exa,get_code_context_exa,crawling_exa,deep_researcher_start,deep_researcher_check"]
    }
  }
}

Note: This uses curl to fetch the remote endpoint. Alternative: use the official Exa MCP GitHub repo if curl doesn’t work with remote MCPs.

Option B: Install locally from Exa GitHub repo

git clone https://github.com/exaai/exa-mcp-server.git
cd exa-mcp-server
npm install
npm run build

Then add to ~/.claude/settings.json:

{
  "mcpServers": {
    "exa": {
      "command": "node",
      "args": ["/path/to/exa-mcp-server/dist/index.js"],
      "env": {
        "EXA_API_KEY": "${EXA_API_KEY}"  // Leave blank for free tier
      }
    }
  }
}

Option C: Use Docker (if npm/curl not working)

docker run -d \
  -e EXA_API_KEY="" \
  -p 3001:3001 \
  exaai/mcp-server:latest

Then configure in ~/.claude/settings.json to connect to localhost:3001.

Recommend: Option A (remote endpoint) — no local installation needed.

Verify:

Run claude in terminal
Type /mcp → should see exa listed
Test with: “Research Acme Corp using company_research_exa” → should return structured company data

Step 3: Set Environment Variable for Exa API Key (Optional — Free Tier Works Without)

If you want to lift rate limits later:

Get free Exa API key: https://api.exa.ai/
Add to your shell profile (~/.bashrc, ~/.zshrc, etc.):

export EXA_API_KEY="your_api_key_here"

Reload shell: source ~/.bashrc (or source ~/.zshrc)

Free tier limits: 1,000 requests/month (~33/day). Plenty for current Solanasis scale.

Step 4: Disable Claude.ai MCP Inheritance (CRITICAL)

Claude.ai may auto-inject MCP servers into your CLI sessions. This can cause conflicts.

Disable with environment variable:

# Run Claude Code with this flag:
ENABLE_CLAUDEAI_MCP_SERVERS=false claude
 
# Or add to your shell profile permanently:
# echo 'export ENABLE_CLAUDEAI_MCP_SERVERS=false' >> ~/.bashrc

Verify: Run claude, then /mcp — should ONLY see the servers you’ve explicitly configured, not extra Claude.ai servers.

Phase 1B: CLAUDE.md Configuration (LOCAL — Can Start Now)

These instructions guide Claude’s behavior persistently across sessions.

Step 5: Create User-Level CLAUDE.md

File location: ~/.claude/CLAUDE.md

cat > ~/.claude/CLAUDE.md << 'EOF'
# Global Instructions — Dmitri Zasage / Solanasis
 
## Identity
- Working with Dmitri Zasage, CEO of Solanasis LLC (Colorado)
- Solanasis = fractional CIO/CSIO/COO (fCIO/fCSIO/fCOO) firm for SMBs and nonprofits
- Core offerings: Security Assessments, Disaster Recovery Verification, Data Migrations, CRM Setup, Systems Integration, Responsible AI Implementation
- Currently a one-person operation with 1099 contractors
- Growth hacking mindset — unconventional/Smartcuts approach
 
## Planning: Proportional to Complexity
 
- **Routine tasks** (research a prospect, draft an email, update a doc): Just do it. No formal planning.
- **Medium tasks** (multi-step research, document creation, config changes): State a brief plan (3-5 bullets) before starting.
- **Complex tasks** (new services, system redesigns, multi-day projects, strategic decisions): Use the `planner` subagent for a structured plan.
 
Don't waste time planning the obvious. Save planning for work where the approach genuinely matters.
 
## Verification: Proportional to Stakes
 
Launch the `senior-reviewer` subagent ONLY for:
- **Client deliverables** — proposals, reports, assessments, anything going to someone outside Solanasis
- **Strategic decisions** — pricing, new offerings, partnerships, major process changes
 
For everything else (internal docs, research, emails, routine tasks), self-review is sufficient. Don't burn tokens verifying draft emails.
 
**Decision tree:**
- Client deliverables → ALWAYS run senior-reviewer
- Strategic decisions → ALWAYS run senior-reviewer
- Internal docs, research, analysis → Self-review only
- Draft emails, messages, routine tasks → No verification
- Quick answers, lookups, conversation → No verification
 
## Research: Use the Research Agent for Depth
 
Launch the `research-agent` subagent for:
- **Prospect/company research** — Before sales calls, when prepping for client meetings
- **Competitor analysis** — Understanding the competitive landscape
- **Market research** — Industry trends, technology landscape, compliance requirements
- **Technology evaluation** — Comparing tools, vendors, platforms before recommending to clients
- **Security/compliance research** — Threat landscape, framework requirements, vulnerability context
 
For simple lookups ("What does Acme Corp do?"), just use WebSearch directly — no need for the full research agent.
 
**Decision tree:**
- Deep prospect prep → research-agent
- Competitor analysis → research-agent
- Market/tech research → research-agent
- Strategic research feeding a decision → research-agent → then senior-reviewer on the recommendation
- Quick factual lookup → WebSearch directly
- Checking a specific URL → WebFetch directly
 
## Tool Routing — Efficiency, Not Restriction
 
The goal is efficiency, not restriction. Use whichever tool gets the job done fastest.
 
### Quick Guide
- **WebSearch** — fastest for finding information, searching topics, news, docs. Use this when you just need to look something up.
- **WebFetch** — fastest for reading a specific URL you already have.
- **Exa/DuckDuckGo MCP** — for deeper semantic search, company research, people research (if installed).
- **Chrome** — for everything interactive: logins, dashboards, forms, screenshots, site testing, web apps, visual inspection, or anything that needs a real browser.
 
### Efficiency tip
WebSearch returns results in one call. Chrome requires navigate → wait → read_page (3+ calls, more tokens). For a simple "what is X?" question, WebSearch is 10x faster. But for "log into ClickUp and check the sprint," Chrome is the only option.
 
**Use Chrome freely.** It's the right tool for interactive work. Just don't use it as a search engine when WebSearch is faster.
 
## Communication Preferences
 
- Bullet points and numbered lists preferred over long paragraphs
- Sub-bullet points for additional context
- Always spell out acronyms on first use (unless super common like AI, IT, CEO)
- Include "pro tips" whenever relevant to help Dmitri learn
- Clarifying questions via multiple choice:
  - Option A = recommended answer (with explanation of why)
  - Remaining options in order of recommendation
  - Include context with each option for learning
  - Provide textarea for additional notes per answer
 
## Technical Conventions
- SQL: lowercase keywords and built-in functions, snake_case for tables/cols/procs/funcs/views/triggers
- C#: snake_case for variables that match DB column names
- Codebehind pattern for Blazor, JSInterop for forms
- HTML generated via C# classes
EOF

Verify:

cat ~/.claude/CLAUDE.md | head -20
# Should show the Identity section

Step 6: Update Project-Level CLAUDE.md

File location: /sessions/admiring-modest-gauss/mnt/_solanasis/solanasis-docs/.claude/CLAUDE.md

Read existing file first:

cat /sessions/admiring-modest-gauss/mnt/_solanasis/solanasis-docs/.claude/CLAUDE.md

Add these sections to the end (don’t overwrite security rules):

## MCP Tools
 
### Available MCP Servers
- DuckDuckGo (web, news, image search — unlimited, free)
- Exa (semantic search, company research, people search — free tier)
- ClickUp connector (project management)
- Google Calendar (event management)
- Gmail (email search, drafts)
- Canva (design)
 
### Preferred Research Tools (in order)
1. WebSearch — For finding info quickly
2. Exa (if available) — For business/company research
3. DuckDuckGo (if available) — Unlimited backup search
4. WebFetch — For reading a specific URL
5. Chrome read_page — Only if above tools fail (JS-heavy pages)
 
DO NOT use Chrome for basic web research. Chrome is for interactive tasks only.
 
## Active Connectors
- ClickUp (project management)
- Google Calendar
- Gmail
- Canva (design)
 
## Tech Stack Context
- ClickUp for PM, Xero for accounting, Coda as wiki
- Google Workspace, Google Voice for business phone
- Website: solanasis.com (cPanel/Namecheap Stellar VPS)
- Brevo for email marketing (List ID: 2)
- Payment: 50% upfront / 50% on delivery (full upfront under $2,500)

Verify:

tail -20 /sessions/admiring-modest-gauss/mnt/_solanasis/solanasis-docs/.claude/CLAUDE.md

Phase 1C: Subagent Files (LOCAL — Create in CLI)

These enable planning, research, and verification workflows.

Step 7: Create Subagent Directory

mkdir -p ~/.claude/agents

Step 8: Create Senior Reviewer Subagent

File: ~/.claude/agents/senior-reviewer.md

cat > ~/.claude/agents/senior-reviewer.md << 'AGENT_EOF'
---
name: senior-reviewer
description: >
  Senior quality reviewer. Use this agent to verify ANY substantive work
  before presenting to the user. This includes: documents, plans, research
  findings, configurations, code, templates, and client deliverables.
  Use proactively after completing work.
tools:
  - Read
  - Grep
  - Glob
  - Bash
  - WebSearch
  - WebFetch
disallowedTools:
  - Write
  - Edit
  - Agent
model: opus
maxTurns: 10
---
 
# Senior Reviewer Agent
 
You are a senior technical and strategic reviewer for Solanasis, a
fractional CIO/CSIO/COO (fCIO/fCSIO/fCOO) firm targeting SMBs and
nonprofits.
 
## Your Job
 
Review the work that was just completed and provide an honest assessment.
You are the quality gate before anything goes to the user (Dmitri, the CEO).
 
## Review Checklist
 
For EVERY review, check:
 
### Accuracy
- Are facts correct? Cross-reference claims with web search if needed.
- Are technical details accurate (framework names, tool capabilities, pricing, configuration syntax)?
- Are there unsupported claims or hallucinations?
 
### Completeness
- Does the output fully address what was asked?
- Are there obvious gaps or missing sections?
- Would Dmitri need to ask follow-up questions to use this?
 
### Best Practices
- Is this truly the best approach, or is there a better way?
- Are there industry best practices being ignored?
- Would an experienced consultant do it differently?
 
### Solanasis Context
- Does this align with Solanasis's business model (small team, 1099 contractors, credibility-building phase)?
- Is this practical for a one-person operation scaling up?
- Does it fit the "unconventional/Smartcuts" approach Dmitri prefers?
 
### Tool Usage Review
- Were the right tools used? (WebSearch vs Chrome, etc.)
- Was the approach efficient or wasteful?
- Could this have been done faster/better?
 
## Output Format
 
Return a structured verdict:

Senior Review Verdict

Status: APPROVED | APPROVED WITH NOTES | NEEDS REVISION

Accuracy: [Pass/Fail] — [brief note] Completeness: [Pass/Fail] — [brief note] Best Practices: [Pass/Fail] — [brief note] Solanasis Fit: [Pass/Fail] — [brief note] Efficiency: [Pass/Fail] — [brief note]

Issues Found: (if any)

[Issue 1]
[Issue 2]

Suggestions: (if any)

[Suggestion 1]
[Suggestion 2]


If status is NEEDS REVISION, be specific about what needs to change.
AGENT_EOF

Verify:

cat ~/.claude/agents/senior-reviewer.md | head -20

Step 9: Create Planner Subagent

File: ~/.claude/agents/planner.md

cat > ~/.claude/agents/planner.md << 'AGENT_EOF'
---
name: planner
description: >
  Task planning agent. Use for any multi-step task to plan the approach
  before execution. Determines which tools to use, what order to work in,
  and identifies potential issues. Use proactively for complex requests.
tools:
  - Read
  - Grep
  - Glob
  - WebSearch
  - WebFetch
disallowedTools:
  - Write
  - Edit
  - Bash
  - Agent
model: opus
maxTurns: 5
---
 
# Task Planner Agent
 
You plan the most efficient approach for completing tasks.
 
## Your Job
 
Given a task description, produce a brief execution plan that:
 
1. **Identifies the right tools** — Specifically:
   - WebSearch for finding information (NEVER Chrome for research)
   - WebFetch for reading specific URLs (NEVER Chrome navigate + read_page)
   - Chrome ONLY for authenticated sites, form filling, or explicit browser automation
   - Read/Glob/Grep for local files
   - Bash for commands and scripts
 
2. **Orders the steps** — What needs to happen first, what can be parallel
 
3. **Flags risks** — What could go wrong, what to watch out for
 
4. **Estimates scope** — Quick (< 5 min), Medium (5-15 min), Complex (15+ min)
 
## Output Format

Execution Plan

Scope: Quick | Medium | Complex Steps:

[Step] — using [Tool]
[Step] — using [Tool]
[Step] — using [Tool]

Tool Routing:

Research: WebSearch (no Chrome needed)
URL reading: WebFetch (no Chrome needed)
[Any Chrome needed?]: [Yes/No — reason]

Risks: [Any gotchas] Verification: [What senior-reviewer should check]


Keep plans concise. 3-7 steps for most tasks.
AGENT_EOF

Verify:

cat ~/.claude/agents/planner.md | head -20

Step 10: Create Research Agent

File: ~/.claude/agents/research-agent.md

cat > ~/.claude/agents/research-agent.md << 'AGENT_EOF'
---
name: research-agent
description: >
  Deep research specialist. Use this agent for ANY research task that requires
  searching multiple sources, synthesizing findings, or building a comprehensive
  picture. This includes: prospect/company research, competitor analysis, market
  research, technology evaluation, security landscape research, compliance
  requirements, vendor comparison, and due diligence. Use proactively whenever
  research depth matters.
tools:
  - WebSearch
  - WebFetch
  - Read
  - Grep
  - Glob
  - Bash
disallowedTools:
  - Write
  - Edit
  - Agent
  - mcp__Claude_in_Chrome__navigate
  - mcp__Claude_in_Chrome__form_input
  - mcp__Claude_in_Chrome__file_upload
model: opus
maxTurns: 15
---
 
# Research Agent — Solanasis Deep Research Specialist
 
You are a senior research analyst for Solanasis, a fractional CIO/CSIO/COO
(fCIO/fCSIO/fCOO) firm targeting SMBs and nonprofits.
 
## Your Job
 
Conduct thorough, multi-source research and return a structured, citation-rich
report. You are the research equivalent of the senior-reviewer — a specialist
that ensures research quality before findings reach the CEO (Dmitri).
 
## Research Tool Priority (Efficiency Order)
 
Use the fastest tool that gets the job done:
 
1. **WebSearch** — First choice for finding information, news, documentation,
   articles. Fast and cheap.
2. **Exa MCP tools** (if available) — For semantic search, company research
   (`company_research_exa`), people search (`people_search_exa`), deep
   research reports (`deep_researcher_start`/`deep_researcher_check`).
3. **DuckDuckGo MCP** (if available) — Alternative search with no rate limits.
4. **WebFetch** — For reading specific URLs you've found. Extract content
   from articles, docs, reports.
5. **Read/Grep/Glob** — For analyzing local files, uploaded documents,
   previous research.
6. **Chrome read-only tools** — `read_page`, `get_page_text` are available
   as FALLBACK for pages that require JavaScript rendering or have
   anti-scraping protection. State why WebFetch didn't work if you use these.
 
## Research Standards
 
### Source Quality
- Prefer official sources (company websites, SEC filings, government databases, vendor documentation)
- Cross-reference claims across 2+ sources when possible
- Flag single-source claims as "unverified" or "reported by [source]"
- Note publication dates — flag anything older than 12 months as potentially outdated
 
### Intellectual Honesty
- Clearly distinguish: verified facts vs. inferences vs. speculation
- Say "I couldn't find information on X" rather than making it up
- Note conflicting information when sources disagree
- Flag gaps in the research — what COULDN'T you find that matters?
 
### Solanasis Context
- We target SMBs (Small and Medium Businesses) and nonprofits
- Our wedge offerings: security assessments, disaster recovery (DR) verification, data migrations
- Goal: become their operational resilience partner with recurring revenue
- Growth hacking mindset — look for unconventional angles
- Currently building credibility — partner/association references are valuable
 
## Output Format

Research Report: [Topic]

Date: [date] Research Depth: Quick Scan | Standard | Deep Dive Sources Consulted: [number]

Key Findings

[Finding 1] — [Source]
[Finding 2] — [Source]
[Finding 3] — [Source]

Analysis

[Synthesized narrative connecting the findings]

Solanasis Relevance

[How this connects to our business — opportunities, risks, action items]

Gaps & Limitations

[What couldn’t be found]
[What needs further investigation]

Sources

Source 1 — accessed [date]
Source 2 — accessed [date]


Adjust depth based on what's asked. Quick prospect lookups don't need the full template — use judgment.
AGENT_EOF

Verify:

cat ~/.claude/agents/research-agent.md | head -20
ls -la ~/.claude/agents/
# Should show: planner.md, senior-reviewer.md, research-agent.md

Phase 1D: Test MCP & Subagents in Claude Code

Step 11: Verify MCP Installation

# Start Claude Code CLI
claude
 
# In the CLI, run:
/mcp
# Should output: duckduckgo, exa (and any other MCPs you configured)
 
# Test DuckDuckGo
# (in Claude chat)
Search for "SOC 2 compliance requirements"
# Should use duckduckgo tool, not Chrome
 
# Test Exa
# (in Claude chat)
Research the company "Anthropic" using Exa
# Should use exa tools, return structured company data

Success criteria:

/mcp shows both duckduckgo and exa
WebSearch is used for basic queries (faster than Chrome)
Exa is used for company/people research (better than WebSearch)
Chrome is NOT used for research-only tasks

Step 12: Test Subagents

# In Claude Code CLI, test each subagent:
 
# Test 1: Planner
Plan a task to research three SMB prospects and draft cold emails
# Should invoke planner, return a structured 5-step plan
 
# Test 2: Research Agent
Research the cybersecurity compliance landscape for SMBs in 2026
# Should invoke research-agent, return a structured report
 
# Test 3: Senior Reviewer
Write a client proposal, then ask for review
# Should invoke senior-reviewer after you finish, return verdict

Success criteria:

Planner delegates complex tasks automatically
Research agent is used for depth research automatically (via description)
Senior reviewer is launched for client work automatically

Phase 2A: Marketplace Plugins Installation (LOCAL CLI)

Step 13: Install Phase 1 Plugins (20 Skills from 6 Plugins)

Location: Local Claude Code CLI (installed on your machine, persists)

Commands to run locally:

claude
# In the CLI, use plugin install or equivalent command
# (Exact syntax depends on your Claude Code version)
 
# Recommended installation order (install in parallel if possible):
 
/plugin install operations@knowledge-work-plugins
/plugin install sales@knowledge-work-plugins
/plugin install marketing@knowledge-work-plugins
/plugin install engineering@knowledge-work-plugins
/plugin install customer-support@knowledge-work-plugins
/plugin install data@knowledge-work-plugins

Note: If these commands don’t work, try:
/plugin marketplace search operations
# Then click install on the Operations plugin

Verify:

/plugin list
# Should show 6 new plugins (operations, sales, marketing, engineering, customer-support, data)
# Plus existing: legal, productivity, brand-voice, cowork-plugin-management

Step 14: Configure Model Invocation Mapping

After plugins install, you need to set 8 of the 20 skills to disable-model-invocation: true to keep context budget manageable.

Location: .claude/plugins/ directory or plugin configuration files

Skills to set to MANUAL-ONLY (8 total):

Plugin	Skill	Reason
Operations	change-management	Invoke only when planning migrations
Operations	resource-planning	Invoke only for staffing/capacity
Sales	create-an-asset	Invoke explicitly for proposals
Sales	competitive-intelligence	Invoke explicitly for battlecards
Sales	daily-briefing	Invoke explicitly each morning
Marketing	campaign-planning	Invoke explicitly for campaigns
Customer Support	knowledge-management	Invoke after resolving issues
Data	data-validation	Invoke explicitly during migration QA

Skills to keep AUTO-INVOKE (12 total):

Plugin	Skill	Reason
Operations	compliance-tracking	Triggers on “SOC 2”, “GDPR” — frequent
Operations	risk-assessment	Triggers on “risk”, “what could go wrong” — frequent
Operations	vendor-management	Triggers on “evaluate vendor” — frequent
Operations	process-optimization	Triggers on “bottleneck” — frequent
Sales	account-research	Triggers on “research [company]” — daily
Sales	draft-outreach	Triggers on “draft outreach to” — daily
Sales	call-prep	Triggers on “prep me for my call” — daily
Marketing	content-creation	Triggers on “write a blog post” — frequent
Engineering	documentation	Triggers on “write docs” — frequent
Engineering	incident-response	Triggers on “incident”, “production down” — critical
Data	data-exploration	Triggers on “profile this dataset” — migration work
Data	sql-queries	Triggers on “write a query” — migration work

How to disable model invocation:

Option A: Via CLI (if available)

# (Syntax varies by Claude Code version — check `/help`)
/plugin skill disable-model-invocation change-management
/plugin skill disable-model-invocation resource-planning
# ... repeat for all 8 skills

Option B: Edit plugin SKILL.md files directly

Find the plugin files in ~/.claude/plugins/ or the marketplace folder, then edit each SKILL.md to add:

disableModelInvocation: true

Or in the skill frontmatter:

---
name: change-management
disableModelInvocation: true
---

Verify:

# In Claude CLI or Cowork, ask:
"What skills can you invoke?"
# Should list 12 skills (auto-invoke), not 20

Phase 2B: Custom Skills Build

Step 15: Create Research-First Skill

File: ~/.claude/skills/research-first/SKILL.md

mkdir -p ~/.claude/skills/research-first
 
cat > ~/.claude/skills/research-first/SKILL.md << 'SKILL_EOF'
---
name: research-first
description: >
  Quick focused research on a specific topic. Research using MCP tools
  (Exa, DuckDuckGo, WebSearch) before falling back to WebFetch or Chrome.
  For deeper research, delegate to the research-agent subagent.
  Invoke manually with /research-first.
disableModelInvocation: true
---
 
# Research-First Skill
 
A lightweight skill for focused research tasks, complementing the deeper `research-agent` subagent.
 
## How to Use
 
Invoke manually: `/research-first [topic]`
 
Examples:
- `/research-first SOC 2 Type II compliance for SaaS`
- `/research-first disaster recovery best practices for SMBs`
- `/research-first latest data migration tools 2026`
 
## Research Sequence
 
1. **Exa semantic search** — Company research, people search, deep researcher tools
2. **DuckDuckGo** — Web and news search if Exa unavailable
3. **WebSearch** — General web search fallback
4. **WebFetch** — Read specific documents you found
5. **Chrome read_page** — Only if above fails (JS-heavy pages)
 
**NEVER use Chrome as the primary research tool.** Chrome is for interactive tasks only.
 
## Output Format
 
Return findings in this format:

Research: [Topic]

Top Findings

[Finding] — Source
[Finding] — Source
[Finding] — Source

Key Insights

[Synthesized summary]

Next Steps

[What to do with this research]


Keep research focused and actionable. Adjust depth based on the topic.
SKILL_EOF

Verify:

# In Claude CLI:
/skill research-first
# Should prompt for topic and return research results

Phase 3: Context Budget Verification

Step 16: Calculate Total Context Load

Before plugin installation, verify context window budget:

Component	Size	Count	Subtotal
Existing skills	~200 chars	6	~1,200 chars
Phase 1 plugins (12 auto-invoke)	~200 chars	12	~2,400 chars
Subagent descriptions	~300 chars	3	~900 chars
CLAUDE.md content	~800 chars	1	~800 chars
Memory/context files	variable	1	~500 chars
TOTAL ACTIVE			~5,800 chars

Available context budget (at 128K context):

System prompt: ~2,000 chars
Reserved for conversation: ~120,000 chars
Available for tool descriptions: ~2,560 chars

Status: ⚠️ OVER BUDGET

Mitigation: Implement Model Invocation Mapping

Keep 12 auto-invoke (high-frequency) skills enabled
Set 8 skills to manual-only (disable model invocation)
Result: ~2,400 chars active (under ~2,560 budget)

Verify after installation:

# In Claude CLI, check how many skills are loaded
/skill list
# Count auto-invoke vs manual-only
 
# Ask Claude:
"Are you seeing 'Excluded skills' warnings?"
# If YES → not enough context budget, disable more skills
# If NO → context budget is OK

SECTION 2: INSTALLATION CHECKLIST (STEP-BY-STEP)

Prerequisites

Claude Code CLI installed locally (NOT Cowork)
Terminal access to your machine (not the VM)
npm installed (npm --version returns a version)
Git installed (git --version returns a version)
Max plan is active (needed for subagent token budget)

Phase 1: MCP Servers (LOCAL)

DuckDuckGo MCP installed (npm install -g duckduckgo-mcp)
DuckDuckGo added to ~/.claude/settings.json
DuckDuckGo tested in Claude CLI (/mcp shows it)
Exa MCP configured (remote endpoint or local)
Exa tested in Claude CLI
ENABLE_CLAUDEAI_MCP_SERVERS=false verified (or disabled)

Phase 1: CLAUDE.md (LOCAL)

User-level CLAUDE.md created (~/.claude/CLAUDE.md)
Project-level CLAUDE.md updated (solanasis-docs/.claude/CLAUDE.md)
Both files verified with cat command

Phase 1: Subagents (LOCAL)

~/.claude/agents/ directory created
senior-reviewer.md created and verified
planner.md created and verified
research-agent.md created and verified
Subagents tested in Claude CLI

Phase 2: Plugins (LOCAL)

operations plugin installed (/plugin install operations@knowledge-work-plugins)
sales plugin installed
marketing plugin installed
engineering plugin installed
customer-support plugin installed
data plugin installed
Model invocation mapping applied (8 skills set to manual-only)
Total skills: 12 auto-invoke, 8 manual-only

Phase 2: Custom Skills

research-first skill created (~/.claude/skills/research-first/SKILL.md)
research-first skill tested

Phase 3: Verification

Context budget verified (no “Excluded skills” warnings)
Chrome permissions set to Allow (not Ask)
DuckDuckGo/Exa preferred for research (tested)
Chrome reserved for interactive tasks (tested)
Subagents auto-invoke correctly (tested)
Plugins show in /plugin list (tested)

Phase 3: Cowork Session

Open Cowork session with solanasis-docs folder
Verify plugins appear in skill menu
Verify MCP tools available in chat
Test full workflow: research → plan → create → verify

SECTION 3: WHAT CAN BE TESTED IN COWORK NOW

These items CAN be tested in the current Cowork session before local installation:

Skill logic validation — Create/test custom skills using skill-creator (already installed)
Plugin compatibility — Read plugin docs, verify they don’t conflict with existing skills
Workflow simulation — Manually run through the research → plan → create → verify loop
CLAUDE.md effectiveness — See if instructions guide tool selection correctly
Context budget impact — Estimate token costs of new skills

These items CANNOT be tested in Cowork (will reset after session):

Permanent MCP server installation
Plugin installation (marketplace plugins only install in local Claude Code)
Subagent file creation (needs to persist to ~/.claude/agents/)
System-wide configuration (ENABLE_CLAUDEAI_MCP_SERVERS flag)

SECTION 4: ROLLBACK PLAN (IF SOMETHING BREAKS)

If Context Window Overflows

Symptom: Claude says “Excluded skills” or responses become evasive

Fix:

Disable more skills from manual-only list (move from 8 to 12 manual)
Remove Phase 2 plugins temporarily
Keep Phase 1A (MCP) + Phase 1B (CLAUDE.md) + 3 critical skills

If MCP Servers Stop Working

Symptom: WebSearch works but DuckDuckGo/Exa don’t appear in /mcp

Fix:

# Check for conflicts
/mcp
 
# Remove and reinstall
npm uninstall -g duckduckgo-mcp
npm install -g duckduckgo-mcp
 
# Restart Claude
claude

If Plugins Cause Conflicts

Symptom: Skill descriptions collide, or unexpected behaviors emerge

Fix:

Note which plugin caused the issue
Run /plugin uninstall [plugin-name]
Verify conflict is gone
Reinstall one plugin at a time

If Subagents Don’t Auto-Invoke

Symptom: You ask research agent to be used, but it’s not invoked

Fix:

Check subagent description field — must include “use proactively”
Verify file location: ~/.claude/agents/[name].md
Verify YAML frontmatter is valid (no syntax errors)
Restart Claude Code: exit then claude

If Chrome Gets Blocked Accidentally

Symptom: “I don’t have permission to use Chrome”

Fix:

/permissions
# Find Chrome tools, set to "Allow"

SECTION 5: RISKS & GOTCHAS

High-Risk Items

Risk	Impact	Mitigation
MCP server command path wrong	Claude can’t find Exa/DDG tools	Verify with `which duckduckgo-mcp` before adding to settings
Context window overflow	”Excluded skills” warnings, degraded performance	Model invocation mapping (already in plan)
Plugin version conflicts	Skill descriptions collide, unexpected behavior	Install one plugin at a time, test between installs
Subagent nesting (Claude limitation)	Senior reviewer can’t delegate to research agent	Design subagents to be self-sufficient (already done)
ENABLE_CLAUDEAI_MCP_SERVERS injection	Unexpected MCP servers appear in sessions	Set environment variable at CLI startup

Medium-Risk Items

Risk	Impact	Mitigation
npm/Docker not available	Can’t install DuckDuckGo locally	Use Exa remote endpoint instead
Brave Search MCP setup friction	Requires API key, more complex	Skip for now, use DuckDuckGo + Exa
Cowork session loss	All unsaved custom skills/plugins reset	Copy custom skills to local machine immediately
Plugin marketplace unavailable	Can’t install plugins in Cowork	All plugins must be installed via local Claude Code CLI
Research-first skill underutilized	Skill exists but not used	Include in CLAUDE.md to remind when to use manually

Low-Risk Items

Risk	Impact	Mitigation
Exa free tier rate limit hit	Search tools slow after 1,000 queries/month	Get free API key to lift limit
DuckDuckGo search quality issues	Some queries return weak results	Use Exa or WebSearch as fallback
Skill descriptions need tweaking	Auto-invoke too aggressive or too passive	Update descriptions after first week of use

SECTION 6: VERIFICATION TESTS (RUN AFTER EACH PHASE)

Test 1: MCP Tools Available

claude
/mcp
# Output should include: duckduckgo, exa

Test 2: Research-First Tool Selection

# In Claude chat:
"What is the latest zero-day vulnerability landscape?"
# Verify: Uses Exa or DuckDuckGo, NOT Chrome

Test 3: Chrome Reserved for Interactive

# In Claude chat:
"Log into ClickUp and show me the sprint board"
# Verify: Uses Chrome (because it requires login)

Test 4: Subagents Auto-Invoke

# In Claude chat:
"Research this prospect before my call: Acme Corp"
# Verify: research-agent subagent invokes automatically

Test 5: Planning on Complex Tasks

# In Claude chat:
"Design a security assessment service offering for SMBs"
# Verify: planner subagent invokes, returns structured plan

Test 6: Verification on Client Work

# In Claude chat:
"Draft a proposal for Acme Corp for a security assessment"
# Verify: senior-reviewer subagent invokes after you finish

Test 7: Plugin Skills Appear

# In Claude chat:
"What skills do I have available?"
# Verify: Lists 20 skills (12 auto + 8 manual)

Test 8: Context Budget OK

# In Claude chat:
"Do you see any 'Excluded skills' warnings?"
# Verify: "No" — context budget is sufficient

SECTION 7: AFTER INSTALLATION — FIRST WEEK OPERATIONS

Day 1: Verify All Connections

Test each MCP tool (Exa, DuckDuckGo)
Test each subagent (planner, research-agent, senior-reviewer)
Test 5 critical plugins (operations, sales, engineering, data)
Verify CLAUDE.md is guiding tool selection

Day 2-3: Run Real Workflows

Research a prospect → draft outreach → create asset (uses sales skills)
Plan a security assessment → document findings (uses operations + engineering)
Extract and validate data from a sample migration (uses data skills)

Day 4-5: Monitor Chrome Usage

Track: How often is Chrome used for research vs. interactive?
Goal: Chrome used < 20% for research-only tasks
If > 20%: Add PreToolUse Chrome gate hook (Section 5 of Architecture doc)

Day 6-7: Adjust Model Invocation Mapping

Identify which manual-only skills should auto-invoke
Re-enable any skills you find yourself invoking > 3x/day
Disable any skills that auto-fire incorrectly

End of Week 1: Measure ROI

Count: How many client deliverables used plugins?
Count: How many were improved by senior-reviewer?
Count: How many research tasks used research-agent?
Decision: Keep all plugins, or disable any that aren’t earning their context?

APPENDIX: EXACT COMMAND REFERENCE

MCP Server Installation Commands

# DuckDuckGo (npm)
npm install -g duckduckgo-mcp
 
# Exa (remote endpoint — no local install needed)
# Just add to ~/.claude/settings.json
 
# Verify MCP is active
claude
/mcp
 
# Test DuckDuckGo
"Search for SOC 2 compliance using DuckDuckGo"
 
# Test Exa
"Research Anthropic Inc using Exa"

Plugin Installation Commands

claude
 
# Install marketplace plugins
/plugin install operations@knowledge-work-plugins
/plugin install sales@knowledge-work-plugins
/plugin install marketing@knowledge-work-plugins
/plugin install engineering@knowledge-work-plugins
/plugin install customer-support@knowledge-work-plugins
/plugin install data@knowledge-work-plugins
 
# List installed plugins
/plugin list
 
# Disable model invocation for a skill
/plugin skill disable-model-invocation [skill-name]

File Creation Commands

# Create directories
mkdir -p ~/.claude/agents
mkdir -p ~/.claude/skills/research-first
 
# Create files (use cat << 'EOF' pattern for multi-line content)
cat > ~/.claude/CLAUDE.md << 'EOF'
[content here]
EOF
 
# Verify files exist
ls -la ~/.claude/
ls -la ~/.claude/agents/

Environment Variables

# Disable Claude.ai MCP inheritance
export ENABLE_CLAUDEAI_MCP_SERVERS=false
claude
 
# Or add to shell profile permanently
echo 'export ENABLE_CLAUDEAI_MCP_SERVERS=false' >> ~/.bashrc
source ~/.bashrc

APPENDIX: DEFINITION OF “DONE”

Installation is complete when:

MCP Tools — All research tools (Exa, DuckDuckGo, WebSearch) are available and tested
CLAUDE.md — Both user-level and project-level instructions are in place
Subagents — All three (planner, research-agent, senior-reviewer) are created and auto-invoke correctly
Plugins — All 6 Phase 1 plugins installed, 12 auto-invoke, 8 manual-only
Custom Skills — research-first skill created and tested
Context Budget — No “Excluded skills” warnings, all 20 skills accessible
Verification Tests — All 8 verification tests pass
First Week — Full week of monitoring Chrome usage, adjusting model invocation mapping

Success Metric: Claude can research → plan → create → verify a full client deliverable without manual intervention, using the right tools efficiently.

Document Status: READY FOR EXECUTION Next Step: Review this plan with Dmitri, then proceed with Phase 1 (local installation) Expected Timeline: 4-6 hours for Phase 1, 2-3 hours for Phase 2, 1 week for Phase 3 (monitoring)

Solanasis Docs

Explorer

INSTALLATION-PLAN-MCP-PLUGINS-SKILLS

Solanasis MCP & Plugins Installation Plan

CRITICAL CONTEXT: COWORK SESSION VS LOCAL INSTALLATION

What Persists vs What Resets

Architecture Implication

SECTION 1: MASTER INSTALLATION SEQUENCE

Phase 1A: MCP Servers (LOCAL CLI — Must Do First)

Step 1: Add DuckDuckGo MCP (Recommended Over Brave)

Step 2: Add Exa MCP (Free Remote Endpoint)

Step 3: Set Environment Variable for Exa API Key (Optional — Free Tier Works Without)

Step 4: Disable Claude.ai MCP Inheritance (CRITICAL)

Phase 1B: CLAUDE.md Configuration (LOCAL — Can Start Now)

Step 5: Create User-Level CLAUDE.md

Step 6: Update Project-Level CLAUDE.md

Phase 1C: Subagent Files (LOCAL — Create in CLI)

Step 7: Create Subagent Directory

Step 8: Create Senior Reviewer Subagent

Senior Review Verdict

Step 9: Create Planner Subagent

Execution Plan

Step 10: Create Research Agent

Research Report: [Topic]

Key Findings

Analysis

Solanasis Relevance

Gaps & Limitations

Sources

Phase 1D: Test MCP & Subagents in Claude Code

Step 11: Verify MCP Installation

Step 12: Test Subagents

Phase 2A: Marketplace Plugins Installation (LOCAL CLI)

Step 13: Install Phase 1 Plugins (20 Skills from 6 Plugins)

Step 14: Configure Model Invocation Mapping

Phase 2B: Custom Skills Build

Step 15: Create Research-First Skill

Research: [Topic]

Top Findings

Key Insights

Next Steps

Phase 3: Context Budget Verification

Step 16: Calculate Total Context Load

SECTION 2: INSTALLATION CHECKLIST (STEP-BY-STEP)

Prerequisites

Phase 1: MCP Servers (LOCAL)

Phase 1: CLAUDE.md (LOCAL)

Phase 1: Subagents (LOCAL)

Phase 2: Plugins (LOCAL)

Phase 2: Custom Skills

Phase 3: Verification

Phase 3: Cowork Session

SECTION 3: WHAT CAN BE TESTED IN COWORK NOW

SECTION 4: ROLLBACK PLAN (IF SOMETHING BREAKS)

If Context Window Overflows

If MCP Servers Stop Working

If Plugins Cause Conflicts

If Subagents Don’t Auto-Invoke

If Chrome Gets Blocked Accidentally

SECTION 5: RISKS & GOTCHAS

High-Risk Items

Medium-Risk Items

Low-Risk Items

SECTION 6: VERIFICATION TESTS (RUN AFTER EACH PHASE)

Test 1: MCP Tools Available

Test 2: Research-First Tool Selection

Test 3: Chrome Reserved for Interactive

Test 4: Subagents Auto-Invoke

Test 5: Planning on Complex Tasks

Test 6: Verification on Client Work

Test 7: Plugin Skills Appear

Test 8: Context Budget OK

SECTION 7: AFTER INSTALLATION — FIRST WEEK OPERATIONS

Day 1: Verify All Connections

Day 2-3: Run Real Workflows

Day 4-5: Monitor Chrome Usage

Day 6-7: Adjust Model Invocation Mapping

End of Week 1: Measure ROI

APPENDIX: EXACT COMMAND REFERENCE

MCP Server Installation Commands