Noah Braun

AI Ops @Tesla. Building at Project Carbon Fiber to end the resume black hole. Launching in October. Join the waitlist:

[Noah Braun @noahbraun127]

https://t.co/dXAJPdRKT5 maps your real experience directly to the requisition schema—eliminating ingestion drops in ~30 seconds. Waitlist: https://t.co/dXAJPdRKT5

[Noah Braun @noahbraun127]

The code is disposable, the constraint is what survives to the next prompt. What do you track between revisions, the wording or the rule?

[Noah Braun @noahbraun127]

Engineers now rewrite prompts many times before shipping, but almost none of those edits touch code syntax. Whats actually being revised is the constraint underneath. Here's the mechanism:

[Noah Braun @noahbraun127]

- Pass 1: vague request, plausible looking output
- Pass 5: named edge case (null field, timeout, retry cap)
- Pass 9: explicit invariant the output must satisfy every time

[Noah Braun @noahbraun127]

The criteria outlives the code. The code gets thrown away and regenerated more casually than anyone expects. Worth keeping for the next spec you write before any code exists.

[Noah Braun @noahbraun127]

Most engineers dont write implementations anymore. They write the acceptance criteria and let the code get regenerated until it passes. Here's the mechanism:

[Noah Braun @noahbraun127]

- Invariant list written first: no nulls, idempotent retries, fail-closed timeouts
- Generator produces a candidate function against that list
- Candidate checked against the list, not read line by line

[Noah Braun @noahbraun127]

The invariant is the real deliverable now. The generated code is just one candidate answer to it. Which constraints do you write into a prompt versus catch later in review?

[Noah Braun @noahbraun127]

A vague AI prompt and a broken AI prompt are not the same problem. A vague one produces code you cant test at all. Here's the mechanism:

[Noah Braun @noahbraun127]

- Vague prompt: infinite valid outputs, zero falsifiable claims
- Invariant: field != null, op.idempotent = true, on_error = fail_closed
- Constrained prompt: one claim, one test, pass or fail

[Noah Braun @noahbraun127]

The doc never answered any of that. The code did, three different ways, in one afternoon. Keep this thread handy for your next interface review.

[Noah Braun @noahbraun127]

An approved design doc doesnt prove your team agrees on the interface. It just proves nobody built it yet. Here's the mechanism:

[Noah Braun @noahbraun127]

- Implementation A assumes retries are idempotent
- Implementation B assumes null means empty
- Implementation C assumes strict ordering on writes

[Noah Braun @noahbraun127]

The debate shifts from who argued longest to what the system actually did under load. Does your team build competing branches, or still settle architecture with a meeting?

[Noah Braun @noahbraun127]

Most architecture debates get won by the most confident voice, not the correct one. Parallel throwaway branches change that before a spec even exists. Heres the mechanism:

[Noah Braun @noahbraun127]

- 2 to 3 full branch implementations of the same feature
- Each branch built on a different architecture assumption
- All branches tested against identical data, load, and edge cases

[Noah Braun @noahbraun127]

The deeper mechanism: AI output still needs constant human judgment, and some argue that correction loop runs faster face to face than over chat. Watch for this becoming the next return to office argument. Does proximity actually speed up AI correction, or just feel faster?

[Noah Braun @noahbraun127]

Hacker News is discussing a post arguing offices are coming back specifically to use AI in person, the logic being that fast, informal correction loops on AI output need bodies in the same room, not better tools.

[Noah Braun @noahbraun127]

The winner isn't the best presenter. It's the variant that survived the same traffic as the other two.

[Noah Braun @noahbraun127]

Architecture debates isnt won by whoever is right. They're won by whoever argues longest. Running three AI-built variants of the same feature changes that math entirely. Here's the mechanism:

[Noah Braun @noahbraun127]

Three variants, one test harness:

- Same feature spec, three architectural assumptions
- Each version runs against identical test traffic
- Failures get logged: timeouts, drops, memory spikes