Summary

Foundational properties

Property	Description
Statelessness	The LLM remembers nothing between turns. The whole conversation is reassembled and resent on every message, so anything the agent must keep has to live on disk.
Non-determinism	The same input always produces different output. The LLM generates plausible text, not correct text.
Degradation	The longer the input, the less precisely the LLM attends to each part. Earlier instructions get buried.
Truncation	When input exceeds the context limit, older content is summarised or dropped. Information is silently lost.

Engineering principles

#	Principle	Chapter	Software engineering name	One-line summary
1	Only build workflows around things you already do well	4	—	You must be an expert at the process before you automate it.
2	One agent, one job	5	Unix philosophy	Separate work into distinct agents with separate folders and instructions.
3	Self-containment	6	Encapsulation	Everything the agent needs must be in its folder, and nothing else.
4	Always verify non-deterministic output	7	—	Every change the LLM makes could be wrong; every change must be checked.
5	One session, one task	8	—	Pick one thing, do it, save the results, end the session.
6	Principle of least privilege	9	Least privilege	Give the agent only the access it needs and can reliably handle.
7	Reduce the surface where non-determinism can act	11	Attack surface reduction	Move procedures into skills, data into separate files, and mechanical work into scripts.
8	Practise supply-chain hygiene	12	Supply-chain security	Every dependency you import is a trust decision. Read, learn, then build your own.
9	Change one thing at a time	15	One Variable at a Time (OVAT) / bisection	When you want to know what caused a change in behaviour, only one thing should change between the working state and the broken one.

Antipatterns

Antipattern	Chapter	What goes wrong
Auto-memory	10	Claude Code’s automatic memory writes rules outside the project folder without review. Casual remarks get escalated to commandments, stale snapshots become lies the agent follows. Breaks self-containment, review, and least privilege.
Mixing tasks in one session	8	Data from a finished task stays in context and contaminates the next one. Thirty messages that say ‘EUR’ are louder than one line in CLAUDE.md saying ‘use USD’.
Overloaded CLAUDE.md	11, 14	Putting all instructions, data, and procedures in CLAUDE.md loads everything every session. Wastes attention and tokens on content the current task doesn’t need.
Using LLM for mechanical work	11	Having the LLM read, parse, count, or sort when a deterministic script could do it. Adds non-determinism to steps that don’t need judgement, costs more, and breaks on model updates.
Enforcing rules via CLAUDE.md instead of hooks	13	Pattern-matching checks (em-dashes, date formats, banned phrases) as CLAUDE.md instructions that the LLM can ignore. A hook running a linter cannot be ignored.
Giving the agent irreversible access	9	Letting the agent send emails, post to APIs, or interact with services where mistakes can’t be undone. A sent email has no git revert.
Desktop agents with full system access	9	Tools like Claude Cowork that access the entire computer, browser, and email. The blast radius is not one folder but everyone the agent can reach.
Importing third-party skills directly	12	Installing someone else’s skill without review. No audit gate, no test harness, and the same skill produces different output on different machines due to non-determinism.
No bounds or circuit breakers for costs	14, 16	Without spending limits or automatic stops, costs compound silently. In one case, two agents in a multi-agent setup got trapped in a dialogue loop for eleven days, running up tens of thousands of dollars before anyone noticed.

Payoff

The principles above limit what non-determinism can damage. One chapter covers what the discipline gives back: once your work is contained in agents, the accumulated record lets you analyse your own work.

Payoff	Chapter	What it enables
Analysing your own work	17	The contained work becomes a complete, version-stamped record of inputs, processing and outputs. It supports self-assessment, job analysis, delegation decisions, and period-over-period comparison. The agent assembles the record; you supply the judgment.