Create Meta-Prompt

Create prompts for complex tasks that benefit from research before implementation.

Task: $ARGUMENTS

If $ARGUMENTS is empty or vague, ask the user: "What task needs exploration? Describe what you want to accomplish." Do not proceed until you have a clear task description.

Use this when:

• •Task needs exploration to understand scope
• •Multiple approaches are possible
• •You need to understand existing code patterns first

If a dedicated slash command exists for the task type (e.g., /ddd-analysis, /extract-requirements), use that instead of creating a meta-prompt. See /pipeline for the complete list of pipeline commands.

Workflow: Research → Plan → Implement (using plan directly)

Steps

1. •

   Read project context

   • •CLAUDE.md (required if exists)
   • •README.md
2. •

   Check for existing workflows

   bash

   ls -d .prompts/*-research .prompts/*-plan 2>/dev/null
   

   If any exist, list them and ask: "Continue one of these, or start new?" When checking, also consider whether the user's current task description might match an existing workflow under a different slug name.

   • •If continuing: use that slug and proceed to step 4 — the decision table will determine which phase to generate next based on the existing artifacts (research.md, plan.md) in that workflow
   • •If new: derive slug from task. If the derived slug matches an existing workflow directory, append a numeric suffix (e.g., refactor-payment-system-2) or ask the user for a different name.
3. •

   Determine slug (if starting new)

   • •Lowercase, hyphenate, drop stopwords (a, an, the, of, for, to)
   • •Example: "Refactor the payment system" → refactor-payment-system
4. •

   Determine which phase to generate

   Condition	Action
   No research.md	Generate research prompt
   research.md exists, Status is partial or failed	Ask user: re-run research or proceed with partial findings? If research.md contains a metadata block but no substantive findings (all questions unanswered or answered with "Could not determine"), recommend re-running with refined questions.
   research.md exists but is empty or has no substantive content	Treat as if research.md does not exist — re-generate research prompt
   research.md exists, Status is success (or no metadata) AND has substantive content, no plan.md	Generate plan prompt
   plan.md exists, Status is partial or failed	Ask user: re-generate plan or implement with current plan?
   plan.md exists, Status is success (or no metadata)	No prompt needed — tell user to implement with /run-prompt {slug}-plan

5. •

   Generate the prompt — create the output directory (and .prompts/ parent if needed) using mkdir -p.

Research Prompt

.prompts/{slug}-research/prompt.md:

# Research: [task]

## Objective
Understand [what needs to be understood] before implementation.

## Context
- Guidelines: `CLAUDE.md`
- Key files: [relevant paths]
- Stack: [if relevant]

## Questions to Answer
1. [Specific question about the codebase]
2. [Specific question about approach]
3. [Specific question about existing patterns]
4. [Specific question about constraints or risks]

## Explore
- [Key files/directories to examine]
- [What to search for]

## Output
Write findings to `.prompts/{slug}-research/research.md`:
- Answers to the questions above
- Existing patterns to follow
- Risks or concerns
- Recommended approach
- Metadata block (append at end):
  ## Metadata
  ### Status
  [success | partial | failed]
  ### Dependencies
  - [files or decisions this relies on, or "None"]
  ### Open Questions
  - [unresolved issues, or "None"]
  ### Assumptions
  - [what was assumed, or "None"]

Plan Prompt

.prompts/{slug}-plan/prompt.md:

# Plan: [task]

## Context
- Research: `.prompts/{slug}-research/research.md`
- Guidelines: `CLAUDE.md`

## Instructions
1. Read research.md
2. Design implementation as checkpoints
3. Each checkpoint must include:
   - Build: what to implement
   - Test: what unit tests to write for THIS checkpoint's code
   - Verify: how to confirm all existing + new tests pass before moving on
4. NEVER design a dedicated "testing" checkpoint at the end. Tests are written alongside the code they verify, within the same checkpoint. Each checkpoint must leave the test suite green.

## Output
Write plan to `.prompts/{slug}-plan/plan.md`:
- Ordered checkpoints (implementation + tests each — no checkpoint without tests unless it is purely non-code work like documentation or configuration)
- Files to create/modify
- Metadata block (Status, Dependencies, Open Questions, Assumptions)

After Plan is Complete

No separate "do" prompt is needed. The plan.md contains the checkpoints. User implements by:

• •Running /run-prompt {slug}-plan (executes the plan)
• •Or asking directly: "Implement the plan in .prompts/{slug}-plan/plan.md"

Progress will be tracked in .prompts/{slug}-plan/progress.md during execution. If interrupted, running the same command again will resume from the last incomplete checkpoint.

After Saving

Report what was created and what's next:

• •After research prompt: "Created research prompt. Run with /run-prompt {slug}-research, then run /create-meta-prompt [same task] again to generate the plan phase."
• •After plan prompt: "Created plan prompt. Run /run-prompt {slug}-plan to generate the plan. Once the plan is generated, run the same command again to implement it checkpoint by checkpoint."