Claude Feature Decision Matrix As of 2026-05

For each use-case pattern, which Claude platform features are essential, conditional, or rarely useful. Hover any cell for rationale.
✓ Essential — pattern doesn't work well without it
~ Conditional — useful in specific sub-cases
· Rarely useful — feature mismatch with pattern
Pattern ↓ Feature → Prompt caching Extended thinking Tool use Computer use 2.0 Files API Skills MCP Plugins Agent SDK Memory tool Batch API Citations
RAG support copilotUser asks → retrieve → ground → answer0–3 mo ✓ · ~ · ✓ ~ ~ ~ · ~ · ✓
Agentic workflowPlan → tool calls → observe → replan3–12 mo ✓ ✓ ✓ ~ ~ ✓ ✓ ✓ ✓ ✓ ~ ~
Batch enrichmentDocument/data classification, extraction0–3 mo ✓ · ~ · ✓ ~ · · · · ✓ ~
Computer-use automationBrowser/desktop control via screenshots3–12 mo ~ ✓ ✓ ✓ · ✓ ~ ~ ✓ ~ · ·
Domain expert assistantVertical assistant w/ deep procedures0–3 mo ✓ ✓ ~ · ✓ ✓ ✓ ✓ ~ ~ ~ ✓
Code automation (Claude Code)CLI + IDE for engineering teams0–3 mo ✓ ✓ ✓ ~ · ✓ ✓ ✓ ~ ~ ~ ·
Embedded copilotChat surface inside an internal app3–12 mo ✓ ~ ✓ · ~ ~ ✓ ~ ~ ~ · ~
Eval / regression testingPrompt + model regression suite0–3 mo ✓ ~ ~ · ~ ~ · · ~ · ✓ ·
How to read this matrix
  1. Pick your pattern — find the row that best describes the workload. If two rows fit, your design is probably blending patterns and that's fine — combine the ✓ columns.
  2. The green cells are your build list. Anything marked ✓ is core to making the pattern work. Skipping a green cell usually shows up as a quality, cost, or maintenance problem within 90 days.
  3. Yellow cells are decisions, not defaults. Adopt only when the sub-condition in the tooltip applies. Adopting all yellows by reflex bloats the architecture.
  4. Gray cells are signal that you're using the wrong tool. If you find yourself reaching for one, re-check the pattern row — you may be solving a different problem than you think.
  5. Time-to-disruption tag on each pattern row classifies how soon shipping that pattern moves the needle: 0–3 MO = mature pattern, ship now; 3–12 MO = still maturing or integration-heavy; 12+ MO = research-stage / not represented in this matrix yet (autonomous-agent-at-scale, multimodal-output-as-product). Use the lens to sequence pilots: 0–3 patterns first to build evidence, 3–12 next once eval discipline lands.

Canonical stacks (one-liner per pattern)

RAG support copilot

Sonnet 4.6 · cached system + retrieved context · Files API for source docs · citations on.
Sonnet 4.6 + caching + Files + citations

Agentic workflow

Sonnet 4.6 baseline (Opus 4.7 escalation) · Agent SDK loop · MCP for system access · Skills for procedures · memory tool · plugins for distribution.
Agent SDK + MCP + Skills + memory + plugins

Batch enrichment

Haiku 4.5 · Batch API · cached schema · Files API for inputs.
Haiku 4.5 + batch + caching + Files

Computer-use automation

Sonnet 4.6 + thinking · computer use 2.0 · Skills for the workflow playbook · Agent SDK for the loop.
computer use 2.0 + thinking + Skills + Agent SDK

Domain expert assistant

Sonnet 4.6 + thinking · Skills for procedures · Files for regs/policies · MCP for systems · citations.
Skills + Files + MCP + citations + plugins

Code automation

Claude Code (Opus 4.7 for hard refactors, Sonnet 4.6 default) · team plugin (commands + skills + hooks + MCP).
Claude Code + plugins + Skills + MCP

Embedded copilot

Sonnet 4.6 (Haiku 4.5 for triage) · cached app context · MCP to host app · memory tool for personalization. See reference-architectures.html Pattern 6 for the diagram.
Sonnet + MCP + caching + memory

Eval / regression suite

Same model as prod · Batch API overnight · cached prompt prefix · evalset versioned in repo.
batch + caching + match-prod-config