Design Patterns: Symptom → Pattern
The Gang-of-Four patterns are answers to recurring design problems. Reaching
for a pattern by name invites over-engineering; reaching for one by symptom
keeps the design honest. This skill maps the pressure you're feeling — "this
switch grows every time we add a type", "callers keep re-implementing the same
multi-step setup" — to the pattern that relieves it, and just as importantly
says when no pattern is the right call.
For the per-pattern intent, structure, and a minimal recipe, see REFERENCE.md. This body is the selector.
When to Use This Skill
| Use this skill when... | Use something else instead when... |
|---|---|
| A structure resists a change that should be easy (new type, new behaviour) | Removing duplication with no design pressure → code-quality-plugin:dry-consolidation |
| Choosing between two patterns, or unsure a pattern is warranted | Mechanical refactor toward pure functions → code-quality-plugin:code-refactor |
| Naming the pattern already latent in a pile of conditionals | Complexity metrics on a function → code-quality-plugin:code-complexity |
| Teaching/justifying a pattern choice in review | UI/component composition patterns → component-patterns-plugin |
Selector
Match the symptom, not the noun. Confirm the pressure is real before adopting any pattern — each adds an indirection that a one-off conditional does not.
| Symptom you feel | Candidate pattern | Confirm before adopting |
|---|---|---|
A switch/if on a type code grows with every new variant | Strategy / State | ≥2 variants exist and more are expected |
| Object behaviour changes with an internal mode, transitions are tangled | State | The modes have distinct transitions, not just a flag |
| Many objects must react when one changes | Observer | Reactors are genuinely decoupled from the source |
| Callers hard-code which concrete class to instantiate | Factory Method / Abstract Factory | The choice varies at runtime or by config |
| An interface you need ≠ the interface you have (3rd-party/legacy) | Adapter | You can't change the callee directly |
| Behaviour to add/remove per-instance at runtime, combinatorially | Decorator | Subclass explosion is the alternative |
| Constructing one object needs many ordered steps/options | Builder | The telescoping constructor actually hurts |
| A subsystem is too many moving parts for callers | Facade | Callers only need a common subset |
| An algorithm's skeleton is fixed but steps vary | Template Method / Strategy | The skeleton truly is stable |
| Traversal logic leaks across a collection's clients | Iterator | The collection's shape is non-trivial |
| Operations sprawl across a type hierarchy you can't grow | Visitor | The hierarchy is stable but operations churn |
When two fit: Strategy (composition) over Template Method (inheritance) unless the skeleton must be enforced; State over Strategy when the variants drive transitions between each other.
Parameters
Parse $ARGUMENTS:
- Symptom or target (optional) — a free-text symptom/requirement, or a
file/diff to assess for latent patterns. If absent, default to the current
change (
git diff HEAD) and scan it for the symptoms above.
Execution
Execute this pattern selection:
Step 1: Name the pressure
State, in one line, the change that is currently hard — the new type, the new
reactor, the runtime choice. If the target is code, locate the growing
conditional / hard-coded new / leaking interface that signals it. If nothing
resists change, stop and say so: no pattern is warranted.
Step 2: Match symptom → candidate
Use the selector to pick 1-2 candidates. Read their intent in REFERENCE.md to confirm fit. State explicitly what each pattern costs (an extra interface, an indirection, a new type to maintain).
Step 3: Justify or reject
Apply the "confirm before adopting" gate for the chosen pattern. If the variants don't yet exist (YAGNI) or only one ever will, recommend the simpler conditional and reject the pattern with the reason. Patterns earn their cost only against real, recurring variation.
Step 4: Recommend
Emit: the symptom, the chosen pattern (or "no pattern — keep the conditional"), the cost it adds, and the minimal shape from REFERENCE.md mapped onto the target's own types. Keep it a recommendation; don't rewrite the code wholesale.
Anti-patterns
| Mistake | Correct approach |
|---|---|
| Picking a pattern by its cool name | Start from the symptom; let it select the pattern |
| Adding Strategy for a single, stable behaviour | One implementation = no pattern; YAGNI |
| Abstract Factory where a function would do | Use the lightest construct that removes the pressure |
| Stacking patterns "to be enterprise-ready" | Each indirection is cost; adopt one only against real variation |
Quick Reference
| Pressure | Pattern | Cost |
|---|---|---|
Type-code switch grows | Strategy / State | One interface + N classes |
| Mode-driven behaviour & transitions | State | State classes + transition wiring |
| Fan-out reactions | Observer | Subscription lifecycle |
Hard-coded new | Factory | An extra creation seam |
| Wrong interface, can't change callee | Adapter | A thin translation layer |
| Per-instance runtime behaviour | Decorator | Wrapper chain |
Related
- REFERENCE.md — per-pattern intent, structure, minimal recipe
code-quality-plugin:dry-consolidation— when the pressure is duplication, not variation, dedupe instead of patterningcode-quality-plugin:code-refactor— the mechanical moves that introduce the chosen patternsoftware-design-plugin:design-deep-modules— a pattern should deepen the interface, not add a shallow indirection layer