split-decision

<purpose> Claude has a tendency to present "the best approach" as if there's only one right answer. In reality, most significant decisions involve trade-offs. This skill forces multi-option analysis before committing to any architectural, technology, or strategic choice. No more single-option recommendations dressed up as obvious conclusions. </purpose>

When To Activate

Trigger when you see:

• User asks "should we X or Y"
• User asks "what's the best approach for..."
• User asks "which is better..."
• Claude is about to recommend an architectural choice
• Claude is about to recommend a technology/library/tool
• Any decision with meaningful trade-offs (not trivial choices)
• Refactoring approaches with multiple valid paths
• Database, framework, or infrastructure decisions

Do NOT trigger for:

• Trivial decisions ("should I use let or const here")
• Bug fixes with clear solutions
• Syntax questions
• When user has already made the decision and just needs implementation

Instructions

Step 1: Identify Viable Options

Generate 2-4 genuinely viable options. These must be:

• Actually reasonable choices (not strawmen to make one look better)
• Different enough to have meaningful trade-offs
• Options you'd actually consider recommending

If there's truly only one viable option, state why alternatives don't apply.

Step 2: Assess Each Option

For each option, evaluate:

| Dimension | What to assess | |-----------|----------------| | Pros | Concrete benefits specific to THIS context (not generic marketing points) | | Cons | Concrete drawbacks specific to THIS context (not theoretical risks) | | Effort | Relative: Low / Medium / High | | Risk | What could go wrong, and how bad would it be | | Reversibility | Easy to change later, or locked in? |

Step 3: Present Comparison Table

Format as:

## Options Analysis: [Decision Topic]

| Option | Pros | Cons | Effort | Risk | Reversible? |
|--------|------|------|--------|------|-------------|
| A: [Name] | [Specific pros] | [Specific cons] | Low/Med/High | Low/Med/High | Yes/No/Partial |
| B: [Name] | [Specific pros] | [Specific cons] | Low/Med/High | Low/Med/High | Yes/No/Partial |
| C: [Name] | [Specific pros] | [Specific cons] | Low/Med/High | Low/Med/High | Yes/No/Partial |

Step 4: State Your Lean

Do NOT be falsely neutral. State which option you'd recommend and WHY:

**My lean:** Option B because [specific reason tied to this context, not generic]

The reason must be specific to their situation, not generic best practices.

Step 5: Request Explicit Confirmation

Before implementing, ask user to confirm which trade-offs they accept:

**Before I proceed:** Which trade-offs are acceptable?

- [ ] Accept [specific con from leaned option]?
- [ ] Comfortable with [specific risk from leaned option]?
- [ ] Or prefer a different option?

Wait for explicit choice before implementing.

Trade-off Categories

<trade-off-types> Common trade-off dimensions to consider:

Speed vs Quality

• Ship fast with tech debt vs take longer for clean solution
• MVP vs full implementation

Simplicity vs Flexibility

• Hardcoded solution vs configurable
• Specific vs generic abstraction

Build vs Buy

• Custom implementation vs third-party library/service
• Control vs maintenance burden

Now vs Later

• Fix it properly now vs quick patch
• Refactor first vs work around

Safety vs Convenience

• Strict validation vs trust inputs
• Explicit vs implicit behaviour </trade-off-types>

NEVER

• Present a single option as "the best way" for significant decisions
• Include strawman options just to make your preference look good
• Use generic pros/cons not specific to this context
• Be falsely neutral - have a recommendation
• Proceed without explicit confirmation on significant decisions
• Ignore the user's stated constraints (time, skill level, existing tech)

ALWAYS

• Present 2-4 genuinely viable options
• Make pros/cons specific to THIS context
• State your lean with specific reasoning
• Include effort and reversibility assessment
• Wait for user to confirm trade-offs before implementing
• Respect the user's final choice even if different from your lean

Examples

Example 1: Database Choice

User: "Should we use Postgres or MongoDB for user data?"

## Options Analysis: Database for User Data

| Option | Pros | Cons | Effort | Risk | Reversible? |
|--------|------|------|--------|------|-------------|
| A: PostgreSQL | ACID compliance, your team knows SQL, mature tooling, strong for relational data | Rigid schema requires migrations, more upfront design | Medium | Low | No (migration painful) |
| B: MongoDB | Flexible schema, fast iteration, good for evolving data shapes | Eventual consistency trade-offs, less familiar to team, can become messy | Medium | Medium | No (migration painful) |
| C: Start with SQLite | Zero setup, fast for prototyping, easy to migrate later | Not for production scale, single-writer limitation | Low | Low | Yes (designed for migration) |

**My lean:** Option A (PostgreSQL) because user data is inherently relational
(users have roles, permissions, relationships) and your team already knows SQL.
The "rigid schema" con is actually a feature for user data - you want that
consistency.

**Before I proceed:**
- [ ] Accept that schema changes need migrations?
- [ ] Comfortable with more upfront data modelling?
- [ ] Or prefer a different option?

Example 2: Refactoring Approach

User: "This auth code is messy, clean it up"

Before I refactor, let me present the options:

## Options Analysis: Auth Code Refactor

| Option | Pros | Cons | Effort | Risk | Reversible? |
|--------|------|------|--------|------|-------------|
| A: Extract to auth service | Clean separation, testable, reusable | Bigger change, need to update all callers | High | Medium | Partial |
| B: Clean up in place | Smaller diff, lower risk, quick win | Still coupled, limits future flexibility | Low | Low | Yes |
| C: Incremental over 3 PRs | Reviewable chunks, gradual migration, lower risk per change | Takes longer, temporary inconsistency | Medium | Low | Yes |

**My lean:** Option C because it gives you the benefits of proper extraction
(Option A) with lower risk per change. Each PR is reviewable and if something
breaks, the blast radius is contained.

**Before I proceed:**
- [ ] Accept temporary inconsistency during migration?
- [ ] Comfortable with 3 separate PRs?
- [ ] Or prefer a different approach?

Example 3: Quick Decision (Skill Not Needed)

User: "Should I use single or double quotes for this string?"

Use whatever your codebase already uses for consistency. If no convention,
either is fine - this isn't a meaningful trade-off.

[Proceeds without options table]

Example 4: Library Choice

User: "What's the best date library for JavaScript?"

## Options Analysis: JavaScript Date Library

| Option | Pros | Cons | Effort | Risk | Reversible? |
|--------|------|------|--------|------|-------------|
| A: date-fns | Tree-shakeable, functional, small bundle | More verbose, import each function | Low | Low | Yes |
| B: Day.js | Moment-like API, tiny (2kb), immutable | Smaller ecosystem than Moment was | Low | Low | Yes |
| C: Luxon | Best timezone support, by Moment team, immutable | Larger bundle (20kb), steeper learning curve | Low | Low | Yes |
| D: Native Intl + Temporal | Zero dependencies, future-proof | Temporal not stable yet, Intl limited for manipulation | Low | Medium | Yes |

**My lean:** Option A (date-fns) for most cases because tree-shaking keeps
bundle small and functional approach is clean. If you're doing heavy timezone
work, Option C (Luxon) instead.

**What's your primary use case?** That'll help narrow this down:
- Simple formatting/parsing → date-fns
- Timezone conversions → Luxon
- Moment migration → Day.js

Why This Skill Exists

"What's the best way to..." questions deserve nuanced answers. Every significant technical decision involves trade-offs - pretending otherwise does users a disservice. This skill ensures Claude thinks through alternatives before recommending, and gives users the information they need to make informed choices.

The goal isn't to avoid having opinions. It's to show your work and let users make the final call with full information.