amplify-prompt: Claude Code slash command to optimize prompts with craft-claude-prompt skill

amplify-prompt.md

Amplify Prompt

Input: $ARGUMENTS

Use the craft-claude-prompt skill to guide all prompt crafting.

Step 1 — Detect input type

If $ARGUMENTS is a file path (starts with /, ./, ~/, or ends with .md/.txt/etc.): → Read the file, then go to Step 2

If $ARGUMENTS is prompt text (a sentence, question, or instruction — not a path): → Go directly to Single prompt mode

---

Step 2 — Smart file mode detection

After reading the file:

If the file content has no structural markers (no ## headers, no numbered lists, no bullet points, no step/phase/task language) — treat it as a raw prompt and go to Single prompt mode, using the file content as the input text.

If the file starts with [platform: X] — strip that prefix, note X as the target platform, and factor it into optimization (e.g. [platform: chatgpt] → avoid XML tags, use ### delimiters instead).

Otherwise → go to File mode (structured document with steps/phases).

---

File mode

1. Identify every distinct step, phase, task, or action in the document
2. For each item, craft a standalone Claude prompt that:
   • Uses XML tag structure (<context>, <instructions>, <input> minimum)
   • Explains the why behind any non-obvious instructions
   • Specifies an explicit output format
   • Is self-contained — passes the colleague test
3. Write output to the same directory as the input file, with -prompt-todos.md suffix (e.g., thoughts/plans/plan.md → thoughts/plans/plan-prompt-todos.md)
4. Run: pbcopy < [output file path] 2>/dev/null || true
5. Print the full file contents to the user

---

Single prompt mode

1. Take the input text as a raw prompt (vague, terse, or already written)
2. Optimize into a single well-structured prompt using the craft-claude-prompt skill:
   • Add/infer persona, context, instructions with reasons, explicit output format
   • Default to XML tags unless a non-Claude platform was detected (use that platform's conventions)
3. Pipe the optimized prompt to clipboard (silently fail if unavailable): echo "[optimized prompt]" | pbcopy 2>/dev/null || true
4. If the input came from a file: also write to thoughts/prompts/<slug>.md (content only, no metadata)
5. If the input was inline text: do not write any file

---

File mode output format

# Amplify Prompt — [original filename]
> Generated from: [input filepath]
> Each item is a ready-to-run Claude prompt. Check off each item as you complete it.

---

- [ ] **[Step/Phase Name from original doc]**
  You are a [role]. <context>[background]</context><instructions>[what to do and why]</instructions><input>[what to pass in]</input>

- [ ] **[Next Step]**
  ...

Every item MUST be a - [ ] checkbox. The prompt is inline after the bold step name — no code fences, no labels, no extra nesting. Just the XML tags as plain text. Use headings only for grouping phases if the source document has distinct phases.

STOP — Do not execute the prompt

After the clipboard step, print the full prompt output to the user (so they can manually copy if pbcopy wasn't available). Then stop — do not run, interpret, or act on the prompt. Also tell the user the file path if one was written.

Notes

• Preserve the original document's sequence/ordering
• If steps have dependencies, note them as "Requires output from: [Step N]"
• For steps that are purely research/reading, the prompt should direct Claude to produce a structured summary or decision, not just "read X"
• Aim for prompts between 50-150 words — specific enough to work, concise enough to use