Starting a project

Monday stack

You have a company project this week. Install in order. One lab per step. One artifact. Skip 04 and 05 if the job does not need them. Skip 07 if this is greenfield. This is not Catch up and not the decoder.

01Pick a coding agent
Agents / harness
Install one. Drive a real ticket on this company’s repo this week. Not a toy.
Lab Claude Code quickstart · Cursor quickstart · Windsurf getting started
Artifact One real PR or ticket the agent helped close. · board Agents / harness
02Put the right things in the window
LLMs / context
Context is what the model can see this turn. Assemble it for THIS repo.
Lab Effective context engineering for AI agents
Artifact A reusable context / prompt assembly checked into the repo. · board LLMs / context
03Give it hands, then bound them
Agents / harness
Tools, permissions, MCP. The wrapper is the product.
Lab Effective harnesses for long-running agents · Code execution with MCP
Artifact Bounded tools + permissions on that project. · board Agents / harness
04Graph only if the work must survive a crash
Agents / harness
Skip unless the job is multi-step and has to resume.
Durable, multi-step loops. Checkpoint so a crash is not a restart.
Lab LangGraph persistence
Artifact A checkpointed loop, or skip. · board Agents / harness
05Company data only if the app needs it
Grounding / RAG
Skip unless the product must answer from internal docs, tickets, or code.
Ground in the product: retrieve with provenance, not a science paper.
Lab Agentic RAG with LangGraph
Artifact Retrieval with provenance, or skip. · board Grounding / RAG
06Eval before you call it shipped
Evals
A test the system can fail. Not a vibe check.
Lab OpenAI evals
Artifact An eval that can fail the loop. · board Evals
07Wrap into the existing system
Production
Skip if this is greenfield.
Wire the agent/model into the company app or system of record.
Lab AI SDK in a Next.js app
Artifact A call from the old app, not a demo URL. · board Production
08Ship a URL
Production
Put the app on cloud. Not a VPC. Not a training cluster.
Lab Next.js App Router + AI SDK · WorkflowAgent
Artifact A URL. · board Production

Heard the slang first? New here is the decoder. Want the four-skill spine? Catch up. The board is still the firehose.

Monday stack

01Pick a coding agent
Agents / harness
Install one. Drive a real ticket on this company’s repo this week. Not a toy.
Lab Claude Code quickstart · Cursor quickstart · Windsurf getting started
Artifact One real PR or ticket the agent helped close. · board Agents / harness
02Put the right things in the window
LLMs / context
Context is what the model can see this turn. Assemble it for THIS repo.
Lab Effective context engineering for AI agents
Artifact A reusable context / prompt assembly checked into the repo. · board LLMs / context
03Give it hands, then bound them
Agents / harness
Tools, permissions, MCP. The wrapper is the product.
Lab Effective harnesses for long-running agents · Code execution with MCP
Artifact Bounded tools + permissions on that project. · board Agents / harness
04Graph only if the work must survive a crash
Agents / harness
Skip unless the job is multi-step and has to resume.
Durable, multi-step loops. Checkpoint so a crash is not a restart.
Lab LangGraph persistence
Artifact A checkpointed loop, or skip. · board Agents / harness
05Company data only if the app needs it
Grounding / RAG
Skip unless the product must answer from internal docs, tickets, or code.
Ground in the product: retrieve with provenance, not a science paper.
Lab Agentic RAG with LangGraph
Artifact Retrieval with provenance, or skip. · board Grounding / RAG
06Eval before you call it shipped
Evals
A test the system can fail. Not a vibe check.
Lab OpenAI evals
Artifact An eval that can fail the loop. · board Evals
07Wrap into the existing system
Production
Skip if this is greenfield.
Wire the agent/model into the company app or system of record.
Lab AI SDK in a Next.js app
Artifact A call from the old app, not a demo URL. · board Production
08Ship a URL
Production
Put the app on cloud. Not a VPC. Not a training cluster.
Lab Next.js App Router + AI SDK · WorkflowAgent
Artifact A URL. · board Production

Heard the slang first? New here is the decoder. Want the four-skill spine? Catch up. The board is still the firehose.

CLAUDE COMPUTING

01Pick a coding agent

02Put the right things in the window

03Give it hands, then bound them

04Graph only if the work must survive a crash

05Company data only if the app needs it

06Eval before you call it shipped

07Wrap into the existing system

08Ship a URL