In one month I rewrote auth around passkeys, changed how an agent loop stored state, and replaced instructions that had been useful a week earlier. The parts I expected to keep were the parts I kept replacing.
The things that stayed were less glamorous: session boundaries, tests, a deployment target, and the user action I needed to preserve.
These are my notes on five artifacts I now expect to replace. Not a stack prescription. A list of outputs I stopped mistaking for assets.
The liquid parts (prepare to swap):
01 .. Agent Harnesses
I have moved the same work between agent runners because one exposed the controls I needed and another did not. The useful boundary was the repo, tests, and task files. Not the runner.
02 .. Models
A prompt that behaves with one model can wobble with another. I test the outcome again when I change the model. I do not treat a passing run as portable evidence.
03 .. Source Code
Source code is tokens spent to produce behavior. I can spend them again quickly now. That does not make replacement safe. The gate, data migration, constraints, and user-visible behavior are the parts I have to defend.
04 .. Docs
Docs are source. When behavior changes, they move with it or they drift. Why a constraint exists is memory, context, or policy. It belongs somewhere built to survive a rewrite, not beside the artifact it explains.
05 .. Prompts
including skills, rules, CLAUDE.md files, reusable instructions
In the deja experiment, careful instructions still let a weak gate pass broken behavior. The prompt was not the proof. Since then I keep prompts shorter and spend more attention on checks an implementation cannot bluff.
Why I'm Looking at This
I maintain my own client code, product code, deployment setup, and support trail. A fast rewrite is useful only when I can tell whether it broke something I still have to answer for.
So this is the division I am trying: artifacts are liquid. Be willing to replace the runner, model, implementation, docs, and prompt. Be stubborn about intent, constraints, authority, data, and proof.
I have evidence for the replacements I made. I do not yet have evidence that this is a general stack model.
Further Reading
- Boris Tane: The SDLC Is Dead .. Not faster. Dead.
Related
- Compaction .. What survives when the context window resets.
- gate-review .. the weak-test failure that changed where I put trust.
- Loop .. a small runnable version with files as state.
- CAMPFIRE .. The context window is a campfire with a token limit.