Override a Prompt

Some built-in prompt texts can be changed. This page lists each one, says whether you can replace it, append to it, or turn it off, and shows how. Most rows were tested live against the local stub (source "Probe"). The rest come from reading the program (source "The program") and were not run.

Replace means the original text is gone and yours takes its place.
Append means the original stays and your text is added.
Turn Off means the text is removed.
Indirect means the text stays the same, but it tells the model to use something you control.
No means we found no way to change the text.

A subscription (OAuth) login gets the same main system prompt text and the same 12 default tools as an API key (live probe live.oauth-parity on 2.1.307 and 2.1.308, 2026-10-22, and on 2.1.309, 2026-10-23, with the prompt Reply with 'OK'.; the text matched once model names, temp paths, and the per-run billing hash were masked). It also got the same Explore subagent prompt in a live run on 2.1.305 (2026-10-21), which was not run again, because the model decides by itself whether to start Explore. The Explore request had the same 7 tools: Bash, ListAgents, Read, ReportFindings, Skill, DeferredToolPlaceholder, and ToolSearch. In the 2.1.305 run, we compared the main request with an API key request on the same model (Opus 5.5). The system prompt differed only in the memory folder path. Other differences:

The two cached system blocks and messages[1] (role system) have a cache ttl of "1h" with a subscription. The API key request had no ttl.
thinking.display was updates with a subscription and omitted with an API key.
The ScheduleWakeup description is different. With a subscription it says requests use a 1-hour prompt-cache TTL. With an API key it says the default 5-minute TTL. Both say the program clamps delays to 60 to 3600 seconds.
The first user message listed RemoteTrigger as a deferred tool and a schedule skill only with a subscription, and the code-review skill text was longer there (cloud review and a --post option).
The request metadata IDs differ, as expected.

Agent Merge Rule

Agents are merged by name. The later source wins: built-in, plugin, user (~/.claude/agents), project (.claude/agents), --agents JSON, managed settings. A same-named file therefore replaces a built-in agent completely: body, tools, and model. This was tested for Explore (project file) and general-purpose (--agents).

Main System Prompt

Prompt	Override	How	Source
Main system prompt	Replace	`--system-prompt` or `--system-prompt-file`.	Probe
Main system prompt	Append	`--append-system-prompt` or `--append-system-prompt-file`.	Probe
Main thread agent	Replace	`--agent`.	Probe
Auto memory section	Turn Off	`CLAUDE_CODE_DISABLE_AUTO_MEMORY=1` (tested). `autoMemoryEnabled: false` setting (not tested).	Probe
Git commit and PR rules in the Bash tool text	Turn Off	`includeGitInstructions: false` (tested) or `CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS=1` (not tested).	Probe
Chrome integration prompt	Turn Off	`CLAUDE_CODE_DISABLE_CFC_PROMPT=1`.	The program
Resume prompt (`Continue from where you left off.`)	Replace	`CLAUDE_CODE_RESUME_PROMPT=<text>`.	The program
Managed memory	Append	`claudeMd` in managed settings only. CLAUDE.md-style text is added as org memory.	The program

Subagents and Built-in Agents

Prompt	Override	How	Source
Explore agent	Replace	A same-named agent file, `.claude/agents/Explore.md`. You get no read-only tool list unless you set `tools`.	Probe
Explore and Plan agents	Turn Off	`CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS=1` removes both from the agent list.	Probe
Plan agent	Replace	Same rule as Explore: a same-named agent file.	The program
general-purpose agent	Replace	`--agents` JSON with the same name, or a same-named agent file.	Probe
statusline-setup agent	Replace	A same-named agent file.	The program
claude agent	Replace	A same-named agent file.	The program
claude agent	Turn Off	`CLAUDE_CODE_DISABLE_AGENT_VIEW=1` or the `disableAgentView` setting.	The program
web-fetch agent	Replace	A same-named agent file. Off by default. `CLAUDE_CODE_WEB_FETCH_AGENT=1` turns it on and `CLAUDE_CODE_DISABLE_WEB_FETCH` turns it off.	The program
claude-code-guide agent	Replace	A same-named agent file. This rests on the shared agent merge code and was not tested for this agent. Not offered when the entry point is `sdk-ts`, `sdk-py`, or `sdk-cli`, so it is not available in `-p`. The interactive screen offers it (checked on 2.1.308).	The program
Workflow subagent	No	No override found.	The program
comment-thread-analyst agent	No	No override found.	The program
Coordinator and worker prompts	No	No text override. Coordinator mode replaces the agent list and the main prompt. `--append-system-prompt` is still added.	The program
All subagent prompts	Append	`--append-subagent-system-prompt <text>` or `--append-subagent-system-prompt-file <file>`. Print mode only. Hidden from `--help`. It reaches nested subagents.	Probe
All built-in agents	Turn Off	`CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS=1` in print mode removes the whole agent list.	Probe
Memory text for subagents	Turn Off	Set `CLAUDE_CODE_MEMORY_SUBAGENT_APPEND` and the memory block is not appended. Only matters with memory stores.	The program
Custom agent fields	Append	Frontmatter fields `initialPrompt`, `observer`, `observerMessage`, `omitClaudeMd`, and `skills` are text you set.	The program
Observer prompt	Append	`observerMessage` in agent frontmatter. The base text is fixed.	The program

The flags --system-prompt and --append-system-prompt do not reach built-in subagents. See Subagent with --system-prompt and --append-system-prompt.

Slash Commands and Skills

Prompt	Override	How	Source
/init	Replace	A command file, `.claude/commands/init.md`. `~/.claude/commands` works too.	Probe
/security-review	Replace	Same rule as `/init`.	The program
/simplify	Replace	A same-named skill, `.claude/skills/simplify/SKILL.md`. The skill wrapper adds a `Base directory` line.	Probe
/debug	Replace	Same rule as `/simplify`.	The program
/code-review and its `/review` alias	Replace	Same rule as `/simplify`.	The program
Other bundled skills: batch, claude-api, dataviz, deep-research, doctor, fewer-permission-prompts, goal, insights, loop, plugin-authoring, run, run-skill-generator, team-onboarding, update-config, verify, workflow-authoring	Replace	A same-named skill file.	The program

/recap, /auto-mode-setup, and /autocompact sent no model request, so they have no prompt to change. /pr-comments does not exist in 2.1.311.

Output Styles

Prompt	Override	How	Source
Built-in styles: Explanatory, Learning, Proactive, Concise	Replace	You cannot edit them. A custom style file with the same `name`, such as `~/.claude/output-styles/explanatory.md`, replaces one.	Probe

Compaction

Prompt	Override	How	Source
Compaction prompt	Append	Text after `/compact` goes under `Additional Instructions:`.	Probe
Compaction prompt	Append	Standard output of a `PreCompact` hook goes in the same place, after your text.	Probe
Compaction prompt	Indirect	A `## Compact Instructions` section in CLAUDE.md is not copied into the compaction text. The compaction text tells the model to look for it in the context it already has.	Probe

The base text stays in every case. The program picks one of several base texts by a server flag. The control text was seen.

Hooks

Prompt	Override	How	Source
Prompt hook	No	You write `prompt` (`$ARGUMENTS` is the hook JSON) and `model`. The system text is fixed. It is different on `Stop`.	Probe
Agent hook	No	You write `prompt` and `model` (default Haiku). The system text is fixed and has a `StructuredOutput` tool.	Probe

Auto Mode

Prompt	Override	How	Source
Auto mode classifier	Append	The `autoMode` setting has `allow`, `soft_deny`, `hard_deny`, `environment`, and `classifyAllShell`. `"$defaults"` keeps the built-in rules at that spot.	Probe
Auto mode classifier	Replace	A list without `"$defaults"` replaces the built-in rules. In the test, `soft_deny` lost 72 default rules.	Probe

The autoMode object is also sent to the server in safeguards.classifier_context on the main request. The test used the local fallback classifier.

Plan Mode

Prompt	Override	How	Source
Plan mode workflow	Replace	`--plan-mode-instructions <text>` (hidden flag) replaces the phases. The read-only preamble and the `ExitPlanMode` footer stay.	Probe

Other

Prompt	Override	How	Source
Tool descriptions	No	`--tools`, `--allowedTools`, and `--disallowedTools` only add or remove tools. `toolAliases` is an SDK option for names, not text. MCP tool text comes from the server.	Probe
Session title and branch name	No	No override found. See also session naming.	The program
Memory selection and extraction	No	No override found.	The program
Surface prompts	No	No override found. The entry point picks the text.	The program
SDK `systemPrompt` presets	No	The SDK package passes flags to the program. The program has only `--system-prompt[-file]` and `--append-system-prompt[-file]`. Not tested.	The program

Examples

Replace the Explore Agent

Save this as .claude/agents/Explore.md. The name must match.

---
name: Explore
description: Find files and answer questions about the code.
tools: Read, Grep, Glob
---

You search the codebase and report what you find. Do not edit files.

Append to Every Subagent

This works in print mode only. The flag is hidden from --help.

claude -p --append-subagent-system-prompt "Answer in one short paragraph." "Use a subagent to find where login is handled."

Keep the Auto Mode Defaults

Add your own rule and keep the built-in rules with "$defaults".

{
  "autoMode": {
    "soft_deny": ["$defaults", "Never push to the main branch."]
  }
}

Replace a Bundled Skill

Save this as .claude/skills/simplify/SKILL.md. The skill name matches the built-in /simplify.

---
name: simplify
description: Simplify the code I changed.
---

Review the changed code. Remove dead code and extra layers. Keep behavior the same.

Cases with No Override

Tool descriptions. You can add or remove tools. You cannot change their text.
Hook prompt system text. You write only the prompt and model.
Subagents and the main thread flags. --system-prompt and --append-system-prompt do not reach built-in subagents. Use --append-subagent-system-prompt.
Compact Instructions in CLAUDE.md. The ## Compact Instructions section is not copied into the compaction text. It works only because the model is told to look for it in its context.

Agent Merge Rule​

Main System Prompt​

Subagents and Built-in Agents​

Slash Commands and Skills​

Output Styles​

Compaction​

Hooks​

Auto Mode​

Plan Mode​

Other​

Examples​

Replace the Explore Agent​

Append to Every Subagent​

Keep the Auto Mode Defaults​

Replace a Bundled Skill​

Cases with No Override​