Notes on software a company can own.
What we learn mapping processes, writing contracts before code, and running software that stays governed after the first release.
-
LLM cascade: basal before Opus for under a third of the cost
LLM cascade test: basal, a Polish open-weights model, as a filter in front of Claude Opus. 98.5% accuracy against 98.6%, at 25 to 31% of the cost.
-
AI agents for marketing: I stopped approving every step
AI agents for marketing, 5 weeks in: 14 Claude Code agents, 3 approval lanes, and why I now approve only what is hard to undo. With a copyable CLAUDE.md block.
-
KPI tree template: connect a company goal to team metrics
A KPI tree template with parent IDs, 2 worked examples, formula links vs hypotheses, and how to use the tree in a weekly review. Copy it into a spreadsheet.
-
Company brain AI hears the chat. Who writes down the rest?
Company brain AI can recall every thread. 4 other layers of company knowledge need writing down: a 5-layer model, survey numbers and a 10-decision test.
-
Leading vs lagging indicators: 30 examples for operations
Leading vs lagging indicators for operations: 30 examples of leading and lagging KPIs in 15 pairs by team, and a simple test that a leading one really leads.
-
Claude Code context window full: what to do, what is lost
Claude Code context window full? A summary replaces your conversation. What survives, what is lost, and the 5 commands to use before it happens.
-
When AI speeds up the work, how do you keep teams aligned?
How to keep teams aligned when AI speeds up the work: what a goal means for each team, who needs to change their plan, and when the plan should change.
-
MEMORY.md vs CLAUDE.md vs AGENTS.md: which file does what
MEMORY.md vs CLAUDE.md is a question of who writes the file. What each one loads, when Claude Code reads AGENTS.md, and how to use all 3 without duplicates.
-
AI agent memory architecture: what to keep, and where
AI agent memory architecture in practice: the 4 kinds of memory my Claude Code agents use, where I keep each one, and what breaks when they mix.
-
Multi-agent orchestration with Claude Code: my setup
Multi-agent orchestration without a framework: how I split marketing work across Claude Code agents that share one memory in Markdown files, and what broke.
-
Custom software vs SaaS: when renting stops fitting the company
Custom software fits one company's process; SaaS fits the average. 4 questions that show which side your company is on, and what a third option costs.
-
Who is accountable for AI-generated code in your company
A tool built with AI works until the person who prompted it leaves. Four things nobody wrote down, why a PDF policy is not evidence, and what to check this week.
Bring the process, not the pitch.
A 30-minute fit call on the one process your team keeps working around. If standard software is the better answer, we say so.