Back to Details

clarify-decisions

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Clarify Decisions

决策澄清

What this is: a reusable interview protocol, not a pipeline stage. Nested under a parent (e.g. 
frame-change
step 2, 
define-project
, 
triage
) you stay in that parent's conversation and checklist — apply these rules, do not announce a mode switch, do not treat the parent as finished when your item is checked off, and run the parent's checklist per Todos below. Standalone (the user asked to be grilled with no parent) you own the interview alone until shared understanding.
Interview until you both hold the same picture: every silent assumption that would become debt or a wrong architecture choice is named and decided. Leading words for this skill: open set, territory, card, close package. The map (prompts, plans, knowns) is not the territory (codebase, runtime, users, history) — clarify-decisions shrinks that gap before wrong guesses get expensive.
这是什么： 一个可复用的访谈协议，而非流水线阶段。如果嵌套在父级流程下（例如
frame-change
步骤2、
define-project
、
triage
），请保持在父级对话和检查清单中——遵循以下规则，不要宣布切换模式，完成当前项后不要将父级流程标记为已完成，并按照下方“待办事项”执行父级检查清单。如果是独立使用（用户主动要求接受盘问且无父级流程），则由你主导整个访谈，直到双方达成共识。
访谈需持续到双方对内容达成一致：所有可能演变为技术债务或错误架构选择的隐性假设都被明确并做出决策。本skill的核心关键词：open set（开放集）、territory（实际场景）、card（问题卡片）、close package（收尾包）。地图（提示、计划、已知信息）不等于实际场景（代码库、运行时、用户、历史情况）——决策澄清技能旨在减少这种差距，避免错误猜测产生高昂成本。

The Iron Law — channel

铁则——沟通渠道

EVERY QUESTION IS INLINE CHAT WITH FULL CONTEXT.
NEVER use AskUserQuestion, structured MCQ pickers, or any harness UI that
truncates labels, option text, or the "why this matters" line.
A tap-friendly UI that strips consequences is not faster — it is a different, worse interview. "House style prefers the picker", "standup in five minutes", "the lead said use the structured UI", and "the option description field is long enough" are not exceptions.
EVERY QUESTION IS INLINE CHAT WITH FULL CONTEXT.
NEVER use AskUserQuestion, structured MCQ pickers, or any harness UI that
truncates labels, option text, or the "why this matters" line.
一个简化的、剥离后果说明的友好UI并非更高效——它是一种更差的访谈方式。“团队风格偏好选择器”“五分钟后开站会”“负责人要求使用结构化UI”“选项描述字段足够长”均不能成为例外。

The Iron Law — open set (no fixed rounds)

铁则——开放集（无固定轮次）

THERE IS NO FIXED ROUND COUNT.
NEVER "Question k of N", "last of 5", or "we budgeted four cards".
Stop only when the open set is empty of judgment calls that change
architecture, data, auth/security, UX flow, or implementation scope.
Open set = high-blast unknowns still undecided + branches the last answer opened + parent known-unknowns still needing a user lock.
Home rule: recompute the open set after every answer, then either the next card or the close package. Every other mention of "recompute" points here.
A pre-listed todo of decision areas is a living map, not a quota: append when a branch opens; drop when resolved; never close because the original list finished while a high-blast item remains. Time, standup, senior "wrap after a handful", and exhaustion change when you report progress — not whether an unstated decision exists.
THERE IS NO FIXED ROUND COUNT.
NEVER "Question k of N", "last of 5", or "we budgeted four cards".
Stop only when the open set is empty of judgment calls that change
architecture, data, auth/security, UX flow, or implementation scope.
开放集 = 尚未决定的高影响未知项 + 上一个回答引出的新分支 + 父级流程中仍需用户确认的已知未知项。
核心规则： 在每个回答后重新计算开放集，然后决定是提出下一张问题卡片还是生成收尾包。所有其他提到“重新计算”的地方均指向此规则。
预先列出的决策领域待办事项是一张动态地图，而非固定配额：当新分支出现时添加项；当决策完成时移除项；绝不能因为原始列表已完成但仍存在高影响项就结束访谈。时间、站会、上级要求“聊几个就结束”以及疲劳感只会改变你汇报进度的时机——不会改变未明确决策是否存在的事实。

Starting map (before the first card)

初始地图（第一张问题卡片前）

Load parent Knowns inventory, Blindspot list, and scan digest when present (e.g. 
.skills/<CODE>/knowns.md /
_pending-<slug>/knowns.md
, `.skills/<CODE>/scan.md / `_pending-<slug>/scan.md
). Then emit one short thought-partner map in ordinary chat — not a question card, not a multi-question dump:
  1. Locked — what you treat as fixed (posture, explicit non-negotiables, digest facts).
  2. Open high-blast — the decision forks you expect to walk first (names only; no options yet).
  3. How you will close unknowns — cards for judgment calls; reference or 
    run-spike
    /
    research
    when the user can only know it when they see it or when the answer is a fact; teach-then-ask when a blindspot blocks a real choice.
Invite a correction only if the map is wrong ("stop me if a lock is false"). Then the first card. Nested under a parent that already stated this map: skip the restate and go to the first card.
加载父级已知信息清单、盲区列表（如果存在），并扫描摘要（例如
.skills/<CODE>/knowns.md /
_pending-<slug>/knowns.md
, `.skills/<CODE>/scan.md / `_pending-<slug>/scan.md
）。然后用普通聊天形式发送一份简短的协作地图——不要用问题卡片，也不要一次性抛出多个问题：
  1. 已锁定项——你视为固定不变的内容（定位、明确的不可协商项、摘要中的事实）。
  2. 高影响开放项——你预计首先要梳理的决策分支（仅列名称；暂不提供选项）。
  3. 如何解决未知项——针对判断类决策使用问题卡片；当用户只有实际看到才能知晓答案或答案为事实时，使用参考资料或
    run-spike
    /
    research
    ；当盲区阻碍真实决策时，先讲解再提问。
仅当地图有误时邀请用户修正（“如果有锁定项不正确，请随时打断我”）。然后发送第一张问题卡片。如果嵌套的父级流程已提供此地图，则跳过重复说明，直接发送第一张问题卡片。

Retrieval package (feature work)

检索包（功能开发）

When this interview is about feature work (neighbors, overlap, reuse), hold a 
load-subgraph
retrieval package (envelope + seeds + fingerprints + schema/recipe) for Territory and grounded claims:
ModeRule
Nested under a parent that already produced a packageReuse that package if it remains valid (same seeds, source fingerprints sha256+present, schema/recipe). If invalid or missing → rederive via REQUIRED SUB-SKILL: use 
load-subgraph
.
Standalone (no parent package)Load retrieval once before the first interview card (REQUIRED SUB-SKILL: use 
load-subgraph
with terms/paths/CODE seeds).
In progressIf any derivation source input changes, or material scope / terms / paths change, or fingerprints differ → rederive. Do not re-run every card when the package is still valid.
No on-disk session cache. Grounded claims (one home): follow 
skills/execution/load-subgraph/references/grounded-claims.md
for every conclusion from the package.
当本次访谈针对功能开发（关联模块、重叠部分、复用逻辑）时，需持有一个 
load-subgraph
检索包（信封 + 种子 + 指纹 + schema/recipe），用于获取实际场景和有依据的结论：
ModeRule
Nested under a parent that already produced a packageReuse that package if it remains valid (same seeds, source fingerprints sha256+present, schema/recipe). If invalid or missing → rederive via REQUIRED SUB-SKILL: use 
load-subgraph
.
Standalone (no parent package)Load retrieval once before the first interview card (REQUIRED SUB-SKILL: use 
load-subgraph
with terms/paths/CODE seeds).
In progressIf any derivation source input changes, or material scope / terms / paths change, or fingerprints differ → rederive. Do not re-run every card when the package is still valid.
No on-disk session cache. Grounded claims (one home): follow 
skills/execution/load-subgraph/references/grounded-claims.md
for every conclusion from the package.

Question card (every turn)

问题卡片（每一轮）

Exactly one decision per message. Emit this shape in ordinary chat — not a tool call. Every slot is required; thinning under time pressure is a channel violation.
  1. Radius — one of: 
    architecture
    · 
    data
    · 
    auth/security
    · 
    UX flow
    · 
    polish-diff
    (label it).
  2. Thread — three short lines the user can scan before the question:
    • Locked so far — 1–3 decisions already taken that constrain this fork (or "none yet").
    • This card — the single fork now.
    • Still open after — remaining high-blast names if this were answered (living open set — never "3 of 5").
  3. Territory — grounded facts from the repo, digest, or parent knowns (paths, middleware, prior PRs, current behavior, landmines) — enough that the options make sense. When a blindspot blocks the choice, teach here (what it is, why it bites in this product) before the question. If you truly have no facts, say so; do not invent them. Never ask the user to recall what you can read.
  4. Question — the decision in plain language.
  5. Why it matters — what changes if the answer flips (queue vs sync, schema, permission boundary, scope, reverse cost). Enough for the user to pick an option without a follow-up. Ground in this repo or product; never "just checking".
  6. Closes — unknown class this card retires: 
    known-unknown
    · 
    unknown-known
    · 
    blindspot-confirm
    .
  7. Options (2–4) — short title plus consequence line (gain, pay, break). Bare labels are not options.
  8. Recommendation — your pick, first or clearly marked, with a one-line reason the user can accept in two words or push back on.
  9. Stop. Wait. After the answer: recompute the open set (Iron Law — open set), then next card or close package.
Do not batch questions. The card is the detail — not a teaser for "more if you want".
Exactly one decision per message. Emit this shape in ordinary chat — not a tool call. Every slot is required; thinning under time pressure is a channel violation.
  1. Radius — one of: 
    architecture
    · 
    data
    · 
    auth/security
    · 
    UX flow
    · 
    polish-diff
    (label it).
  2. Thread — three short lines the user can scan before the question:
    • Locked so far — 1–3 decisions already taken that constrain this fork (or "none yet").
    • This card — the single fork now.
    • Still open after — remaining high-blast names if this were answered (living open set — never "3 of 5").
  3. Territory — grounded facts from the repo, digest, or parent knowns (paths, middleware, prior PRs, current behavior, landmines) — enough that the options make sense. When a blindspot blocks the choice, teach here (what it is, why it bites in this product) before the question. If you truly have no facts, say so; do not invent them. Never ask the user to recall what you can read.
  4. Question — the decision in plain language.
  5. Why it matters — what changes if the answer flips (queue vs sync, schema, permission boundary, scope, reverse cost). Enough for the user to pick an option without a follow-up. Ground in this repo or product; never "just checking".
  6. Closes — unknown class this card retires: 
    known-unknown
    · 
    unknown-known
    · 
    blindspot-confirm
    .
  7. Options (2–4) — short title plus consequence line (gain, pay, break). Bare labels are not options.
  8. Recommendation — your pick, first or clearly marked, with a one-line reason the user can accept in two words or push back on.
  9. Stop. Wait. After the answer: recompute the open set (Iron Law — open set), then next card or close package.
Do not batch questions. The card is the detail — not a teaser for "more if you want".

Worked shape

Worked shape

**architecture** · export generation locus

Thread
- Locked so far: comment API stays stable; posture = Run Spike
- This card: where PDF generation runs
- Still open after: guest export auth · stroke storage · plan quota

Territory
- Export is `POST /api/reviews/:id/export`, session-auth only today.
- Gateway idle timeout is 30s; large reviews with drawings already time out
  similar heavy handlers on the request thread.
- Workers already exist for transcode (`jobs/transcode`); no export job yet.

Where should export generation run?

↳ This decides whether we need a job queue, a ready-notification path, and
  an artifacts bucket — or none of those. Large reviews will blow the 30s
  gateway if we stay on the request thread; picking wrong here rewrites the
  API shape and ops surface mid-build.

Closes: known-unknown

- **Sync in the API request** — simplest; a 400-comment export times out.
- **Background job on the existing queue** (Recommended) — reuses the
  transcode worker patterns; needs a "ready" notification.
- **Client-side only** — zero backend; caps formats and helps support less.

Recommended: background job — export size is unpredictable with drawings.
**architecture** · export generation locus

Thread
- Locked so far: comment API stays stable; posture = Run Spike
- This card: where PDF generation runs
- Still open after: guest export auth · stroke storage · plan quota

Territory
- Export is `POST /api/reviews/:id/export`, session-auth only today.
- Gateway idle timeout is 30s; large reviews with drawings already time out
  similar heavy handlers on the request thread.
- Workers already exist for transcode (`jobs/transcode`); no export job yet.

Where should export generation run?

↳ This decides whether we need a job queue, a ready-notification path, and
  an artifacts bucket — or none of those. Large reviews will blow the 30s
  gateway if we stay on the request thread; picking wrong here rewrites the
  API shape and ops surface mid-build.

Closes: known-unknown

- **Sync in the API request** — simplest; a 400-comment export times out.
- **Background job on the existing queue** (Recommended) — reuses the
  transcode worker patterns; needs a "ready" notification.
- **Client-side only** — zero backend; caps formats and helps support less.

Recommended: background job — export size is unpredictable with drawings.

Order and coverage

Order and coverage

  • Blast-radius first. Next open-set item that can change architecture, data model, public API, auth/security, UX flow, or implementation scope — even when the user asks to start with polish.
  • Walk every branch. Dependency order; opened sub-branches before the trunk. Stop rule = open-set empty (Iron Law — open set).
  • Judgment only to the user. Facts load in Territory; only forks that need a human lock become cards.
  • Right-size to posture. When the parent supplies it, or 
    docs/agents/project.md
    has Project posture: skip data migration / backward compatibility / deprecation on Run Spike · Research · Learning; press them on Cut Released · Scaling · Maintenance. Absent posture, walk every branch. Posture and Team band are orthogonal.
  • Package to team band. When 
    ## Team
    has a non-empty roster or a Workflow band override, read band and packaging from that section. Small/Multi: optional ownership/reviewer probes when relevant. Solo or Team absent: no multi-person assignee theater. Never invent a team; never hard-fail for missing Team.
  • Blast-radius first. Next open-set item that can change architecture, data model, public API, auth/security, UX flow, or implementation scope — even when the user asks to start with polish.
  • Walk every branch. Dependency order; opened sub-branches before the trunk. Stop rule = open-set empty (Iron Law — open set).
  • Judgment only to the user. Facts load in Territory; only forks that need a human lock become cards.
  • Right-size to posture. When the parent supplies it, or 
    docs/agents/project.md
    has Project posture: skip data migration / backward compatibility / deprecation on Run Spike · Research · Learning; press them on Cut Released · Scaling · Maintenance. Absent posture, walk every branch. Posture and Team band are orthogonal.
  • Package to team band. When 
    ## Team
    has a non-empty roster or a Workflow band override, read band and packaging from that section. Small/Multi: optional ownership/reviewer probes when relevant. Solo or Team absent: no multi-person assignee theater. Never invent a team; never hard-fail for missing Team.

Pre-implementation interview map

Pre-implementation interview map

Clarify Decisions owns the interview leg of pre-implementation unknowns work. Other legs are open-set sources or handoffs — not extra fixed rounds:
LegClarify Decisions doesDoes not re-own
Blindspot passLoad parent Blindspot; high-blast items → teach-then-ask cards or explicit locks. No parent list + low familiarity → short territory teach-pack on landmines before preference cards.Full scan / knowns inventory (
frame-change
step 1)
Frame Change / scopeIf the real issue is multi-subsystem scope, hand back to parent decomposition.Approach menus and tier (
frame-change
steps 4–5)
InterviewRich cards; blast-radius first (slots and order above).—
ReferencesBest reference is source code (folder, module, prior PR, even another language). Restate semantics; lock accept / adapt / reject. Diagrams and screenshots are weaker fallbacks.Implementing the reference
Unknown knownsNo abstract taste grind. Reference path or parent 
run-spike
/ 
research
, then one card on the result.		Running the run-spike session
Plan readinessClose package high-tweak surface (data model, type interfaces, user-facing flows).Writing 
tasks.md
(
plan-tasks
)
"Just make something sensible" is not a decision while a concrete reference exists: surface it, restate semantics, accept / adapt / reject. Inventing industry defaults is a fact failure.
Clarify Decisions owns the interview leg of pre-implementation unknowns work. Other legs are open-set sources or handoffs — not extra fixed rounds:
LegClarify Decisions doesDoes not re-own
Blindspot passLoad parent Blindspot; high-blast items → teach-then-ask cards or explicit locks. No parent list + low familiarity → short territory teach-pack on landmines before preference cards.Full scan / knowns inventory (
frame-change
step 1)
Frame Change / scopeIf the real issue is multi-subsystem scope, hand back to parent decomposition.Approach menus and tier (
frame-change
steps 4–5)
InterviewRich cards; blast-radius first (slots and order above).—
ReferencesBest reference is source code (folder, module, prior PR, even another language). Restate semantics; lock accept / adapt / reject. Diagrams and screenshots are weaker fallbacks.Implementing the reference
Unknown knownsNo abstract taste grind. Reference path or parent 
run-spike
/ 
research
, then one card on the result.		Running the run-spike session
Plan readinessClose package high-tweak surface (data model, type interfaces, user-facing flows).Writing 
tasks.md
(
plan-tasks
)
"Just make something sensible" is not a decision while a concrete reference exists: surface it, restate semantics, accept / adapt / reject. Inventing industry defaults is a fact failure.

Close package (required)

Close package (required)

When the open set has no remaining high-blast judgment call — and before returning control to a parent or claiming shared understanding — emit:
  1. Decisions table — rows: radius · topic · decision (user's words) · unknown class closed.
  2. Constraints block — ready-to-paste locks (architecture and data first; polish-diff last). Flag lower-radius answers that conflict with higher-radius locks.
  3. High-tweak surface — locks most likely to change under real implementation pressure (data model, type interfaces, UX flows). Mechanical refactors stay buried; do not re-interview them here.
  4. Explicit confirmation — is this the shared picture? Only an affirmative on this package counts.
Not confirmation: "any other questions?", "we're aligned, skip the table", "just go write requirements", senior pressure to skip ceremony, or silence. If they correct a row, edit and re-confirm. If confirmation opens a new high-blast fork, return to cards.
Do not enact anything — no production code, no scaffolding, no plan execution — until that confirmation lands. (Glossary/
CONTEXT.md
updates via 
define-domain
as a passive side effect are allowed when a term settles mid-interview.)
When the open set has no remaining high-blast judgment call — and before returning control to a parent or claiming shared understanding — emit:
  1. Decisions table — rows: radius · topic · decision (user's words) · unknown class closed.
  2. Constraints block — ready-to-paste locks (architecture and data first; polish-diff last). Flag lower-radius answers that conflict with higher-radius locks.
  3. High-tweak surface — locks most likely to change under real implementation pressure (data model, type interfaces, UX flows). Mechanical refactors stay buried; do not re-interview them here.
  4. Explicit confirmation — is this the shared picture? Only an affirmative on this package counts.
Not confirmation: "any other questions?", "we're aligned, skip the table", "just go write requirements", senior pressure to skip ceremony, or silence. If they correct a row, edit and re-confirm. If confirmation opens a new high-blast fork, return to cards.
Do not enact anything — no production code, no scaffolding, no plan execution — until that confirmation lands. (Glossary/
CONTEXT.md
updates via 
define-domain
as a passive side effect are allowed when a term settles mid-interview.)

Todos

Todos

Nested: no competing list. You run inside the parent's checklist — interview item stays in-progress until the close package is confirmed. Open-set progress is that item's progress, not a second channel.
Standalone: a living open-set list of decision areas is fine — still one card per message; still recompute after each answer (Iron Law — open set). If a parent skill is already in flight, never open a second channel.
Nested: no competing list. You run inside the parent's checklist — interview item stays in-progress until the close package is confirmed. Open-set progress is that item's progress, not a second channel.
Standalone: a living open-set list of decision areas is fine — still one card per message; still recompute after each answer (Iron Law — open set). If a parent skill is already in flight, never open a second channel.

Rationalizations

Rationalizations

ThoughtReality
"House style / the lead said use AskUserQuestion"Channel is the Iron Law. Inline cards are the interview; pickers truncate the why and the consequences.
"Standup in five minutes — short labels only"A truncated decision is slower than one clear card. Time pressure changes when you report, not what a decision needs.
"Option description field is long enough"If the tool caps text, it is the wrong channel. Full context goes in chat.
"I'll AskUserQuestion and also paste context"Dual channel. One inline card; no picker.
"Recommended + one-line reason is enough"Without Thread, Territory, and consequences per option, the user cannot analyze — only accept a default.
"Keep Why to one line so it stays snappy"Rich enough to decide without a follow-up is the floor. A sentence budget is not a thinness license.
"Context can be a follow-up if they ask"The card is the detail. Follow-up-only context is a thin-card failure.
"We finished the 4 areas on the todo — close"The todo is a living map. Open-set empty is the stop; precommitted N is not.
"Question 3 of 5, then package"No fixed N. Countdown framing is a red flag.
"User asked for button color first"Blast-radius first still holds. Polish Diff after architecture, data, and auth forks.
"We're aligned — skip the decisions table"Shared understanding is the package + yes. Alignment theater without the table is not confirmation.
"Senior said just write requirements"User/senior can override process ownership; they cannot make an unstated decision exist. Emit the package; get the yes.
"I'll assume the safe default and mark done"Assumptions are not decisions. One card; wait.
"Just pick industry best practice — they said sensible"Look up the territory reference first; restate; lock with the user.
"Senior said switch cleanly into clarify-decisions and park the parent"Nesting is the clean switch. Parking the parent and opening a clarify-decisions checklist is dual-channel thrash.
"A short decision checklist under clarify-decisions isn't a competing list"It is a second list. Decision areas live as the parent's in-progress interview item.
"Announce Using clarify-decisions so the user sees the write-handoff"Nested: no mode-switch announcement. Standalone (no parent): you may name clarify-decisions once.
"Parent already loaded neighbors — re-run every card for freshness"Reuse the valid package; rederive only when fingerprints/seeds/scope change
"Standalone interview — skip load-subgraph, Territory is enough"Feature work: load once before the first card
ThoughtReality
"House style / the lead said use AskUserQuestion"Channel is the Iron Law. Inline cards are the interview; pickers truncate the why and the consequences.
"Standup in five minutes — short labels only"A truncated decision is slower than one clear card. Time pressure changes when you report, not what a decision needs.
"Option description field is long enough"If the tool caps text, it is the wrong channel. Full context goes in chat.
"I'll AskUserQuestion and also paste context"Dual channel. One inline card; no picker.
"Recommended + one-line reason is enough"Without Thread, Territory, and consequences per option, the user cannot analyze — only accept a default.
"Keep Why to one line so it stays snappy"Rich enough to decide without a follow-up is the floor. A sentence budget is not a thinness license.
"Context can be a follow-up if they ask"The card is the detail. Follow-up-only context is a thin-card failure.
"We finished the 4 areas on the todo — close"The todo is a living map. Open-set empty is the stop; precommitted N is not.
"Question 3 of 5, then package"No fixed N. Countdown framing is a red flag.
"User asked for button color first"Blast-radius first still holds. Polish Diff after architecture, data, and auth forks.
"We're aligned — skip the decisions table"Shared understanding is the package + yes. Alignment theater without the table is not confirmation.
"Senior said just write requirements"User/senior can override process ownership; they cannot make an unstated decision exist. Emit the package; get the yes.
"I'll assume the safe default and mark done"Assumptions are not decisions. One card; wait.
"Just pick industry best practice — they said sensible"Look up the territory reference first; restate; lock with the user.
"Senior said switch cleanly into clarify-decisions and park the parent"Nesting is the clean switch. Parking the parent and opening a clarify-decisions checklist is dual-channel thrash.
"A short decision checklist under clarify-decisions isn't a competing list"It is a second list. Decision areas live as the parent's in-progress interview item.
"Announce Using clarify-decisions so the user sees the write-handoff"Nested: no mode-switch announcement. Standalone (no parent): you may name clarify-decisions once.
"Parent already loaded neighbors — re-run every card for freshness"Reuse the valid package; rederive only when fingerprints/seeds/scope change
"Standalone interview — skip load-subgraph, Territory is enough"Feature work: load once before the first card

Red flags — stop and rewrite the turn

Red flags — stop and rewrite the turn

  • Calling 
    AskUserQuestion
    or any truncated MCQ tool for a clarify-decisions decision
  • More than one question mark aimed at the user in a single message (except clarifying examples inside option text)
  • Options that are labels only — no consequence lines
  • A card missing Thread, Territory, Why it matters, or Closes
  • "Question k of N", "final round", or closing because a precommitted count finished while high-blast remains
  • Leading with polish-diff while architecture / data / auth branches remain open
  • Closing with "any other questions?" instead of the decisions package
  • Handing back to the parent or starting requirements without an explicit yes on the package
  • Route Tasking the user for a fact present in the repo or the parent's scan digest
  • Abstract taste cards for an unknown-known when a reference or run-spike path exists
  • Nested re-derive every card while the parent package fingerprints still match
  • Standalone feature interview with no retrieval before the first card
  • Calling 
    AskUserQuestion
    or any truncated MCQ tool for a clarify-decisions decision
  • More than one question mark aimed at the user in a single message (except clarifying examples inside option text)
  • Options that are labels only — no consequence lines
  • A card missing Thread, Territory, Why it matters, or Closes
  • "Question k of N", "final round", or closing because a precommitted count finished while high-blast remains
  • Leading with polish-diff while architecture / data / auth branches remain open
  • Closing with "any other questions?" instead of the decisions package
  • Handing back to the parent or starting requirements without an explicit yes on the package
  • Route Tasking the user for a fact present in the repo or the parent's scan digest
  • Abstract taste cards for an unknown-known when a reference or run-spike path exists
  • Nested re-derive every card while the parent package fingerprints still match
  • Standalone feature interview with no retrieval before the first card