Learn the
loop, not
the tool.
Working setup. Portable framework.
anjanps.com
@anjanps5Effective
Claude setup.
Six moving parts. Set them up once, and every session after runs faster.
Six parts of a real setup.
CLAUDE.md
Project memory Claude reads every session: rules, structure, conventions.
Permissions
settings.json sets what runs without asking you first.
CLI tools
Commands Claude runs on demand, without loading schemas into context.
MCP servers
Structured access to your data and SaaS, for when a CLI won't cut it.
Skills
Reusable playbooks Claude loads when a task needs them.
Subagents
Parallel workers, each with its own context. Fan out, then combine.
Reach for a CLI
before an MCP.
Schemas load up front
Every server's tool definitions sit in your context window every turn, whether you use them or not.
Called on demand
Runs through the shell Claude already has. Called only when you need it, so your context stays clean.
Watch for: what to hard-code once vs. what to decide per task.
Tool-agnostic
AI maxing.
Tools churn every quarter. The skill that transfers is the loop underneath them.
One loop. Any model, any stack.
Context
The context you load is your leverage.
Spec
Define "done" before you build.
Build
Let the model draft the work.
Verify
Run it and confirm it works.
Ship
Commit the slice. Then loop.
Separate knowledge from code.
├── 00 Studio company · skills · PRDs ├── 01 Products strategy · roadmaps ├── 02 Clients docs only, per engagement ├── 03 Hiring └── 99 Inbox triage
├── ACCOUNTS.md per-folder git identity ├── acolyte/ day-job repos ├── studio/ studio repos ├── connect-america/ └── personal/
Two homes: one to think in, one to build in. Keep them apart and every session — and every subagent — starts from one clean, focused context, not a pile of unrelated docs and code.
Skills transfer.
Tools don't.
Swap Claude for the next model or Cursor for the next editor. Context, spec, and verify still work — that's where your learning compounds.
Loops &
building.
Stop typing every prompt. Design the job, then let it run.
Four loops you'll reuse.
Plan → Build → Verify → Commit
The default for any real feature. Small slices, each one shipped before the next.
Fan-out review
Run several reviewers in parallel, each trying to break the change. Keep it only if it holds up.
Research → Synthesize → Cite
Many searches in parallel, then one agent reconciles and sources it.
Loop-until-dry
Keep hunting until two rounds turn up nothing new.
Four ways to keep it running — but autonomous loops burn tokens fast. Give /goal a clear exit or a turn cap, or it runs forever.
Your turn.
What shipped?
Build the
workflow you want.
Pick one loop from tonight. Point it at a real problem. Start it before you leave.
Thank you, Chennai.
Great building with you tonight. Keep the loop going — and let's stay in touch.
anjanps.com@anjanps5Scan to connect →