Trigger
/fomo-qa dogfood today (2026-07-31, main@e58fef9, mock persona sample_ai_holder, first_review, zh-TW) hit real friction at the cash-anchor ask (input.cash_anchor.status: "absent" → step 6's "what's your current cash balance" question). The user's first answer to "現在的現金餘額(USD 數字)" was 30% — not a dollar amount. A first clarification round re-stated "這題要的是實際美元數字,不是百分比" and got a second still-ambiguous answer (30% 部位是美元). A third, more constrained multiple-choice clarification was needed before the user's actual intent surfaced: cash is roughly 30% of the current position market value, not a literal $30.
Three question/answer round-trips to resolve one pre-card input is exactly the friction #357 and #507 are trying to keep out of this beat — it happened on the one input this product still asks as free-form prose rather than a structured contract, and it burned turns on a mock persona where nothing was actually at stake.
Why this is a wording/contract gap, not user error
The prompt asks for "current cash balance" but portfolio apps commonly express cash both ways — an absolute balance, or a percentage of total account value ("30% cash") — and nothing in the ask disambiguates which one is wanted or offers the second as a first-class path. A user thinking in percentage terms (which is a completely normal way to reason about an account) has no signal that the field wants dollars until they get corrected.
Proposal (for discussion, not decided)
At the cash-anchor ask (flows/first-review.md step 6 / flows/weekly-review.md equivalent), do one of:
- Wording fix only: make the ask explicit that it wants an absolute currency amount, with a short example in the same sentence (e.g. "現在的現金餘額大概是多少(美元金額,例如 8000;不是百分比)"). Cheapest, no schema change.
- Accept both, convert to a stored amount: let the user answer either an absolute amount or "N% of the current position value," and have the agent (or a small engine helper) compute the dollar figure from the already-known position market value before calling
add-cash — but the engine should own the arithmetic and disclose the derivation, not have the agent silently do unverified mental math.
- Leave it as free text but add one clarification turn to the fixed budget specifically for this ask, so a wrong-format answer doesn't have to compete with the question's normal
max_clarifications.
Evidence
Mock-only session, no real account data. Session id and exact Q&A available in this dogfood's ux_receipt trace under the isolated dogfood coach root if needed for follow-up.
Privacy
No real trade data — mock persona sample_ai_holder.csv only.
Trigger
/fomo-qadogfood today (2026-07-31,main@e58fef9, mock personasample_ai_holder, first_review, zh-TW) hit real friction at the cash-anchor ask (input.cash_anchor.status: "absent"→ step 6's "what's your current cash balance" question). The user's first answer to "現在的現金餘額(USD 數字)" was30%— not a dollar amount. A first clarification round re-stated "這題要的是實際美元數字,不是百分比" and got a second still-ambiguous answer (30% 部位是美元). A third, more constrained multiple-choice clarification was needed before the user's actual intent surfaced: cash is roughly 30% of the current position market value, not a literal$30.Three question/answer round-trips to resolve one pre-card input is exactly the friction #357 and #507 are trying to keep out of this beat — it happened on the one input this product still asks as free-form prose rather than a structured contract, and it burned turns on a mock persona where nothing was actually at stake.
Why this is a wording/contract gap, not user error
The prompt asks for "current cash balance" but portfolio apps commonly express cash both ways — an absolute balance, or a percentage of total account value ("30% cash") — and nothing in the ask disambiguates which one is wanted or offers the second as a first-class path. A user thinking in percentage terms (which is a completely normal way to reason about an account) has no signal that the field wants dollars until they get corrected.
Proposal (for discussion, not decided)
At the cash-anchor ask (
flows/first-review.mdstep 6 /flows/weekly-review.mdequivalent), do one of:add-cash— but the engine should own the arithmetic and disclose the derivation, not have the agent silently do unverified mental math.max_clarifications.Evidence
Mock-only session, no real account data. Session id and exact Q&A available in this dogfood's
ux_receipttrace under the isolated dogfood coach root if needed for follow-up.Privacy
No real trade data — mock persona
sample_ai_holder.csvonly.