next-determine-what-to-do-next

name	next - Determine What to Do Next
description	Sub-orchestrator for next-step requests. Classifies intent, checks context, selects the single highest-value immediate action, and routes to execution or the right analysis skill.
output	{"format":"prose"}

Next - Determine What to Do Next

Input: $ARGUMENTS

---

Core Principles

1. One step, not a roadmap. This skill returns ONE next action. If the user needs a sequence, route to /fonss. The temptation to give five steps "for context" dilutes the recommendation and overwhelms the user.

2. Classify before routing. "What should I do next?" is almost never actually a next-step request. It's usually a decision, a diagnostic, a goal question, or method-seeking in disguise. Classify first, then route.

3. Context determines everything. The same words from different states need different skills. "What next?" when stuck = /diagnose. "What next?" when exploring = /search. "What next?" when everything is clear = /action.

4. Backtracking is a valid next step. Sometimes the right next action is to undo, reconsider, or restart. The skill must be able to recommend going backward, not just forward.

5. Missing context is a finding, not a blocker. If you can't determine the next step because the goal is unclear, THAT is the finding: "Your next step is to clarify the goal."

---

Phase 1: Intent Classification

Is this actually a next-step request? Test against common disguises:

[N1] SURFACE_REQUEST: [what the user literally said]
[N2] ACTUAL_INTENT: [next-step | decision | method | diagnostic | exploration | goal-clarification | emotion]

Classification Table

Pattern	Actual Intent	Route
"Should I X or Y?"	Decision	→ INVOKE: /decide $ARGUMENTS
"How do I X?"	Method-seeking	→ INVOKE: /how $ARGUMENTS
"Why is X happening?" / "What went wrong?"	Diagnostic	→ INVOKE: /diagnose $ARGUMENTS
"What are my options?"	Exploration	→ INVOKE: /search $ARGUMENTS
"I want X" / "I need X"	Goal clarification	→ INVOKE: /want $ARGUMENTS
"I'm stuck" / "I'm frustrated"	Emotion + diagnostic	→ INVOKE: /emotion $ARGUMENTS
"What should I do next?" / "Now what?"	Genuine next-step	→ Continue to Phase 2
"Is this good?" / "Did I do it right?"	Evaluation	→ INVOKE: /evaluate $ARGUMENTS
"Make this better"	Iteration	→ INVOKE: /iterate $ARGUMENTS

If routing away: output the reclassification and invoke immediately. Do not continue.

---

Phase 2: Context Assessment

For genuine next-step requests, assess what's known:

[N3] GOAL: [stated or inferred goal]
[N4] DONE_SO_FAR: [what's been completed]
[N5] CURRENT_POSITION: [where in the process the user is]
[N6] BLOCKERS: [what's preventing progress, if anything]
[N7] AVAILABLE_INFORMATION: [what the user has to work with]

Context Completeness

Level	Has	Missing	Route
Complete	Goal + state + no blockers	—	→ Phase 3 (select action)
Blocked	Goal + state + identified blocker	Blocker resolution	→ /diagnose the blocker
Partial	Goal but unclear state or progress	Current position	→ Ask user to state where they are
Thin	Vague goal, unclear state	Goal + state	→ /want to clarify
Recovery	Prior attempt failed	Why it failed	→ /sbfow to diagnose failure

[N8] CONTEXT_LEVEL: [complete | blocked | partial | thin | recovery]
[N9] CONTEXT_ACTION: [proceed to Phase 3 | route to skill | ask user]

---

Phase 3: Next-Step Selection

Given complete context, determine the single highest-value next action.

[N10] CANDIDATE_ACTIONS:
  [N11] ACTION: [specific action] — TYPE: [concrete task | analysis | decision | validation] — VALUE: [high | medium | low]
  [N12] ACTION: [specific action] — TYPE: [type] — VALUE: [level]
  [N13] ACTION: [specific action] — TYPE: [type] — VALUE: [level]

Selection Criteria

Criterion	Question
Dependency	Does anything else need to happen before this can work?
Unblocking power	Does this action unblock other actions?
Information gain	Does this resolve uncertainty that's holding things up?
Reversibility	If wrong, how hard is this to undo?
Timeliness	Is there a reason to do this NOW vs later?

Routing Table

Situation	Route
Concrete task is obvious and ready	→ /action $ARGUMENTS
Need method before acting	→ /how $ARGUMENTS
Need evaluation before acting	→ /evaluate $ARGUMENTS
Need to choose between options	→ /decide $ARGUMENTS
Need more options before choosing	→ /search $ARGUMENTS
Need to understand root cause	→ /diagnose $ARGUMENTS
Need to undo or backtrack	State what to undo and why
Multiple steps needed	→ /fonss $ARGUMENTS

---

Phase 4: Output

NEXT STEP:
  ACTION: [one specific thing to do]
  TYPE: [concrete task | analysis | decision | validation | backtrack]
  WHY_THIS_FIRST: [why this action is the highest-value next step]

  → INVOKE: /skill-id $ARGUMENTS

AFTER THIS STEP:
  CHECK: [what result to verify immediately after]
  IF_GOOD: [likely next step after this one]
  IF_BAD: [what to do if this step fails or produces weak output]

---

Failure Modes

Failure	Signal	Fix
Misclassification	Treated a decision as a next-step request	Phase 1 classification table must be checked first
Roadmap output	Returned 5 steps instead of 1	Enforce single-step rule; route to /fonss for sequences
Forward-only bias	Never recommends backtracking	Explicitly check: is the right move to go backward?
Context assumption	Assumed goal/state without checking	Phase 2 must extract context from the prompt, not infer
Generic routing	Always routes to /action	Consider all routing options; /action is only for clear, concrete tasks
Missing recovery path	No IF_BAD in output	Always specify what to do if the step fails

---

Depth Scaling

Depth	Candidates Generated	Classification Rigor	After-Step Analysis
1x	2	Pattern match only	IF_GOOD only
2x	4	Full classification table	IF_GOOD + IF_BAD
4x	6	Classification + confidence	Full branch analysis
8x	8	Classification + evidence	Multi-path scenarios

Default: 2x. These are floors.

---

Pre-Completion Checklist

• Intent classified (not assumed to be a next-step request)
• If reclassified: routed to correct skill immediately
• Context assessed with completeness level
• If context incomplete: addressed (ask user or route to clarifying skill)
• Single next step selected (not a roadmap)
• WHY_THIS_FIRST explicitly states why this beats alternatives
• Recovery path specified (IF_BAD)
• Backtracking considered as a valid option

---

Integration

• Use from: user directly, /meta, /handle
• Differs from /fonss: fonss builds a multi-step chain; next picks ONE step
• Differs from /wsib: wsib picks the best skill for a prompt; next picks the best ACTION for a state
• Routes to: whatever skill the next step requires
• Escalates to: /fonss when multi-step chain is clearly needed