Opus 4.6 Prompting

Opus 4.6 builds on the 4.5 prompting philosophy with key additions: adaptive thinking (4 effort levels), 1M token context (beta), 128K output, agent teams, and server-side compaction. The core principle remains—natural language over directives.

What Changed from 4.5 to 4.6

Feature	Opus 4.5	Opus 4.6
Thinking	Extended thinking (binary on/off + budget_tokens)	Adaptive thinking (low/medium/high/max)
Context	200K tokens	1M tokens (beta), context rot eliminated
Output	64K tokens	128K tokens
Agent architecture	Single agent, sub-agent delegation	Agent teams (parallel coordination)
Compaction	Client-side	Server-side compaction API (beta)
Prefilling	Supported	Disabled (400 error), use structured outputs
"Think" sensitivity	Highly sensitive	Reduced (adaptive thinking handles calibration)

Key Behavioral Patterns

Behavior	Implication
Literal instruction following	Be explicit about desired behaviors
Adaptive thinking	Let the model calibrate reasoning depth; specify effort level only when needed
Highly responsive to system prompt	Use natural language, not directives
Tool trigger sensitivity	Replace MUST use with Use when...
1M context awareness	Context rot eliminated; long conversations maintain quality

Quick Migration Checklist

When upgrading prompts from 4.5 or earlier:

1. Remove aggressive language (CRITICAL, MUST, NEVER, ALWAYS)
2. Replace budget_tokens extended thinking with adaptive thinking effort levels
3. Replace assistant prefilling with output_config.format or system prompt instructions
4. Add context/motivation ("because..." explanations)
5. Match prompt formatting to desired output style
6. Trust the model more—reduce over-specification
7. Use XML tags for agent workflows (<current-state>, <blocked>, <workflow>)
8. Consider agent teams for parallel workloads instead of sequential sub-agents

Reference Files

File	Use When
patterns.md	Optimizing any prompt—transformation rules and examples
agentic-patterns.md	Building agents, agent teams, and long-horizon workflows

Utility Script

The scripts/analyze-prompt.py script analyzes prompts for deprecated patterns:

echo "Your prompt text" | python3 scripts/analyze-prompt.py

Configure thresholds in scripts/config.json:

• min_words: Minimum words to trigger analysis (default: 1)
• enabled: Enable/disable the analyzer (default: true)

Pattern Categories

The patterns.md reference covers:

1. Aggressive language — Soften MUST, NEVER, CRITICAL
2. Think variants — Less critical with adaptive thinking, but still worth cleaning up
3. Tool invocation — Prevent over-triggering
4. Formatting — Match prompt style to output style
5. Context/motivation — Add "because..." explanations
6. Over-specification — Trust the model more
7. Verbosity — Optimize for conciseness
8. XML structure — Use semantic tags for agent workflows

Agentic Best Practices

The agentic-patterns.md reference covers:

1. Context management — 1M context window reduces compaction pressure
2. Agent teams — Parallel coordination for complex projects
3. Sub-agent delegation — Preserve main context for research tasks
4. Tool design — Avoid over-triggering in descriptions
5. State tracking — JSON for structured, text for notes
6. Long-horizon workflows — Incremental progress patterns
7. Parallel execution — When to parallelize vs sequence
8. Error handling — Fail fast, document blockers