Why we don't ship a chat box (and what we ship instead).

Open most "AI trading apps" launched in 2026 and the first thing you see is a chat input. Ask anything. Swing Deck deliberately doesn't. Not because chat is bad technology — it's because chat is the wrong UX for a discipline tool. A chat box lets you talk yourself into trades. Surface-bound AI talks you out of them.

The chat-box failure mode

The pattern you'll see if you trade with a generic LLM open in another tab: you have a setup. You're 70% sure. You ask the model "what do you think of NVDA at $211?" and the model — which is helpful, polite, and completionist by training — gives you a thoughtful, balanced response that includes both the bull case and the bear case. You read the bull case more carefully because that's the one that matches your prior. You take the trade.

Twenty trades later, you've reinforced your existing framework with what feels like an outside perspective but is actually a mirror. The chat didn't make the call worse. It made you more confident about a call you were already going to make. The selection bias is one-sided: you ask the chat when you're already leaning bullish. You don't ask when you're already leaning bearish, because the bearish read is the one you avoid in the first place.

A chat box, in a discipline tool, is a confirmation-bias amplifier wearing a verification costume.

The Bloomberg analogy

Power users who actually pay for trading terminals don't have chat boxes. They have bound function keys. F8 brings up the pricing page. F9 brings up news. They learn the keys, they fire the keys at the right moments, and the data comes back in the same shape every time. That predictability is the whole point. Chat boxes have a different shape every time, which is fine for exploration and lethal for repeated decisions.

The thing power users want from a trading app is the opposite of what consumer LLM apps optimize for. Consumers want flexibility — say anything, ask anything, get a customized answer. Power users want bounded surfaces — defined inputs, predictable outputs, no surprises. Bloomberg has been right about this since 1981.

The Surface-Bound AI Principle

What we ship instead of a chat box: ten AI surfaces, each bound to a specific moment in the trading workflow. Bound means three things:

• Triggered, not summoned. Each surface fires only when a specific framework event happens. You don't open a chat to ask about an exit — the Exit Coach reveals itself when your position hits a TP rung or breaches the chandelier stop. You don't ask if the entry's clean — the Entry Coach reveals itself when shares == 0 and the entry trigger is ready.
• Fixed input set. Each surface gets a defined slice of state — pillar IDs, your trade log, the regime, the per-ticker scores — and nothing else. The user can't shape the input by phrasing.
• Fixed output shape. Each surface returns a known structure. JSON-then-render for the heavy ones, structured paragraphs for the narrative ones. You can A/B test outputs across a population. You can't A/B test "the response to whatever the user typed."

The corollary: when no trigger is firing, the AI is invisible. There's no panel asking to be talked to. The dashboard goes quiet. No panels means nothing to do, and that's the feature.

The eight surfaces

Each one fires on a specific event. None of them are reachable except through their trigger. The full list:

1. Pillar Coach. Fires when a risk pillar veto is active for one of your positions. Narrates the veto, cites your history, says what to do.
2. Exit Coach. Fires when shares > 0 and price has hit a TP rung or breached the chandelier stop. Tells you which scale-out tier to take and flags any pattern in your past exit behavior on this ticker.
3. Entry Coach. Fires when shares == 0 and the entry trigger reads as ready (W+D ✓, score ≥80, sleeve room, no earnings within 3 sessions). Confirms the armed state and recommends sizing.
4. Devil's Advocate. Fires automatically alongside every AI thesis generation. Reads the same inputs as the bull thesis and writes the bear case. Pair them side-by-side; never read just one.
5. Thesis Drift. Fires when your thesis tone shifts ≥N% bullish over recent generations while the underlying score has dropped ≥N points. The yellow flag for "you're talking yourself into something the data is talking you out of."
6. Position History Audit. Fires when you expand the deep-dive panel on a ticker you've traded before. Tells you how you've actually traded this name historically — entries, exits, deviations from framework — pulled from your trade log.
7. Comparable Setups. Fires on audit refresh when a near-match historical setup is found. "Last time this ticker scored 78 with VIX at 18 and W+D ✓, the next 10 sessions did X."
8. Catalyst Interpreter. Fires when new headlines arrive for a ticker you hold or have on the watchlist. Returns {material:int, noise:int, sentiment, magnitude} plus prose.

Plus two later additions in the v6.3-v6.4 line: the Whale Confirmation Coach (verifies the framework's whale_sentiment score against SEC filings + dark-pool reports + analyst sentiment) and the Trap & Structure Coach (six-gate check: whales, gamma, dark tape, structure, exhaustion, timing).

Notice what's missing from the list: "answer this question." There's no surface that takes free-text input. Every surface has a trigger that defines what it answers.

The prompt-design rules that follow

Once you commit to bound surfaces, your prompt design changes shape:

The bottom row matters most. A chat box can never tell you something you didn't think to ask, because it only fires when you ask. A bound surface can — because it fires when its condition matches, not when your attention matches. The Thesis Drift coach is the cleanest example: it fires when your tone has drifted bullish while the data drifts bearish. You'd never type "am I rationalizing?" into a chat box. You'd never ask the question because the rationalization is invisible from the inside. The surface fires because the metric crosses, regardless of whether you're paying attention.

The per-call cost numbers

Bounded surfaces have one concrete economic property worth calling out: per-call cost is finite and predictable. From the v6.3 calibration window:

A typical trading session fires 3-6 of the heavy coaches and 5-15 of the light ones — call it $0.40 to $0.80 per session of LLM cost. With a daily budget cap (default $5/day, configurable in llm_budget.py) the worst-case is bounded. There's no "user typed a 2,000-word question" risk because there's no input box where users type anything.

Compare to a free-text chat box where a single rambling question can run a dollar of tokens just by being verbose. In a discipline tool that's used dozens of times per session, that adds up to numbers that change which model you can afford to use.

The trade-off, named

Bound surfaces have one obvious downside: they can only answer the questions you've defined. If you have a novel question — "is there a sector pattern in this week's drawdown that the framework hasn't surfaced?" — the bound surfaces don't have a hook for it. A chat box would let you ask. We don't.

That's a real trade-off and we've kept making it for the same reason: the cost of getting that occasional novel insight is that 95% of your AI usage becomes confirmation-biased self-talk. The math doesn't work. Disciplined traders need the AI to refuse to answer the questions they wish it would, more than they need it to answer the ones they happen to ask.

The dashboard reflects this commitment: there's no chat input anywhere on swing-deck.com or the local framework. There never will be. The product the framework refuses to ship is what makes the framework defensible.

The math floor, not motivation

Other trading apps will keep shipping chat boxes because chat boxes test better with new users. The first session feels magical — "oh, I can ask anything!" — and the retention curve looks great for two weeks. Then the user has had thirty unprofitable conversations and the curve breaks.

Swing Deck's first-session experience is harder. There's no input box waiting for your question. You have to actually trade and let the surfaces fire. The Pillar Coach reveals itself the day you have a position in distress. The Devil's Advocate fires the moment you generate a thesis. The pace of the framework, not your urge to ask, sets the cadence.

That's slower to demo and slower to onboard. It's also the only UX shape we've found that doesn't reliably convert traders into worse traders.

Related: The ten AI surfaces — full I/O contracts · Reading the Whale Confirmation Coach output · The 13 risk pillars (the framework the surfaces narrate)