Prompt

Turn raw user input into a well-engineered prompt, review it, then plan and execute.

Critical Rules

• •NEVER execute without explicit "Do it" from the user. The review step is mandatory. Do not skip it.
• •NEVER repeat or summarize what you just did. Keep the conversation flowing forward. The user saw your output — don't recap it. After completing a step, move to the next step or wait for input. No "Here's what was created" or "Done. Here's what I built" summaries.
• •Stay in flow. Each message should either present new information or ask for a decision. Nothing else.

Workflow

1. Clean & Structure

Take the user's raw input and return:

• •Goal: One sentence at the top summarizing intent
• •Body: Their ideas restructured for scannability. Reformatted, not rewritten. Do not add complexity or assumptions. This is their thinking, clarified.

Do not repeat the full prompt later in conversation — the user already read it. If they need to see it again, that's what "What did I say?" is for.

2. Review

Present via AskUserQuestion with two options:

• •Do it — proceed to iteration and execution
• •What did I say? — re-display the full cleaned-up prompt (user's ideas only, no Claude additions)

Stop here and wait. Do not proceed to step 3 until the user explicitly selects "Do it". Any other input means they want to revise — incorporate their feedback and re-present.

3. Iterate & Warn

Only after "Do it". Iterate on the prompt one more time using conversation context. This version is yours — add analysis the user didn't surface.

Warn the user if their suggestions indicate:

• •Architectural changes that could destabilize the app
• •Design hacks or workarounds
• •Patterns that introduce unnecessary complexity

Surface warnings before executing so the user can reconsider.

4. Plan & Execute

Enter plan mode (EnterPlanMode) with the final iterated prompt. Explore the codebase, design the implementation, and present the plan for user approval before making changes.

Execute in the main conversation — no subagent.

5. Save Prompt File

Create workflows/prompts/<unix-timestamp>.md:

---
status: executed
goal: "<one sentence>"
author: "<user>"
project: "<project name>"
tags: [<framework>, <domain>, <theme>]
issue: ""                                    # optional — GitHub issue number (e.g. 12) or URL
conversation_id: "<claude code conversation id if available>"
commit_hash: "<filled at commit time>"
created: "<ISO 8601 timestamp>"
---

# <Goal as title>

## Human Prompt

<The cleaned-up prompt the human approved before hitting "do it">

## Plan

<The iterated prompt Claude built from the human's input — includes analysis, architecture considerations, and implementation approach>

## Autonomous Considerations

<Decisions Claude made without explicit human direction during planning/execution. What was inferred, assumed, or chosen independently. Reviewers should verify these align with intent.>

## Flagged Issues & Resolutions

<Issues Claude surfaced during iteration/planning and how the developer chose to handle them. Empty if none were flagged.>

## Result

<Filled at commit time: outcome, different approach taken, discarded, etc.>

Prompt File Indexing

Frontmatter fields are the index. Searchable by:

• •author, project, tags, issue, conversation_id, commit_hash, status

Use Grep across workflows/prompts/ frontmatter to query the library.

Prompt Design Philosophy

When iterating on the agent prompt:

• •Show architecture, models, and key components — not implementation code
• •Build from first principles based on current app state
• •Optimize for robustness and stability
• •No unnecessary complexity
• •Every change should strengthen the foundation