Experimental

gosm

Goal-Oriented State Machine. Context-adaptive dispatcher that routes to the right analysis depth based on urgency, stakes, and expertise.

Usage in Claude Code: /gosm your question here

GOSM - Goal-Oriented State Machine

Input: $ARGUMENTS

Principle

Match analysis depth to the situation. Don’t overthink low-stakes decisions. Don’t underthink high-stakes ones.

Context Assessment

Assess these five factors to select the right variant:

FactorOptions
UrgencyURGENT (minutes) / SOON (hours) / NORMAL (days+)
StakesHIGH (major impact) / MEDIUM / LOW (minor)
ExpertiseEXPERT (user knows domain) / INTERMEDIATE / NOVICE
Action CostCHEAP (reversible) / EXPENSIVE (one-shot)
InformationRICH (lots known) / MODERATE / SPARSE

Variant Selection

ContextVariantWhat It Does
URGENT + anyLiteFind the ONE critical assumption. Test it. Act.
LOW stakes + CHEAPQuickSanity check. If nothing’s obviously wrong, proceed.
EXPERT + HIGH confidenceCheckValidate the user’s plan. Find blind spots.
Post-action / reflectingAfterLearn from what happened. Update model.
MEDIUM stakes + NORMAL timeStandardBalanced analysis: key claims, brief ARAW, action.
HIGH stakes + EXPENSIVE + NOVICEFullComprehensive: invoke /procedure_engine

User can override: “quick”, “full”, “just check this”, “reflect on what happened”.

Lite (< 5 minutes)

For urgent situations. Find the ONE assumption that matters most.

  1. What’s the core claim? (one sentence)
  2. What’s the critical assumption? (the one thing that changes everything if wrong)
  3. Quick ASSUME WRONG: If that assumption is wrong, what’s the best alternative? What’s the risk?
  4. Action: What to do right now. How to verify it worked.
  5. Revisit when: Time permits / new info arrives / action didn’t work.

Quick (< 2 minutes)

For low-stakes, reversible decisions. Don’t overthink.

  1. Sanity check: Any obvious constraints violated? Any clearly better option being ignored? Will I regret not thinking harder?
  2. If all clear → Proceed. Note the undo condition.
  3. If anything flagged → Decide if it’s worth deeper analysis.

Check (< 10 minutes)

For experts who have a plan. Don’t explore — validate.

  1. Blind spot scan: Obvious failure modes? Missing stakeholders? Unrealistic assumptions? Resource gaps? Timeline issues?
  2. Strongest argument against: What’s the best case for NOT doing this?
  3. Verdict: PROCEED / PROCEED WITH CAUTION (issues noted) / PAUSE (significant issue) / RETHINK (fundamental problem)

After (< 15 minutes)

For reflecting on actions already taken.

  1. Expected vs actual: What did you think would happen? What did happen? What’s the gap?
  2. What this reveals: For each learning — what to do differently.
  3. Updated model: What did you believe before? What do you believe now?
  4. Next action: Based on what was learned.

Standard (15-30 minutes)

Balanced analysis for medium-stakes situations.

  1. Classify: GOAL / PROBLEM / QUESTION / DECISION / SITUATION
  2. Key claims: Surface + hidden assumptions (2-4 claims)
  3. ARAW the most important claim: Assume Right → what follows? Assume Wrong → what alternatives?
  4. Goal journey: Current state → desired state → what serves this?
  5. Contrarian view: What’s the strongest challenge?
  6. Actionable filter: What can the user actually do?
  7. Crux question: The one question that matters most.
  8. Recommended action: Specific, with verification.

Full (30-60+ minutes)

For high-stakes, expensive, novel situations. Invokes the full procedure engine.

→ /procedure_engine $ARGUMENTS

Pre-Completion Check

  • Context assessed (urgency, stakes, expertise, cost, information)
  • Variant matches context (or user override applied)
  • Output depth matches variant (Lite = focused, Full = comprehensive)
  • Specific action recommended