This skill should be used when requirements are unclear, brainstorming ideas, or when "pathfind", "brainstorm", "figure out", "clarify requirements", or "work through" are mentioned.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
This skill should be used when requirements are unclear, brainstorming ideas, or when "pathfind", "brainstorm", "figure out", "clarify requirements", or "work through" are mentioned.
Transition: Mark current completed, add next in_progress
High start (4+): Skip directly to Clarify or Deliver
Early delivery: Skip to Deliver + โณ Caveats
Calibrate first โ user may have already provided context (docs, prior conversation, pointed you at files). If enough context exists, skip to level 3โ4. Don't re-ask what's already clear.
If gaps remain, explore focus areas (pick what's relevant):
Purpose: What problem? Why now?
Constraints: Time, tech, team, dependencies
Success: How will we know it works?
Scope: What's in, what's out?
When multiple approaches exist:
Propose 2โ3 options with trade-offs
Lead with recommendation โ and reasoning
Let user pick, combine, or redirect
Principles:
YAGNI โ cut what's not needed
DRY โ don't duplicate effort or logic
Simplest thing โ prefer boring solutions
Use EnterPlanMode for each question โ enables keyboard navigation of options.
Structure:
Prose above tool: context, reasoning, โ recommendation if clear lean
Inside tool: options only (concise, scannable)
At level 0 โ start with session intent:
Quick pulse check vs deep dive?
Exploring possibilities or solving a specific problem?
What does "done" look like?
Levels 1โ4 โ focus on substance:
2โ4 options per question + "5. Something else"
Inline [โ ] on recommended option + italicized rationale
User replies: number, modifications, or combos
Loop: Answer โ Restate โ Update Confidence โ Next action
After each answer emit:
Confidence: {BAR} {NAME}
Assumptions: { if material }
Unknowns: { what we can clarify; note unknowables when relevant }
Decisions: { what's locked in }
Concerns: { what feels off + why }
Next action by level:
0โ2: Ask clarifying questions
3: Summarize (3 bullets max), fork toward 5
4: Offer choice: refine or proceed
5: Deliver
When answer reveals a concern mid-stream:
Pause before next question
Surface with โณ + brief description
Ask: clarify now, note for later, or proceed with assumption?
Example: "โณ This assumes the API supports batch operations โ clarify now, note for later, or proceed?"
If user proceeds despite significant gap โ escalate to pushback protocol.
Escalate when choice conflicts with goals/constraints/best practices:
โ Alternative: Minor misalignment. Present option + reasoning.
โโ Hazard: High failure risk. Require mitigation or explicit override. Triggers Resolve Conflicts.
Override: Accept "Proceed anyway: {REASON}" โ log in next reflection โ mark Resolve Conflicts complete.
Integrate skeptic agent for complexity sanity checks:
Recommend (offer choice):
Level 5 reached with โณ Caveats > 2
Red flag language in decisions: "might need later", "more flexible", "best practice"
Before finalizing โ you have {N} caveats. Want to run skeptic for a sanity check?
[AskUserQuestion]
1. Yes, quick check [โ ] โ I'll challenge complexity interactively
2. Yes, deep analysis โ launch skeptic agent in background
3. No, proceed โ deliver as-is
Auto-invoke (no choice):
Level 4+ with 3+ unknowns persisting across 2+ question cycles
โโ Hazard escalation triggered during session
When auto-invoking:
[Auto-invoking skeptic โ {REASON}]
Launch with Task tool:
subagent_type: "outfitter:skeptic"
prompt: Include current decisions, unknowns, and caveats
run_in_background: false (wait for findings before delivery)
After skeptic returns:
Present findings to user
If verdict is block โ add Resolve Conflicts stage
If verdict is caution โ offer choice to address or acknowledge
If verdict is proceed โ continue to delivery
Level 5: Produce artifact immediately (doc, plan, code, outline). If none specified, suggest one.
After delivering, ask where to persist (if applicable):
[EnterPlanMode]
1. { discovered path } [โ ] โ { source: `CLAUDE.md` preference | existing directory | convention }
2. Create issue โ { Linear/GitHub/Beads based on project context }
3. ADR โ { if architectural decision }
4. Don't persist โ keep in conversation only
5. Something else โ different location or format
Discovery order for option 1:
CLAUDE.md or project instructions with explicit plan storage preference
Existing .agents/plans/ directory
Existing docs/plans/ directory
Fall back to .agents/plans/ if nothing found
Always suggest filename based on topic. Match existing conventions if present.
Mark Deliver completed after artifact is delivered (persistence is optional follow-up).
Below 5: Append โณ Caveats:
Open questions + context
Assumed decisions + defaults
Known concerns + impact
Deferred items + revisit timing
ALWAYS:
Task stage matching initial confidence at start
EnterPlanMode for each question (keyboard nav)
Prose above tool for context + โ recommendation
One question at a time, wait for response
Restate + update confidence before next move
Update todos at level 4, level 5 thresholds
Apply pushback protocol on conflicts
Check skeptic triggers at level 4+ (unknowns, caveats, red flags)
NEVER:
Proceed from 0โ3 without clarifying questions
Hide uncertainty below level 5
Stack questions or bury decisions in paragraphs
Put recommendation inside plan tool (keep in prose)
