← karajancode.com

The method

The Karajan method, in seven pieces

What a senior brings is judgment and discipline. Judgment no system can guarantee; discipline it can. Each piece below turns one habit of a good team into a mechanism you cannot skip.

writes reviews diff 82a5504
01

Mandatory cross-review

Every change is reviewed by an AI different from the one that wrote it. The verdict is bound to the sha256 of the exact diff: change a single line and the verdict is void — a new review is required. Nothing enters self-approved, and a green that was not earned cannot exist.

Read: the git gates →
pass, or no commit
02

Rules as mechanism

Norms don't live in the prompt, where they are forgotten. They live in git guards and in the project's policy: it passes, or the commit does not enter. And what is marked security is never overridable — not by arbitration, not even by the author of the system.

Read: hardening against AI →
03

The Sentinel

A watchman inside the session checks every command before it runs and blocks the forbidden action in the moment — not at the end, when the damage is done. And it explains why, with a link to the exact rule, so the block teaches instead of just stopping.

Read: the Sentinel →
04

Claims with a source

Every hard fact the agent is about to publish is cross-checked against what its own session actually executed: backed, unbacked, or disproven by its source. A fact its own terminal disproves is a proven hallucination — caught and rejected before it ships.

Read: working with your agent →
each links the last
05

The hash-chained ledger

Everything executed and decided — commands, verdicts, exceptions — is sealed in a ledger where each entry carries the cryptographic fingerprint of the previous one. Neither an agent nor a human can rewrite what was done or decided without breaking the chain, and a broken chain shows.

Read: the HU-Board →
06

Trust expires

The project declares its guarantees — the suite passes, there is a recent audit — and the Steward re-verifies them periodically, so stale evidence stops counting. Because projects don't rot when someone is watching; they rot exactly when nobody is.

Read: working with your agent →
07

Project memory (RAG)

The agent queries the indexed code before it assumes — never guessing what the codebase does. A sensitivity policy decides which model may see what, and the secrets are excluded from the index, so the memory helps without leaking.

Read: the MCP server →