The most useful AI workflow tricks are usually not the flashy ones.
They are the little control surfaces. The subtle moves that keep a session from collapsing into drift, repetition, or polite nonsense.
That is how I think about steer in Codex and /btw in Claude.
Not as magic.
As small interventions that protect context.
The Real Problem Is Drift
You start with a live project, a strong idea, a repo full of history, and a decent sense of what you want.
Then an hour later the AI is still "helping," but the work has begun to slide sideways.
The code is technically fine but tonally off.
The copy is clean but generic.
The structure solves the wrong problem.
The assistant is no longer following the shape of your intent. It is following the shape of its own last few responses.
That is drift.
Most people try to solve it by dumping more instructions into the top of the conversation. Sometimes that works. Often it just makes the session heavier.
The better move is often smaller.
The Side-Channel That Changes How the Main Road Gets Read
I use Claude constantly. Test it, red-team it, build with it, write with it. Claude has been watching my work for a while now and has not called yet, but the relationship is functional.
I think of /btw as a side-channel for important nuance.
Not the main road. The aside that changes how the main road gets read.
The useful part is not the syntax. The useful part is the pattern: you inject a clarifying aside before the model hardens around the wrong assumption. Models overcommit once they think they understand the assignment. An honest "by the way" can save twenty minutes of cleanup.
Here is the range of what that actually looks like in practice.
Tone correction before it calcifies:
/btw this should read like it was written angry, not processed
The model will otherwise smooth everything into editorial calm. That one line protects the original voice.
Hidden constraint the files don't reveal:
/btw the client is non-technical — don't surface this logic in the interface
/btw we're not shipping this month — optimize for clarity, not performance
The code doesn't know the business situation. The /btw bridges that gap without restructuring the whole prompt.
Audience context mid-session:
/btw the reader has already seen the technical version — skip the basics
/btw this is for security professionals, not general readers — don't soften the red team section
Emotional truth behind the design decision:
/btw this feature exists because the founder hates the competitor's version, not because users asked for it
That context changes every suggestion the model makes about tone, positioning, and framing.
The guardrail is the point:
/btw the refusal behavior here is the interesting part — I'm not trying to bypass it, I'm documenting it
This one matters in red team work. Without it the model sometimes treats the research as the threat.
Where /btw Has Paid Off Most: Game Design
This is where I have relied on it most.
Game design is a domain where the gap between what looks correct in a document and what feels right in a playtest is enormous. You can spec a mechanic perfectly and have it collapse the first time a real person touches it. The model cannot know that from the code or the design doc. You have to tell it.
Building Neon Leviathan and Hack, Love, Betray, playtesting surfaced things that no amount of structured prompting would have caught — and /btw became the fastest way to inject that lived information back into the session without losing the accumulated context.
