决策升级评分卡

Purpose

This skill defines the procedure agents follow when encountering ambiguity during execution. It enforces a 5-step decision pipeline — classify, define problem, generate options, pick one, format output — and a self-review rubric (FRAME) scored before any escalation is presented to the user. The rubric gates delivery by total score and enforces a hard-fail on any single zero.

Procedure

1. Classify the ambiguity. When you encounter missing or unclear information, determine which of these 5 types applies:

   • Approach choice — multiple valid approaches exist and the user did not specify one, but all lead to the same outcome. Example: whether to use a helper function or inline the logic.
   • Minor ambiguity — the intent is clear but a detail is vague and a reasonable default exists. Example: "add logging" without specifying log level — default to info.
   • Ambiguous requirement — the user's intent could mean two or more meaningfully different things. Example: "make it faster" could mean optimize the algorithm, add caching, or reduce payload size.
   • Missing information — a required fact is absent and cannot be inferred. Example: "deploy to the server" with no server specified and no convention to fall back on.
   • Risk confirmation — the action is destructive, expensive, or irreversible and the user has not explicitly authorized it. Example: dropping a database table, force-pushing to main.

2. Apply the branch.

   • Approach choice and minor ambiguity → proceed with a documented default. Record your choice in your handoff. Do not block.
   • Ambiguous requirement → escalate using the 1-3-1 method (step 3) and output template (step 4).
   • Missing information or risk confirmation → stop and ask directly. For non-interactive sessions, return a handoff explaining the gap.
   • The dividing line: if the ambiguity changes what you build, escalate. If it only changes how you build the same thing, proceed.

3. 1-3-1 analysis. When escalating, build your escalation in this order:

   • 1 problem — Write one sentence stating the single problem.
   • 3 options — Generate three meaningfully different approaches. Each with a one-sentence trade-off.
   • 1 recommendation — Pick the strongest option. Write the reason.

4. Apply the output template. Format the escalation exactly as follows:

   The [problem statement in one sentence].
   
   I'm [what I'm working on] and [where I am in the process].
   
   The choice is [decision in plain language]. No jargon.
   
   RECOMMENDATION: [A|B|C] because [reason].
   
   A) [option] — [one-sentence trade-off]
   B) [option] — [one-sentence trade-off]
   C) [option] — [one-sentence trade-off]
   

5. Score each criterion. After preparing the escalation, read the FRAME rubric below and assign a score of 0, 1, or 2 to each letter. Show the scoring breakdown to yourself (internal reasoning, not to the user).

6. Apply the hard-fail rule. If any letter scores 0, do not deliver — go to step 7 immediately.

7. Determine action by total score:

   • 9 – 10 — DELIVER — Escalation meets all criteria. Deliver to user.
   • 7 – 8 — FIX the scored < 2 criteria. a. Identify which letters scored below 2. b. Fix those gaps automatically (do NOT consult the user). c. Re-score, then deliver if 9-10. d. If still below 9, retry once more. e. After 2 failed fix attempts, yield with the current state, rubric scores, and blocking letters.
   • 0 – 6 — RESTART — The escalation is fundamentally broken. Discard and rebuild with corrected understanding, or yield with an explanation of what went wrong.

FRAME Rubric

F — FLAG THE BRANCH

Did I classify the ambiguity type and apply the correct branch (proceed vs escalate)?

• 0 — No classification performed. Jumped straight to acting or asking without determining the ambiguity type.
• 1 — Classified the type but picked the wrong branch: proceeded when should escalate (ambiguous requirement, missing info, risk), or escalated when should proceed (approach choice, minor ambiguity).
• 2 — Classified correctly. Applied the right branch: proceed for approach choice and minor ambiguity; escalate for ambiguous requirement; stop-and-ask for missing info and risk confirmation.

R — ROOT THE PROBLEM

Did I state exactly one problem before generating solutions?

• 0 — Skipped problem definition. Jumped from noticing ambiguity straight to options or open-ended questions.
• 1 — Stated a problem but it is vague ("something is unclear") or bundles multiple problems into one statement.
• 2 — One sentence. One problem. Specific enough that someone could evaluate whether a solution addresses it.

A — ARRAY THE OPTIONS

Did I generate 3 meaningfully different options?

• 0 — Options are all variants of the same approach. Or presented fewer than 3. Or presented more than 5.
• 1 — Three options exist but two are too similar to be meaningful alternatives. No trade-off stated per option.
• 2 — Three options. Each takes a distinctly different approach. Each has a one-sentence trade-off.

M — MAKE THE PICK

Did I choose one option and state the reason?

• 0 — Listed options without picking one. Or picked one but gave no reason. Or used "it depends" or hedging language.
• 1 — Picked one but the reason is generic ("it's better") or does not connect to the root problem.
• 2 — Exactly one recommendation stated as RECOMMENDATION: [option] because [reason]. The reason traces back to the root problem defined in R.

E — ENFORCE THE FORMAT

Did I follow the 5-part structured question format exactly?

• 0 — Wrong format entirely. Asked an open-ended question. Missing RECOMMENDATION line.
• 1 — Most sections present but one is wrong or out of order (e.g., context statement missing, or Options numbered instead of lettered A/B/C).
• 2 — All sections in exact order: Problem → Context → Choice → RECOMMENDATION → A) B) C). Lettered options (not numbered). One sentence per option trade-off.

Guardrails

• Never ask open-ended questions when concrete options exist
• Never present options without a recommendation
• Never treat every gap as a blocker (approach choice + minor ambiguity proceed inline)
• Simple yes/no and single-fact clarifications are exempt from this format
• The rubric is fixed — do not add or remove criteria
• Never deliver with any letter scoring 0