Patterns

Production techniques that transfer. Each one was paid for by a failure in a running system, and each is written to stand on its own — you should not need to know anything about the system it came from.

They are deliberately free of client, product and vendor identity. Where a concrete implementation matters it is named as an example, never as a prerequisite.

The contract

Every pattern has six sections, in order:

Section	What it must do
Problem	the failure, concretely, including how it presents
Technique	what to do, specifically enough to apply
When it applies	the conditions that make it worth the cost
When it does NOT apply	where it is wrong, and what it does not cover
Evidence	the scar it came from
See also	related patterns

When it does NOT apply is the section that makes the rest trustworthy. A technique with no stated limits is a slogan, and a reader cannot tell whether their case is one the author considered.

Running agents in production

Pattern	The short version
Keep stdout clean on a stdio protocol channel	Redirect logging to stderr before any other import, or a library's stray print corrupts the protocol
A resilience ladder for streaming agent calls	Skip parse errors, retry process errors, fall back to single-turn — three failures that behave differently
Capture the subprocess stderr you were told to check	The library reports "check stderr" and discarded it by default

Generated content you intend to keep

Pattern	The short version
Validate generated artifacts, then regenerate rather than discard	Conversational leakage is a shape, not a phrase; re-ask with the failure fed back
Inject the literal date, never "today"	The model has no clock and will answer anyway
A gate ladder for auto-merging generated changes	Size cap, CI, read-only reviewer, merge-or-comment — each disqualifying
Never let an automated actor write to a protected branch	Enforce it in the writer class, not the credential

Data and pipelines

Pattern	The short version
MERGE on identity alone, then SET the rest	The MERGE pattern must contain exactly what the uniqueness constraint covers
Incremental index with an automatic full-rebuild fallback	Every failure mode routes to rebuild rather than to a subtly wrong index
The submit-poll-postprocess-review pipeline	Four teams built this independently and none of them named it

Contributing

A pattern earns its place by having cost something. If you cannot write the Evidence section from a real failure, it is a preference, not a pattern — and if you cannot write When it does NOT apply, you have not finished thinking about it yet.