Notes from the work.
Deep dives on the handoffs, data leaks, and automation decisions that turn into revenue systems.
- Written for
- People running the work
- Latest posts
- 68
- Inside
- Systems, decisions, lessons
Showing 51-60 of 68 notes
AI coworker pattern: role, memory and a seat at standup
An AI coworker is an agent with a clear role, memory that lasts across sessions, and a slot in your team's standup. Here is how I build each piece.
Read the note
Claude Agent SDK vs LangChain: which to build agents on
Claude Agent SDK vs LangChain for a Claude agent: what each layer gives you, where LangChain and langchain-anthropic still win, and how to pick one.
Read the note
Claude skills vs tools vs MCP: when to use each
Claude skills vs tools vs MCP in plain terms: what each one holds, what it costs in context, and a simple decision tree for picking the right one first.
Read the note
Claude Slack integration that runs 16 cron agents
My Claude Slack integration runs 16 cron agents and lifts every blocker into one shared queue, so I clear them in one sitting instead of hunting threads.
Read the note
Ahrefs MCP server and Google Ads MCP: a reference design
A reference design for an Ahrefs MCP server that also pulls Google Ads and internal content data, so several AI workflows share one set of tools.
Read the note
Claude Code for business: a team force multiplier
Claude Code for business is not a faster chat tab. Here is what changes when a whole team shares its repo access, CLAUDE.md file, hooks and skills.
Read the note
AI resume screening filters out your best candidates
AI resume screening can reject qualified people before a recruiter sees them. Here is why, and the safer setup where AI reads and a person decides.
Read the note
When a custom MCP server is worth building
A custom MCP server pays off in five cases, from private APIs to audit logging. Here is when to build an MCP server and when an existing one is enough.
Read the note
MCP vs function calling vs REST API: which to use
MCP vs function calling vs a REST API: all three expose data and actions to an LLM. Here is who each one is for, plus a five-rule decision tree.
Read the note
n8n queue mode and an architecture that survives 10x
n8n queue mode, a staging table and split credentials: the n8n architecture I use so a workflow that is fine at 100 runs a day survives 1,000.
Read the note
