Find Prompt

Find the right prompt and run with it

You route the current task to the right file(s) in this library, read them yourself, and then continue the existing work under those instructions — not from scratch, but picking up the conversation and any task already in progress.

${CLAUDE_SKILL_DIR} is this skill's own directory (<library-root>/.claude/skills/find-prompt/). The library root is ${CLAUDE_SKILL_DIR}/../../... All prompt files are under ${CLAUDE_SKILL_DIR}/../../../prompts/english/. Never hardcode an absolute path — always build it from ${CLAUDE_SKILL_DIR} so this works for anyone who clones the library, wherever they put it.

If ${CLAUDE_SKILL_DIR} is not substituted (it shows up literally), fall back: the skill sits next to SKILL.md; run ls up the tree from your working directory to find the folder containing prompts/english/agents/claude-agent-system-prompt.md.

Steps

1. Determine the task. Use $ARGUMENTS if given. Otherwise use the task already being worked on in this conversation (the user's last real request, the in-progress work, the files open). If genuinely unclear, ask one short question and stop.

2. Match the task to the routing table below. Pick the base prompt plus specialists by tier: 1 for a single-domain task (default), 2 only if the task spans two genuinely independent domains, or route to Multi-Agent Orchestration if the work needs isolation/review. See ${CLAUDE_SKILL_DIR}/../../../prompts/english/workflows/prompt-selector-guide.md for the tier definitions and the conflict-precedence rule if loaded prompts disagree.

3. Read the file(s) with the Read tool, using paths built from ${CLAUDE_SKILL_DIR}:

   • base: ${CLAUDE_SKILL_DIR}/../../../prompts/english/agents/claude-agent-system-prompt.md
   • specialist: ${CLAUDE_SKILL_DIR}/../../../prompts/english/<path from the table> If a path does not exist, tell the user to run /update-prompts and stop.

4. Adopt them and continue. Treat the base prompt as your operating system prompt and the specialist as the domain protocol for this task. Then:

   • Restate the current task in one line (from the conversation, not invented).
   • State the success criteria you'll hold yourself to.
   • Carry over every session constraint already established (no git, no server start, static-only, don't touch X, etc.).
   • Pick up the work where it stands — continue the in-progress task under the new protocol. Do not restart, do not re-ask what was already answered, do not re-explore what was already explored.

5. Confirm in one short block which prompts you loaded and what you're now doing, then proceed.

Routing table

Paths are relative to prompts/english/.

Running Claude Code itself (base prompt not needed — the Claude Code prompt is enough alone)

Building software (base = Agent System + the one specialist below)

Improving existing code (base = Agent System + the one specialist below)

Deciding (base = Agent System + the one specialist below)

Rules

• Composition by tier, not a hard cap. Tier 1 (base + 1 specialist) covers most tasks. Move to Tier 2 (base + 2 specialists) only when the task genuinely spans two independent domains — say so explicitly. Route to Multi-Agent Orchestration (Tier 3) when the work needs isolated workers or adversarial review.
• The base is always agents/claude-agent-system-prompt.md, except for a task purely about running Claude Code (then the single Claude Code prompt is enough).
• Tiny task (typo, rename, one-line fix): don't load anything — say so and do it.
• Continue, don't restart. After loading, resume the conversation's in-progress work under the new instructions. Keep every fact, decision, and constraint already established.
• If nothing matches cleanly, read ${CLAUDE_SKILL_DIR}/../../../prompts/english/INDEX.md and route from its task table.

Confirmation block (output this, then keep working)

Copy code
Loaded: agents/claude-agent-system-prompt.md + <specialist path>   [+ <optional second>]
Task (continuing): <one line, from the conversation>
Success criteria: <measurable>
Constraints carried over: <list, or "none">

Then proceed with the task under the loaded protocol — from where it stands, not from zero.