Repository / project instructions
DocumentedOfficial docs define managed, user, project, and local instruction scopes.
View prompt evidenceAgent intelligence
Claude Code
Official Claude Code evidence covers its CLAUDE.md hierarchy, project memory, tools, and permission controls without establishing a public full system prompt.
Claude Code is documented as a multi-surface coding-agent product with scoped CLAUDE.md instructions, repository auto memory, codebase tools, and permission rules. The selected official sources do not establish a complete default base system prompt.
Scope: Claude Code terminal CLI · Claude Code IDE integrations · Claude Code desktop app · Claude Code web · Selectable Claude model · reviewed Aug 31, 2026
Reviewed product surfaces
Product Surface Schematic
Each layer links to the detailed evidence map. Its state applies only to the reviewed evidence.
Reviewed surfaces Claude Code terminal CLI · Claude Code IDE integrations · Claude Code desktop app · Claude Code web
- Base system promptNot established
- CLAUDE.md instructionsDocumented
- Auto memoryDocumented
- ToolsDocumented
- Permission controlsDocumented
- Not established
- Documented
Source-backed answers
Key behaviors
Each answer is scoped to the selected evidence. “Not established” means the evidence does not support the claim; it does not assert absence.
Memory / persistent guidance
DocumentedA separate repository-scoped memory store can preserve useful learnings across sessions.
View prompt evidenceFilesystem and shell tools
DocumentedOfficial product documentation says Claude Code reads code, edits files, and runs commands.
View prompt evidenceApprovals / permissions
DocumentedDeny, ask, and allow rules plus permission modes constrain tool execution.
View prompt evidenceBase prompt visibility
Not establishedAnthropic does not publish a complete default Claude Code system prompt in the selected sources.
Review evidence scopeEditorial guidance
What this means in practice
These conservative editorial implications come from the linked evidence. They add no vendor claims or generic prompt advice.
Editorial implication
Use the appropriate CLAUDE.md scope for durable guidance.
Managed, user, project, and local instruction scopes are documented rather than treated as one undifferentiated prompt.
Supporting evidenceEditorial implication
Keep human-authored CLAUDE.md instructions separate from auto memory.
Repository auto memory preserves selected learnings in a distinct store from CLAUDE.md guidance.
Supporting evidenceEditorial implication
Account for permission evaluation before tool execution.
Deny, ask, and allow rules plus permission modes can constrain read, edit, command, and external-tool workflows.
Supporting evidenceUseful next actions
Continue your research
Compare agent behaviorsReview the same evidence-backed questions across all six products.
Inspect instruction evidenceThe instruction page maps the documented architecture; it is not a full or extracted system prompt.
How does Claude Code compare with Codex?Compare prompt evidence, project context, tools, permissions, and memory claim by claim.
How does Claude Code compare with Gemini CLI?Compare base-prompt posture, project instruction scopes, dynamic context, and controls.
Review source and provenance detailsVerify origin, retrieval, completeness, publication limits, and rights.
Product scope
Claude Code is a coding-agent product; its selected Claude model is a separate input.
- Organization
- Anthropic
- Product
- Multi-surface agentic coding product
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations · Claude Code desktop app · Claude Code webThe full-featured terminal surface for agentic coding work.
- Model context
- Selectable Claude modelClaude Code exposes model selection; this package does not pin one model as universal for the product.
Evidence status
Coverage and review
Official documentation covers instruction scopes, memory, capabilities, and permissions. The review does not rely on independent prompt captures or claim a complete default prompt.
- Coverage
- partial
- Primary source
- Tier A · vendor-published
- Review freshness
- currentThe official overview was re-reviewed for terminal, IDE, desktop, and web surfaces; the existing CLAUDE.md, auto-memory, and permission evidence remains scoped to its cited sources. Targeted planning and web evidence was consulted on September 14, 2026; see the separately dated sources. Other evidence was not re-reviewed.
- Reviewed
- Aug 31, 2026
Detailed instruction layers
Instruction-system map
The instruction map separates base prompt availability, CLAUDE.md context, auto memory, tools, and client-enforced permission controls.
Base system promptnot established
Anthropic does not publish a complete default Claude Code system prompt in the selected sources.
Source detailCLAUDE.md instructionsdocumented
Official docs define managed, user, project, and local instruction scopes.
EvidenceAuto memorydocumented
A separate repository-scoped memory store can preserve useful learnings across sessions.
EvidenceToolsdocumented
The product reads code, edits files, runs commands, and connects to external tools.
EvidencePermission controlsdocumented
Deny, ask, and allow rules plus permission modes constrain tool execution.
EvidenceSupporting instruction evidence
Prompt and snapshot evidence
The instruction page maps the documented architecture; it is not a full or extracted system prompt.
Inspect instruction evidence- Selected snapshot
- Claude Code instruction architecture · reviewed August 11, 2026
- Evidence coverage
- Through Aug 11, 2026Claude Code instruction architecture · reviewed August 11, 2026
- Latest known vendor version
- Not established by selected evidence
- Current product state
- Not established by selected evidence
- Publication
- Full prompt unavailable
- Instruction page
- Available, excluded from indexingUseful original intelligence page, but no full first-party prompt artifact is established.
Evidence-backed capabilities
Capabilities with evidence
Each capability links to its supporting prompt section or first-party source and keeps its stated confidence level.
| Capability | What the source establishes | Confidence | Evidence |
|---|---|---|---|
| Plan mode | Plan mode researches and proposes changes before source edits. It can run exploratory shell commands and write a plan. Outside sessions with bypass permissions available, source edits remain blocked until plan approval. Command review depends on mode, settings, and version; approval exits planning into the selected execution mode. CLI evidence is not a universal read-only sandbox claim. | high | Primary source |
| Web tools | WebSearch returns search titles and URLs; WebFetch retrieves and processes a page, with lossy summarization and possible truncation. Tool/domain rules and provider support apply. Amazon Bedrock does not expose the server-side search tool; an Azure-hosted Foundry deployment does not support it. Web tools do not establish arbitrary network access. | high | Primary source |
| Codebase operations | Official product documentation says Claude Code reads code, edits files, and runs commands. | high | Prompt section |
| Scoped CLAUDE.md instructions | Managed, user, project, and local instruction scopes are officially documented. | high | Prompt section |
| Repository-scoped auto memory | Auto memory stores selected learnings separately from human-authored CLAUDE.md instructions. | high | Prompt section |
| Permission rules and modes | Deny, ask, and allow rules govern tool access with documented precedence. | high | Prompt section |
Version evidence
Snapshots and meaningful changes
No compatible public prompt snapshots are established. Product changelog entries are not treated as prompt versions.
- selected Claude Code instruction architecture · reviewed August 11, 2026Reviewed Aug 11, 2026Instruction evidence
Provenance
Source ledger
Every source lists its origin, retrieval date, completeness, publication limits, and rights separately.
Tier A · vendor-published
Claude Code permission modes — Plan mode
- Publisher
- Anthropic
- Retrieved
- Sep 14, 2026
- Upstream ref
- Live official documentation consulted 2026-09-14
- Rights posture
- Original summary; no prompt text republished.
Plan mode researches and proposes changes before source edits. It can run exploratory shell commands and write a plan. Outside sessions with bypass permissions available, source edits remain blocked until plan approval. Command review depends on mode, settings, and version; approval exits planning into the selected execution mode. CLI evidence is not a universal read-only sandbox claim.
Open primary source Tier A · vendor-published
Claude Code tools reference — WebSearch and WebFetch
- Publisher
- Anthropic
- Retrieved
- Sep 14, 2026
- Upstream ref
- Live official documentation consulted 2026-09-14
- Rights posture
- Original summary; no prompt text republished.
WebSearch returns search titles and URLs; WebFetch retrieves and processes a page, with lossy summarization and possible truncation. Tool/domain rules and provider support apply. Amazon Bedrock does not expose the server-side search tool; an Azure-hosted Foundry deployment does not support it. Web tools do not establish arbitrary network access.
Open primary source Tier A · vendor-published
Claude Code overview
- Publisher
- Anthropic
- Retrieved
- Aug 31, 2026
- Upstream ref
- Live official documentation reviewed 2026-08-31
- Rights posture
- Original summary of official documentation; no raw prompt copied.
Establishes product surfaces and capabilities, not the full default system prompt.
Open primary source Tier A · vendor-published
How Claude remembers your project
- Publisher
- Anthropic
- Retrieved
- Aug 11, 2026
- Upstream ref
- Live official documentation reviewed 2026-08-11
- Rights posture
- Original summary of official documentation.
Documents instruction and memory mechanisms; it is not a system-prompt disclosure.
Open primary source Tier A · vendor-published
Claude Code permissions
- Publisher
- Anthropic
- Retrieved
- Aug 11, 2026
- Upstream ref
- Live official documentation reviewed 2026-08-11
- Rights posture
- Original summary of official documentation.
Documents permission enforcement but not every tool or deployment policy.
Open primary source