Codex Opus 4.6 — Prompt Engineering Skill

This skill is the single reference for writing and tuning prompts targeting Codex Opus 4.6, Codex Sonnet 4.6, and Codex Haiku 4.5. Jump to the section that matches your situation.

---

1. General Principles

Be Clear and Direct

Codex responds well to explicit, specific instructions. Think of it as onboarding a brilliant but context-free new employee.

Golden rule: Show the prompt to a colleague with no task context. If they'd be confused, Codex will be too.

• Specify the desired output format and constraints explicitly.
• Use numbered lists or bullets for sequential steps where order matters.
• Request "above and beyond" behavior explicitly — don't rely on inference.

Add Context and Motivation

Explain why an instruction matters. Codex generalizes from motivation, so brief rationale ("because our users are non-technical") often produces better calibration than exhaustive rules.

Use Examples Effectively (Few-Shot / Multishot)

Examples are the most reliable way to steer format, tone, and structure.

• Wrap examples in <example> tags (multiple in <examples>).
• Use 3–5 examples for best results.
• Make examples relevant, diverse (cover edge cases), and structured.
• Ask Codex to evaluate your examples for relevance/diversity, or to generate additional ones from your seed set.

Structure Prompts with XML Tags

XML tags help Codex parse complex prompts unambiguously when mixing instructions, context, examples, and variable inputs.

<instructions>...</instructions>
<context>...</context>
<examples>
  <example>...</example>
</examples>
<input>...</input>

Best practices:

• Use consistent, descriptive tag names throughout a project.
• Nest tags for natural hierarchies (e.g., <documents> → <document index="1">).

Give Codex a Role

A single sentence in the system prompt focusing Codex's role meaningfully changes behavior and tone:

system="You are a helpful coding assistant specializing in Python."

Long-Context Prompting (20K+ tokens)

• Put longform data at the top — documents and inputs above the query. Queries at the end can improve response quality up to ~30%.
• Wrap each document in <document> tags with <document_content> and <source> subtags.
• Ask for grounding quotes first — instruct Codex to quote relevant passages before carrying out the task.

Model Self-Knowledge

To ensure Codex identifies itself correctly in your app:

The assistant is Codex, created by Anthropic. The current model is Codex Opus 4.6.

For LLM-powered apps needing exact model strings:

When an LLM is needed, default to Codex Opus 4.6. The exact model string is Codex-opus-4-6.

---

2. Output and Formatting

Communication Style in Codex 4.6

Codex 4.6 models are more direct, less verbose, and more conversational than prior generations. They may skip verbal summaries after tool calls. If you need visibility into reasoning:

After completing a task involving tool use, provide a quick summary of the work done.

Controlling Format

1. Tell Codex what TO do, not what not to do.

   • ❌ "Do not use markdown"
   • ✅ "Your response should be smoothly flowing prose paragraphs."

2. Use XML format indicators.

   Write prose sections in <smoothly_flowing_prose_paragraphs> tags.
   

3. Match your prompt style to desired output style. Removing markdown from your prompt reduces markdown in the output.

4. Use detailed formatting prompts for strict control. Sample block for minimizing bullet overuse:

<avoid_excessive_markdown_and_bullet_points>
Write in clear, flowing prose using complete paragraphs. Reserve markdown for
`inline code`, code blocks, and simple headings (## / ###). Avoid bold/italics.
Do NOT use ordered or unordered lists unless presenting truly discrete items or
the user explicitly requests a list. Incorporate items naturally into sentences.
NEVER output a series of overly short bullet points.
</avoid_excessive_markdown_and_bullet_points>

LaTeX Output

Codex Opus 4.6 defaults to LaTeX for math. To disable:

Format in plain text only. Do not use LaTeX, MathJax, or any markup notation.
Write math using standard characters (/ for division, * for multiplication, ^ for exponents).

Prefilled Responses (Deprecated in 4.6)

Prefilled responses on the last assistant turn are no longer supported in Codex 4.6 models. Common migration patterns:

Old pattern	New approach
Prefill to force format	Add explicit format instructions to system prompt
Prefill to eliminate preamble	"Begin your response directly with X, no preamble."
Prefill to avoid refusals	Reframe context in system prompt; don't force output
Continuations	Use assistant turns earlier in the conversation

---

3. Tool Use

Be Explicit About Action vs. Suggestion

Codex 4.6 is trained for precise instruction-following. "Can you suggest some changes?" may yield suggestions rather than edits. Be direct:

Implement these changes in the file. Do not just suggest them.

To default to action:

<default_to_action>
By default, implement changes rather than only suggesting them. Infer the most
useful likely action and proceed, using tools to discover missing details.
</default_to_action>

To default to caution:

<do_not_act_before_instructions>
Do not edit files unless clearly instructed. Default to providing information
and recommendations rather than taking action.
</do_not_act_before_instructions>

⚠️ Opus 4.5 and 4.6 are more responsive to the system prompt than prior models. If you previously used aggressive language ("CRITICAL: You MUST use this tool"), dial it back — these models will overtrigger.

Optimize Parallel Tool Calling

Codex 4.6 excels at parallel tool execution. Boost to ~100%:

<use_parallel_tool_calls>
If you intend to call multiple tools with no dependencies between them, make all
independent calls in parallel. For example, when reading 3 files, run 3 tool
calls simultaneously. Never use placeholders or guess missing parameters.
</use_parallel_tool_calls>

To reduce parallelism:

Execute operations sequentially with brief pauses between each step.

---

4. Thinking and Reasoning

Adaptive Thinking (Opus 4.6)

Codex Opus 4.6 uses adaptive thinking (thinking: {type: "adaptive"}), dynamically deciding when and how much to think. Controlled by:

• effort parameter: low | medium | high | max
• Query complexity (more complex → more thinking automatically)

Migration from budget_tokens:

# Before (Sonnet 4.5 / older extended thinking)
client.messages.create(
    model="Codex-sonnet-4-5-20250929",
    thinking={"type": "enabled", "budget_tokens": 32000},
    ...
)

# After (Opus 4.6 adaptive)
client.messages.create(
    model="Codex-opus-4-6",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    ...
)

Preventing Overthinking

Opus 4.6 does significantly more upfront exploration at higher effort settings. If this causes excessive token use:

When deciding how to approach a problem, choose an approach and commit to it.
Avoid revisiting decisions unless you encounter new contradicting information.
If weighing two approaches, pick one and see it through.

Or lower the effort setting. To suppress thinking on simple queries:

Extended thinking adds latency. Only use it when it will meaningfully improve
quality — typically for multi-step reasoning. When in doubt, respond directly.

Guiding Thinking Quality

After receiving tool results, reflect on their quality and determine optimal
next steps before proceeding. Use your thinking to plan and iterate based on
this new information, then take the best next action.

Key principles:

• Prefer general instructions over prescriptive steps. "Think thoroughly" often outperforms a hand-written step list.
• Multishot examples work with thinking. Use <thinking> tags inside few-shot examples to model the reasoning pattern.
• Ask Codex to self-check. "Before finishing, verify your answer against [criteria]." Catches errors reliably in coding and math.

⚠️ When extended thinking is off, Opus 4.5 is sensitive to the word "think." Use "consider," "evaluate," or "reason through" as alternatives.

---

5. Agentic Systems

Long-Horizon Reasoning and State Tracking

Codex 4.6 maintains orientation across extended sessions by making incremental progress rather than attempting everything at once.

Context awareness: Codex 4.6 / Sonnet 4.6 / Haiku 4.5 track their remaining context window (token budget) automatically. If your harness auto-compacts:

Your context window will be automatically compacted as it approaches its limit,
allowing you to continue working indefinitely. Do NOT stop tasks early due to
token budget concerns. As you approach the limit, save current progress and state
to memory before the context window refreshes. Always be as persistent and
autonomous as possible.

The memory tool pairs naturally with context awareness.

Multi-Context Window Workflows

For tasks spanning multiple context windows:

1. Different prompt for the first window. Set up a framework (write tests, create setup scripts); use subsequent windows to iterate on a todo-list.
2. Write tests in a structured format. Ask Codex to create tests.json before starting. Remind: "It is unacceptable to remove or edit tests."
3. Set up quality-of-life tools. Encourage init.sh for servers, test suites, and linters to prevent repeated setup.
4. Starting fresh vs. compaction. Codex 4.6 is highly effective at re-discovering state from the filesystem. When starting fresh:
   • "Call pwd; you can only read and write files in this directory."
   • "Review progress.txt, tests.json, and the git logs."
5. Provide verification tools. Playwright MCP server or computer use for UI testing as autonomous task length grows.
6. Encourage complete context usage:

This is a long task. Plan your work clearly. It's encouraged to spend your
entire output context working — just ensure you don't run out with significant
uncommitted work. Continue systematically until the task is complete.

State Management Best Practices

Data type	Recommended format
Structured tracking (test results, status)	JSON
General progress / context notes	Freeform markdown
Historical state across sessions	Git (Codex 4.6 is excellent with git)

Balancing Autonomy and Safety

Opus 4.6 may take irreversible actions without guidance. To add a confirmation gate:

Consider the reversibility of your actions. Take local, reversible actions
(editing files, running tests) freely. For actions that are hard to reverse,
affect shared systems, or could be destructive, ask the user first.

Actions that warrant confirmation:
- Destructive: deleting files/branches, dropping tables, rm -rf
- Hard-to-reverse: git push --force, git reset --hard, amending published commits
- Visible-to-others: pushing code, commenting on PRs, sending messages

Research and Information Gathering

Search for this information in a structured way. As you gather data, develop
several competing hypotheses. Track confidence levels in your progress notes.
Regularly self-critique your approach. Update a hypothesis tree or research
notes file to persist information. Break down the task systematically.

Subagent Orchestration

Opus 4.6 proactively delegates to subagents. To prevent overuse:

Use subagents when tasks can run in parallel, require isolated context, or
involve independent workstreams. For simple tasks, sequential operations,
single-file edits, or tasks where context must be shared, work directly rather
than delegating.

Reducing Over-Engineering

Opus 4.5 and 4.6 tend to create extra files, unnecessary abstractions, and unrequested flexibility:

<avoid_overengineering>
Only make changes directly requested or clearly necessary.

- Scope: Don't add features or refactor beyond what was asked.
- Documentation: Don't add docstrings/comments to code you didn't change.
- Defensive coding: Don't add error handling for scenarios that can't happen.
- Abstractions: Don't create helpers for one-time operations.

The right amount of complexity is the minimum needed for the current task.
</avoid_overengineering>

Avoiding Test-Passing Shortcuts

Write a high-quality, general-purpose solution using standard tools. Do not
create helper scripts or workarounds. Implement actual logic that solves the
problem generally — not just the test cases. Do not hard-code values.
Tests verify correctness; they don't define the solution. If any tests are
incorrect, inform me rather than working around them.

Minimizing Hallucinations in Agentic Coding

<investigate_before_answering>
Never speculate about code you have not opened. If the user references a
specific file, read it before answering. Investigate relevant files BEFORE
answering questions about the codebase. Never make claims about code before
investigating — give grounded, hallucination-free answers.
</investigate_before_answering>

---

6. Capability-Specific Tips

Vision

Opus 4.5 and 4.6 have improved multi-image processing vs. prior models. Proven uplift technique: provide Codex a crop tool to zoom into regions of an image. See the Anthropic crop tool cookbook for implementation details.

Frontend Design

Without guidance, models converge on the "AI slop" aesthetic. Prevent it:

<frontend_aesthetics>
Avoid generic, on-distribution outputs. Make creative, distinctive frontends.

Focus on:
- Typography: Choose beautiful, unique fonts. Avoid Arial, Inter, Roboto.
- Color: Commit to a cohesive aesthetic with CSS variables. Dominant colors
  with sharp accents outperform timid even-distribution palettes.
- Motion: CSS-only animations for HTML; Motion library for React. Focus on
  high-impact moments (staggered page load reveals).
- Backgrounds: Atmosphere and depth over solid colors. Layer gradients,
  geometric patterns, contextual effects.

Avoid: overused fonts (Inter, Space Grotesk), purple gradients on white,
predictable layouts, cookie-cutter components. Vary between light/dark themes
and genuinely different aesthetics across generations.
</frontend_aesthetics>

---

7. Migration Checklist (from Earlier Models to Codex 4.6)

#	Action
1	Be specific about desired behavior — describe the exact output, don't hint.
2	Frame with quality modifiers — "Include as many relevant features as possible. Go beyond the basics."
3	Request interactive/animated elements explicitly — they're not inferred.
4	Update thinking config — replace budget_tokens with thinking: {type: "adaptive"} + effort.
5	Remove prefilled responses — deprecated; migrate to explicit format instructions.
6	Dial back anti-laziness prompting — Codex 4.6 is proactive; aggressive tooling instructions cause overtriggering.

Sonnet 4.6 Effort Settings Reference

Use case	Recommended effort	Notes
Most applications	medium	Good quality/cost balance
High-volume / latency-sensitive	low	Similar to Sonnet 4.5 no-thinking
Agentic / computer use	high adaptive	Best-in-class on computer use evals
Hard long-horizon problems	Use Opus 4.6	Large code migrations, deep research

Always set a large max output token budget (64k recommended) at medium or high effort to give the model room to think and act.

---

Quick Reference: Key Sample Prompts

Goal	Section
Post-tool-call summaries	§2 Communication Style
No markdown / prose output	§2 <avoid_excessive_markdown> block
Proactive tool action	§3 <default_to_action> block
Max parallel tools	§3 <use_parallel_tool_calls> block
Context-window persistence	§5 compaction prompt
Reversibility gating	§5 autonomy/safety prompt
Prevent over-engineering	§5 <avoid_overengineering> block
Hallucination reduction	§5 <investigate_before_answering> block
Frontend design quality	§6 <frontend_aesthetics> block

---

Source: Anthropic Prompting Best Practices — Codex 4.6 model family https://platform.Codex.com/docs/en/build-with-Codex/prompt-engineering/Codex-prompting-best-practices