Chat To Code Prompt

chat-to-code-prompt

Purpose

Bridge the gap between a Claude.ai chat and Claude Code. Take whatever has been discussed, decided, or scoped in the current chat and convert it into a clean, self-contained prompt that a fresh Claude Code session can execute with zero additional context from the user.

The output is always a single fenced markdown block, ready to paste into a new Claude Code session.

Activation

Triggered by any of:

• "chat-to-code-prompt"
• "send to Claude Code" / "send this to CC"
• "make a CC prompt" / "Claude Code prompt for this"
• "package this for Claude Code"
• "handoff to Claude Code" / "handoff to CC"
• "convert this to a Claude Code prompt"

Execution

Step 1 — Pick the format

Call ask_user_input_v0 with a single question. Do not infer or skip — always ask.

Question: "What kind of prompt am I packaging?" Options (single_select):

• "Task — build/fix/refactor"
• "Plan — implement this design"
• "Context — continue this work"

Step 2 — Extract chat content

Pull only what's relevant to the chosen format:

• Task: action verb + deliverable, repo/branch, files involved, constraints, success criteria, notes
• Plan: plan name, repo, overview, phases, decisions already made, out of scope, definition of done
• Context: topic, repo, background, what's been done, current state, open questions, suggested next move

Do not invent files, paths, or facts. If the chat lacks a piece (e.g., no repo named), either omit the field, mark it <TBD by Ryan>, or use the most-recently-mentioned repo and flag the assumption above the code block.

Step 3 — Apply Claude Code framing

• State the working directory / repo at the top of the prompt
• Reference any CLAUDE.md if it was mentioned in the chat
• Front-load the deliverable
• Use action verbs (Build, Fix, Refactor, Migrate, Implement, Continue)
• List constraints explicitly — CC will execute, not clarify
• Mention any tool conventions (e.g., git format-patch + git am fallback if push fails in the container)
• Strip all Claude.ai chat artifacts: no "as we discussed", "in our chat", "I mentioned earlier"
• Make the prompt 100% self-contained — assume the receiving CC session has zero memory of this chat

Step 4 — Output

Above the code block, write 2–3 lines max:

• Format mode chosen
• One-line "paste into a fresh CC session in <repo>" instruction
• Any flagged assumptions

Then output the full prompt as a single fenced markdown block so it's tap-and-hold copyable on mobile.

After the block, optionally offer one follow-up: save to file, scaffold the target repo with gh-repo-create, or chain into architect-plan-for-dispatch if the scope is large.

Templates

Task

# <Action verb + deliverable>

**Repo:** <repo>
**Branch:** <branch or "current">

## Context
<1–3 sentences of background — what and why>

## Goal
<one clear sentence>

## Files involved
- `path/to/file` — <what changes>
- ...

## Constraints
- <don't touch X>
- <follow CLAUDE.md / convention Y>

## Success criteria
- [ ] <observable outcome>
- [ ] <test passes>

## Notes
<anything else CC needs cold>

Plan

# Implement: <plan name>

**Repo:** <repo>
**Reference:** <CLAUDE.md / spec / link>

## Overview
<2–4 sentences>

## Phases
1. <Phase 1> — <what + why>
2. <Phase 2> — ...

## Decisions already made
- <decision> — <rationale>

## Out of scope
- <thing not to do>

## Definition of done
- [ ] <outcome>
- [ ] <outcome>

Context handoff

# Continue: <topic>

**Repo:** <repo or N/A>

## Background
<3–5 sentences on what was being worked on and why>

## What's been done
- ...

## Current state
- ...

## Open questions
- ...

## Suggested next move
<1–2 sentences>

Rules

• One fenced code block. Always at the end of the response.
• Self-contained. Never reference "the chat" or "earlier" inside the prompt.
• Never invent. No fabricated files, paths, repos, or facts. Mark unknowns <TBD by Ryan>.
• Push back on empty chats. If there's nothing usable, ask what to package — don't generate filler.
• Flag assumptions. Default to the most-recently-mentioned repo only when obvious; call it out above the block.
• No emojis. Markdown headings, bullets, and checkboxes only.

Chaining

• Upstream: Any chat — planning, debugging, scoping, design.
• Downstream: Pasted into Claude Code.
• Adjacent: prompt-architect for general-purpose Claude prompts; this skill is CC-specific and assumes an executing agent.
• Composes with: For very large multi-task bodies of work, prefer architect-plan-for-dispatch instead — it produces a structured execution file rather than a single prompt.

Example

User: We've been scoping the DTST migration — let's pack this for CC.

Claude: [calls ask_user_input_v0 with the format question]

User: Plan — implement this design

Claude: Mode: Plan. Paste into a fresh Claude Code session opened in ~/code/dtst (ryankolean/dtst). Assumed repo from chat — flag if wrong.

# Implement: DTST migration from Tauri to local-first web app

**Repo:** ryankolean/dtst
**Reference:** CLAUDE.md, MIGRATION-EXECUTE.md (both at repo root)

## Overview
Migrate DTST from Tauri to a pure local-first web app. Stack: React 18 + Vite + TS + Tailwind + Zustand + TanStack Table + Recharts (frontend), Node/Express (backend). Backups live at `~/Library/Application Support/com.summitsoftware.dtst/backups/` and must remain readable.

## Phases
1. Scaffold Vite + React 18 + TS frontend — replace Tauri shell
2. Stand up Node/Express backend — local-only, no auth
3. ...

## Decisions already made
- Local-first only — no cloud sync in v1
- Keep existing JSON schema for backups — no migration needed

## Out of scope
- Multi-user
- Cloud deployment

## Definition of done
- [ ] `npm run dev` boots frontend + backend
- [ ] All 13 Dispatch tasks complete and committed
- [ ] Existing backups load cleanly

Want me to save this to a file or chain into architect-plan-for-dispatch?