Skip to content
Claude Code Internal Analysis

Prompt template catalog

This page is a non-table index into the long string / template literals embedded in cli.renamed.js that look prompt-shaped (instructions to the model, tool descriptions, hook docs, agent files, etc.). The actual prompt bodies live under prompts/ — one file per category, each carrying the full text the bundle ships (template-literal interpolation sites are preserved as ${…}).

The catalog is regenerated from the Babel AST by scripts/extract-prompt-catalog.mjs; machine-readable data lives at docs/99-research-atlas/data/prompt-catalog.json.

Source provenance

  • Input: claude-code-pkg/src/entrypoints/cli.renamed.js (23,273,325 bytes)
  • SHA-256: fd212af5897bf4f5b2c4eee2863ad46140d003abd8569adda2dd32b5857a495b
  • Generated: 2026-05-18
  • AST parser: @babel/parser with typescript + jsx plugins, walked by @babel/traverse
  • Minimum literal length: 120 characters
  • String / template literals scanned: 167,464
  • Candidates above length threshold: 2,265
  • Kept after domain / noise filter: 1,086

Categories

system-context-memory — 46 prompts

Top-level system prompts, memory file blocks, <environment> / <user_*> reminders, and rolling conversation-summary scaffolding.

tool-description-or-guard — 74 prompts

Tool descriptions, usage guards, and tool-output guidance — typically attached to a LK({ ... }) tool registration.

security-permission — 72 prompts

Permission policy text, sandbox / credential rules, and security-relevant guardrails surfaced to the model or hooks.

task-and-agent — 27 prompts

Agent (subagent) definitions, Task tool descriptions, and managed-agents reference material.

mcp-plugin-hook — 190 prompts

MCP server / plugin / hook descriptions, event documentation, and schema .describe(...) text.

slash-command-or-agent-file — 8 prompts

Embedded slash-command / sub-agent files (----fenced frontmatter + body) shipped inside the bundle.

slash-output-style — 3 prompts

Output-style modifier prompts that change how Claude formats its reply for the current session.

structured-output — 8 prompts

Prompts that demand strict JSON-only or fixed-shape responses (RESPOND WITH ONLY A VALID JSON OBJECT, etc.).

misc-prompt-like — 658 prompts

Long literals that look prompt-shaped but did not match a more specific category.

How candidates are filtered

A literal is kept if it satisfies every rule:

  1. Length — at least 120 characters.
  2. Domain keyword — contains at least one in-domain term (Claude, agent, tool, MCP, hook, plugin, skill, sandbox, permission, slash-command, memory, session, context window, model, prompt, /loop, /dream, /schedule, /output-style, worktree, elicitation, PreToolUse/PostToolUse/UserPromptSubmit/SessionStart/SessionEnd, etc.).
  3. Prose ratio — the count of {}()<>;= characters is no more than 25 % of the alphabetic character count, dropping snippets that are mostly code.
  4. Not a known SDK / vendored noise pattern — strings beginning with Anthropic:, The model ', Using Claude with , Cannot , Could not , Unable to , Error: , TypeError: , generic SDK validation strings, and bare URLs are dropped.

Each surviving literal is then assigned one category by checking, in order: slash-command-or-agent-file (YAML-fenced body), slash-output-style, structured-output, mcp-plugin-hook, security-permission, task-and-agent, tool-description-or-guard, system-context-memory. Anything that misses every probe lands in misc-prompt-like.

Caveats

  • The catalog is exhaustive for literal prompts above the length threshold that pass the filter. It does not capture prompts assembled at runtime from many short fragments.
  • enclosing is the closest named function in the AST (FunctionDeclaration / VariableDeclarator / ObjectProperty / ClassMethod). When a literal lives at module top level the entry is marked top-level.
  • Line numbers index into the regenerated cli.renamed.js. The byte offset (0x…) is stable across reformatting.
  • Categories are heuristic. A prompt may legitimately belong to several categories; the rule order above picks one.
  • Template-literal interpolation sites are shown as ${…} rather than rendered — the runtime substitutes those at call time, so the body below is the static skeleton, not the rendered prompt sent to the model.

Created and maintained by Yingting Huang.