xoai Quick Elicit
Guide the human to articulate a high-quality specification through 4 focused rounds of questions. Takes ~3 minutes, not 20.
Core Principle: The quality of output is determined by the quality of input. The most expensive failure is building the wrong thing correctly. Round 0 challenges the framing BEFORE the other rounds capture requirements.
When to Use
BUILD mode, before planning. When the human describes something they want to build, add, or change — and no specification exists yet.
Do NOT use when:
- A spec already exists in
.sage/work/(skip to planning) - The human has provided a detailed written spec (accept it, don't re-elicit)
- FIX mode (the bug report IS the spec)
- The human says "just do it" (respect their autonomy, note the risk)
Process
Pre-Elicit: Infer What You Can
Before asking anything, scan the relevant codebase area to understand:
- What tech stack is in use?
- What patterns exist in the codebase?
- What does the area being changed look like?
- What conventions are established?
If a codebase-scan has been run previously, use its output. Otherwise,
scan the relevant files directly (read project config, entry points,
and the area being changed).
Also scan .sage/docs/ for existing research artifacts (jtbd-, ux-audit-,
opportunity-, user-interview-). If research exists, note the known pain
for use in Round 0.
This determines WHAT TO ASK. Don't ask about things you can already see.
Round 0: Framing (30 seconds)
PURPOSE: Challenge whether the request is framed around the right problem BEFORE committing to a specification direction.
Ask ONE question:
"What goes wrong today without this?"
This is not "what does this feature do?" It's "what specific frustration, failure, or friction makes this worth building?" The answer should be a concrete struggling moment, not an abstract category.
- Good answer: "Sellers upload dark, cluttered photos and their listings don't sell. They blame the platform."
- Thin answer: "Users need to upload photos." (restates the request, not pain)
- Acceptable: "There's no way to add images right now." (genuine gap)
If research context exists: Reference the known pain instead of asking open-ended: "Research identified [pain]. Is that still the driver, or has something changed?"
If the answer restates the feature instead of the pain, probe once: "Got it — but what's the actual problem that causes? What goes wrong for the user?"
Do NOT probe more than once. If the pain remains unclear after one follow-up, note it and proceed.
The challenge: Based on the pain + codebase scan, surface one or two implicit premises the user is making:
Before we lock this in — I notice two assumptions worth checking:
1. [Premise]: [one sentence stating the assumption]
→ [one sentence describing why it might be wrong]
2. [Premise]: [one sentence stating the assumption]
→ [one sentence describing why it might be wrong]
Does the original framing still feel right, or should we adjust?
What counts as an implicit premise:
- "This needs a new screen" (maybe the existing screen can be extended)
- "This is a frontend problem" (maybe it's a data problem)
- "This is one feature" (maybe it's three, and only one matters now)
- "Users want this" (maybe they want something adjacent)
- "This needs to be built" (maybe it can be configured)
What does NOT count (don't waste time on these):
- Technical implementation choices (that's for planning)
- Architecture decisions (that's for spec or architect)
- Obvious facts about the codebase (already visible from scan)
User responds in one of three ways:
"Framing is right." → Proceed to Round 1 with the original request.
Note: Framing: Original request confirmed. Pain: [stated pain].
"Let's adjust." → User provides adjusted framing. Round 1 uses the
adjusted framing, not the original request.
Note: Framing: Adjusted from [original] to [adjusted]. Pain: [pain].
"I don't know / let me think." → Note uncertainty. Proceed with
original framing but flag it:
Framing: Original request (low confidence). Pain: unclear. Consider /research or /jtbd before building.
At Round 0 completion: Prepend framing decision to .sage/decisions.md
BEFORE proceeding to Round 1:
### YYYY-MM-DD — Framing: [initiative]
[Chose framing]. Pain: [pain]. Challenged: [premise names].
(quick-elicit Round 0)
Round 0 bypass conditions
Round 0 is skipped in exactly these cases:
- Lightweight scope — Task classified as Lightweight. One component, no design decisions. Framing adds no value to "fix the typo."
- Spec already exists — Framing was already considered.
- User says "just do it" — Agent notes: "Skipping framing challenge. Spec inherits original request without challenge."
- FIX mode — The bug report IS the spec. Pain is self-evident.
No other bypass conditions. Standard and Comprehensive ALWAYS get Round 0.
Anti-rationalization contract for Round 0 skip: Do NOT skip Round 0 because:
- "The request is clear enough" — clarity of request ≠ correctness of framing
- "This is a small change" — Standard classification means framing matters
- "The user already explained the pain" — restate and verify, don't assume
- "We discussed this in a previous session" — disk is truth, not memory
Round 1: Intent (30 seconds)
Ask TWO questions:
"What should this do when it's working perfectly?" Get the happy path. What does success look like from the user's perspective?
"Who uses this, and when?" Get the actor and trigger. This prevents building features nobody uses.
Wait for answers. Then draft a 3-5 sentence intent statement and show it:
INTENT: [feature name]
Framing: [original | adjusted | original (low confidence)]
Pain: [the struggling moment, from Round 0]
When [actor] does [trigger], the system [behavior].
This enables [value/outcome].
Ask: "Does this capture what you mean? Anything to adjust?"
Note: If Round 0 adjusted the framing, Round 1's questions are asked about the ADJUSTED framing, not the original request.
Round 2: Boundaries (60 seconds)
Ask up to THREE questions, adapting based on what you inferred from the codebase:
"What should this explicitly NOT do?" Boundaries prevent scope creep and gold-plating.
"What existing code does this touch? Any concerns?" Skip if the codebase scan already identified the impact area clearly.
"Any security, performance, or compatibility constraints?" Skip if the change is purely internal with no user-facing impact.
Draft boundary conditions and show them:
BOUNDARIES:
- This feature WILL: [list]
- This feature WILL NOT: [list]
- Constraints: [any non-functional requirements]
- Affected areas: [files/modules identified]
Ask: "Anything missing or wrong here?"
Round 3: Verification (30 seconds)
Ask ONE question:
"How will we know this works? What would you test manually?"
The answer becomes acceptance criteria. Draft them and show:
ACCEPTANCE CRITERIA:
1. [testable criterion]
2. [testable criterion]
3. [testable criterion]
Each criterion MUST be testable — not "works well" but "returns 200 OK with user.id in response body." If the human gives vague criteria, make them specific.
Output: Feature Spec
Combine all four rounds into a spec document saved to
.sage/work/<YYYYMMDD>-<slug>/spec.md using the minimal spec template.
The spec includes a ## Framing section at the top (from Round 0),
followed by Intent, Boundaries, and Acceptance Criteria.
Show the complete spec. Ask: "Ready to plan implementation, or anything to adjust?"
Compression Behavior
If the user is visibly impatient, compress to 1 round. The compressed version merges pain + intent into one message:
"Before we spec this — what's the actual pain this solves? And then: what should it do when working perfectly?"
One message, two questions. Infer boundaries and criteria from context. Show the complete spec with a Framing section for validation.
Even compressed, the framing question survives. Round 0 can be compressed but NEVER eliminated. A compressed Round 0 still produces a Framing section in the spec.
Rules
MUST (violation = bad spec or lost trust):
- MUST NOT ask more than 8 questions total across all rounds.
- MUST NOT ask about things the codebase scan already revealed.
- MUST NOT ask technical implementation questions (that's for planning, not spec).
- MUST show the drafted output after each round for validation.
- MUST make acceptance criteria testable and specific.
- MUST execute Round 0 for Standard+ tasks (see bypass conditions).
- MUST prepend framing decision to decisions.md before Round 1.
- MUST include a non-empty Framing section in spec output.
SHOULD (violation = suboptimal experience):
- SHOULD respect one-word answers — work with what you have, don't interrogate.
- SHOULD make a reasonable choice when the human says "you decide," and state it explicitly so they can override if it's wrong.
- SHOULD compress to 1 round if the human is visibly impatient.
- SHOULD reference existing research context in Round 0 if available.
MAY (context-dependent):
- MAY skip elicitation entirely if the human provides a detailed written spec.
- MAY recommend ARCHITECT mode if requirements are genuinely complex.
Failure Modes
- Human is impatient: Compress to 1 round — merge pain + intent question, infer boundaries and criteria from context, show complete spec with Framing.
- Human has already thought deeply: Skip elicitation, help them write their thoughts into the spec template structure, then validate.
- Requirements are genuinely complex: Recommend switching to ARCHITECT mode for full elicitation with deep-elicit skill.
- Conflicting requirements surface: Don't resolve them yourself. Present the conflict clearly and ask the human to decide.
- Low-confidence framing: Include "Consider /research or /jtbd before building" in the Framing section. Surface at the checkpoint: "The pain behind this isn't fully clear. Want to proceed, or run research first?" This is advisory — do NOT block the user from proceeding.