"LLM Reasoning Modes (4/6) — Claude Code's effort in Practice: low·medium·high·xhigh·max and ultracode"

Part 3 took the API's effort parameter (low–max) head-on. Part 4 looks at the same dial once it moves into the Claude Code CLI. The values are identical — what differs is how you set them, how long they persist, and what ultracode is.

Claude Code's effort isn't a new concept. It's the exact same values from the API's output_config.effort (Part 3), surfaced in the CLI. So Part 4's goal isn't to introduce "another knob" — it's to lay out how you turn that knob in practice: the /effort command, the environment variable, session persistence, and ultracode, which appears in the menu but isn't actually an effort level.

In One Paragraph

Claude Code exposes five effort levels, conservative → capable: low, medium, high, xhigh, max. These are the same effort values the API accepts, surfaced in the UI, and the default is the same as the API's: high. xhigh is Anthropic's recommendation for real coding and agent sessions; max is the deepest reasoning with no constraint on token spend. But max applies to the current session only (unless set via the CLAUDE_CODE_EFFORT_LEVEL environment variable) — while low/medium/high/xhigh persist across sessions. ultracode appears in the menu but is not a sixth effort level. It's a Claude Code packaging of xhigh plus standing permission to launch multi-agent workflows; the API's complete effort set is still low–max.

1. Five Levels — The API's Values, Surfaced

Claude Code's effort runs from conservative to capable across five levels.

low → medium → high → xhigh → max

The first thing to note is that these five are not new CLI-only values. The set the API's effort parameter accepts (Part 3) is exactly low / medium / high / xhigh / max, and Claude Code simply surfaces those values in the UI. The set the API accepts and the set the CLI shows are identical.

So the default is identical too. Claude Code's default effort is high — exactly the API's default. This high default is the launch default on Fable 5, Opus 4.8, Opus 4.6, and Sonnet 4.6. In other words, open Claude Code with no configuration and it runs at high.

2. How Behavior Changes — Mirroring the API

When you raise or lower effort, the way Claude Code's behavior shifts mirrors the API behavior from Part 3. Effort controls not just thinking depth but how the whole response spends tokens — the number of tool calls, the preamble, the detail of the completion report.

The lower the effort: - Fewer tool calls. - Starts working without preamble. - Gives a brief completion report.

The higher the effort: - Explains the plan before starting. - Summarizes changes in detail. - Writes more thorough code comments.

So for the same task, lower effort shifts toward "work quietly and fast, report briefly," while higher effort shifts toward "lay out a plan, go through more tools, summarize thoroughly."

3. The Five Levels at a Glance — Persistence, Behavior, Use

Level Persists across sessions? Behavior When to use
low Persists Minimal tool calls, starts without preamble, brief report Narrow one-answer work — a rename, drafting a single line
medium Persists Balanced between low and high The no-think-about-it default
high Persists Plan explanation, detailed summaries, thorough comments (= default) Most substantive work
xhigh Persists Deeper reasoning and longer horizon than high Real coding / agent sessions (Anthropic's recommendation)
max Current session only (env var is the exception) Deepest reasoning, no constraint on token spend Only the hardest, latency-insensitive problems

One row stands apart in this table — max's persistence.

  • low / medium / high / xhigh persist across sessions once set. Reopen Claude Code later and the value holds.
  • max applies to the current session only. Close the session and it clears. Except: when set via the CLAUDE_CODE_EFFORT_LEVEL environment variable — pin it through the env var and max carries across sessions too.

This asymmetry reads as a deliberate safeguard. Because max has no constraint on token spend, the default behavior is designed to keep it from quietly hardening into a session-wide default and burning cost indefinitely.

4. How to Set It — Three Configuration Surfaces

There are three ways to set effort in Claude Code, each with a different persistence scope.

  • /effort (interactive, per-session): type /effort mid-conversation to pick a level. This path applies to the current session.
  • CLAUDE_CODE_EFFORT_LEVEL environment variable (persistent): pin it through the env var and it persists across sessions — as noted above, this is the one path that carries even max across session boundaries.
  • Per-repo settings: you can pin an effort level in a project's settings (e.g., pinning an effortLevel in the project settings). That keeps the working tone for a repo right next to the code.

To distinguish the three in a line — /effort is "this session only," the env var is "persistent across my environment," and per-repo settings are "always, in this project."

5. What ultracode Actually Is — Not a Sixth Effort Level

Here's the most common misconception to clear up. Claude Code's effort menu shows ultracode alongside low / medium / high / xhigh / max. So it's easy to think "ultracode is a sixth effort level above max." It is not.

Here's what's actually true.

  • The API's effort set is still complete at low–max. There is no additional effort level called ultracode in the API.
  • ultracode pairs the xhigh effort level with standing permission for Claude Code to launch multi-agent workflows. That permission is granted via mid-conversation system messages.
  • In other words, ultracode is not a new level — it's a Claude Code packaging of xhigh + that permission.

Put simply, on the "depth" axis of effort, ultracode equals xhigh. What's added on top is permission — letting Claude Code resolve larger jobs across multi-agent workflows — not reasoning depth. It sits in the menu side by side, so it feels like the same kind of dial, but it's more accurate to read it as "five effort levels, plus one execution mode layered on top."

6. So Which Level, When — Matching in Practice

Map each level's meaning to task difficulty and you get the following. The principle from Part 1 — not "always maximum" but match the dial to task difficulty — applies just as much in the CLI.

  • low — work that narrows to essentially one answer. A rename, drafting a single line — tasks with almost no room to reason.
  • medium — the no-think-about-it default. A solid choice for general flows that need balance.
  • high — most substantive work. The default, and the balance point between quality and token efficiency.
  • xhigh — real coding sessions, agent runs. Anthropic's recommendation, and the starting point when you put it to real work.
  • max — only for the hardest, latency-insensitive problems. Since it has no token constraint, it's also the costliest, so you don't leave it as a standing default (which is why it applies to the current session only by default).

In practice: make xhigh the default starting point for coding and agent work, drop down for light tasks (low/medium), and raise to max only for a single genuinely hard problem — which matches the source's recommendation.

What Comes Next

Parts 3 and 4 have now covered the Claude side of the dial in full — the API's effort and Claude Code's effort in practice.

Part 5 switches camps to OpenAI and Codex's reasoning_effort — how minimal / low / medium / high / xhigh differ across GPT-5 / 5.5 / 5.2-codex, how you set it in config.toml, and what verbosity is, the separate knob OpenAI splits output length onto, unlike Claude's single effort dial.

Claude Code's effort levels, defaults, session persistence, and ultracode's behavior are grounded in Anthropic's primary documentation (effort, Claude Code model config) and the model migration guide.

Series overview: Series index

댓글

댓글 쓰기

이 블로그의 인기 게시물

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »

"ML Foundations (6/9) — Neural Networks: From Perceptron to MLP"

이미지
← 5/9 Regularization & Model… 📚 Series Index 7/9 Deep Learning Training → A neural network is logistic regression from Part 4 stacked into layers. Each layer passes through a new nonlinear activation, expanding the family of functions the network can represent. In this part we put how a neural network learns onto one page. Perceptrons, XOR, MLPs, backpropagation, activation functions, and the universal approximation theorem all meet here. 0. Learning Objectives Write the perceptron's learning rule and explain why it cannot solve XOR. Express an MLP's forward pass as one line of matrix products. Derive backpropagation by hand on a two-layer MLP. Compare the definitions, derivatives, and trade-offs of sigmoid, tanh, ReLU, and GELU. State the universal approximation theorem (Cybenko 1989 / Hornik 1991) precisely. Train an MLP end-to-end in PyTorch. 1. 핵심 요약 Perceptron (1958): \(\hat{y} = \mathrm{step}(w^\top x + b)\). Converges only on linearly separabl...
Read more »