Author: Vicky Lalwani
Profile: https://www.linkedin.com/in/vicky-lalwani/
30 days ago I shared the marketer's .claude/ folder.
3,000+ marketers downloaded it.
Anthropic has shipped 4 updates since.
Subagents. Plugins. Hooks events. Skill metadata.
Time to update the blueprint.
|
|
ā
If you're still running Claude Code with just CLAUDE.md +Ā
a few prompts ā you're on V1.
Here's V2.
The full architecture nobody is mapping:
ā CLAUDE.md is ADVISORY
ā³ Project memory. Loaded every task. Under 500 lines.
ā³ Vague rules = generic output. Specific rules = your output.
ā HOOKS are DETERMINISTIC
ā³ Fire every single time. No skipping. No model judgment.
ā³ PreToolUse, PostToolUse, SessionStart, SessionEnd, PreCommit.
ā³ Hooks = AI. Skills = AI. Different jobs entirely.
ā SKILLS are ON-DEMAND
ā³ Auto-activate when Claude detects the task matches.
ā³ One .md file per workflow. 9-section structure or fail.
ā³ Drop in. Run forever.
ā SUBAGENTS run in ISOLATED CONTEXT
ā³ Reviewer, researcher, analyst ā each in their own thread.
ā³ Main thread stays clean. Parallel work happens in the background.
ā³ The unlock most marketers haven't found yet.
ā MCPs connect EXTERNAL TOOLS
ā³ GitHub, Slack, GSC, Semrush, GA4, HubSpot, Postgres.
ā³ Claude reads live data. No more CSV exports.
ā³ Configure once in .mcp.json. Auto-connects every session.
ā PLUGINS bundle ALL OF THE ABOVE
ā³ Skills + hooks + agents + MCPs in one distributable folder.
ā³ One install. Whole team gets the same setup.
ā³ The piece that makes this team-ready, not just personal.
That's the full architecture.
|
|
ā
The shift since V1:
The BLUEPRINT V1 post was about getting started.
The V2 is about scaling ā going from one marketer'sĀ
workspace to a team-deployable system.
Solo operator ā /skills + /hooks + CLAUDE.md
Team operation ā add /agents + /plugins + .mcp.json
Multi-client agency ā all of the above + per-client skill packages
|
|
ā
Anti-patterns I see most often:
ā CLAUDE.md over 500 lines (context bloat ā Claude stops reading)
ā Vague instructions like "write good code" (give specific rules)
ā No test guidance (Claude skips testing without explicit prompts)
ā Duplicating docs into CLAUDE.md (link instead, don't copy)
ā Treating hooks like skills (hooks fire every time, skills wait)
|
|
ā
If you got the BLUEPRINT V1 last month ā V2 is the upgrade.
Inside V2:
ā Full 15-section .claude/ architecture map
ā Subagents directory with 3 working examples
ā Plugins manifest template
ā Updated CLAUDE.md template (under 500 lines, no bloat)
ā All hooks events explained with real use cases
ā MCP configuration with GSC + GTM + Semrush + Windsor.ai
ā Anti-patterns checklist
Free. No course. No upsell.
1. Like this post
2. Comment "BLUEPRINT V2"
3. Repost for the marketers in your network still on V1.
4. Connect with me ā add "BLUEPRINT V2" in the connection note
I'll DM the updated package this weekend.
MUST BE CONNECTED & REPOSTED.
3,000+ marketers downloaded it.
Anthropic has shipped 4 updates since.
Subagents. Plugins. Hooks events. Skill metadata.
Time to update the blueprint.
|
|
ā
If you're still running Claude Code with just CLAUDE.md +Ā
a few prompts ā you're on V1.
Here's V2.
The full architecture nobody is mapping:
ā CLAUDE.md is ADVISORY
ā³ Project memory. Loaded every task. Under 500 lines.
ā³ Vague rules = generic output. Specific rules = your output.
ā HOOKS are DETERMINISTIC
ā³ Fire every single time. No skipping. No model judgment.
ā³ PreToolUse, PostToolUse, SessionStart, SessionEnd, PreCommit.
ā³ Hooks = AI. Skills = AI. Different jobs entirely.
ā SKILLS are ON-DEMAND
ā³ Auto-activate when Claude detects the task matches.
ā³ One .md file per workflow. 9-section structure or fail.
ā³ Drop in. Run forever.
ā SUBAGENTS run in ISOLATED CONTEXT
ā³ Reviewer, researcher, analyst ā each in their own thread.
ā³ Main thread stays clean. Parallel work happens in the background.
ā³ The unlock most marketers haven't found yet.
ā MCPs connect EXTERNAL TOOLS
ā³ GitHub, Slack, GSC, Semrush, GA4, HubSpot, Postgres.
ā³ Claude reads live data. No more CSV exports.
ā³ Configure once in .mcp.json. Auto-connects every session.
ā PLUGINS bundle ALL OF THE ABOVE
ā³ Skills + hooks + agents + MCPs in one distributable folder.
ā³ One install. Whole team gets the same setup.
ā³ The piece that makes this team-ready, not just personal.
That's the full architecture.
|
|
ā
The shift since V1:
The BLUEPRINT V1 post was about getting started.
The V2 is about scaling ā going from one marketer'sĀ
workspace to a team-deployable system.
Solo operator ā /skills + /hooks + CLAUDE.md
Team operation ā add /agents + /plugins + .mcp.json
Multi-client agency ā all of the above + per-client skill packages
|
|
ā
Anti-patterns I see most often:
ā CLAUDE.md over 500 lines (context bloat ā Claude stops reading)
ā Vague instructions like "write good code" (give specific rules)
ā No test guidance (Claude skips testing without explicit prompts)
ā Duplicating docs into CLAUDE.md (link instead, don't copy)
ā Treating hooks like skills (hooks fire every time, skills wait)
|
|
ā
If you got the BLUEPRINT V1 last month ā V2 is the upgrade.
Inside V2:
ā Full 15-section .claude/ architecture map
ā Subagents directory with 3 working examples
ā Plugins manifest template
ā Updated CLAUDE.md template (under 500 lines, no bloat)
ā All hooks events explained with real use cases
ā MCP configuration with GSC + GTM + Semrush + Windsor.ai
ā Anti-patterns checklist
Free. No course. No upsell.
1. Like this post
2. Comment "BLUEPRINT V2"
3. Repost for the marketers in your network still on V1.
4. Connect with me ā add "BLUEPRINT V2" in the connection note
I'll DM the updated package this weekend.
MUST BE CONNECTED & REPOSTED.
Images:
