Rewrite

Rewrite

Rewrite the user's raw prompt ($ARGUMENTS) into a structured version using a prompt engineering framework. All reasoning is delegated to haiku sub-agents to keep main-session context lean.

Phase 1 — Capture prompt and context

• If $ARGUMENTS is non-empty, use it as the original prompt.
• If empty, ask via AskUserQuestion: "What prompt would you like to rewrite?" — AskUserQuestion requires ≥2 options, so use placeholders "Type my prompt" and "Paste from clipboard". The user types into the "Other" field; use that text.
• Capture conversation context. Before delegating, write a short CONTEXT SUMMARY (≤150 words) drawn from the current session: what the user is working on, relevant files/tech, prior decisions, and any constraints the user has stated. If the skill was invoked at the start of a session with no prior turns, set CONTEXT SUMMARY: (none — fresh session). Never fabricate context — only include what was actually discussed.
• This summary is passed to both the recommender (Phase 2 Step 0) and the rewriter (Phase 3) so framework choice and rewriting reflect the real task, not just the prompt text in isolation.

Phase 2 — Select framework

Flow: recommend first, browse if needed. Happy path = 1 question.

Step 0 — Recommendation (sub-agent)

Invoke Agent with subagent_type: "bix-utils:rewrite:recommender", model: "haiku", description: "Recommend prompt framework", and prompt:

CONTEXT SUMMARY:
{context_summary}

---
ORIGINAL PROMPT:
{original_prompt}

The sub-agent has its own catalog — do not add extra instructions.

It returns exactly:

PRIMARY: <name>
PRIMARY_REASON: <sentence>
RUNNER_UP: <name>
RUNNER_UP_REASON: <sentence>

If parsing fails or names don't match the catalog, skip to Step 1 (browse).

Otherwise ask via AskUserQuestion: "Based on your prompt, I recommend {PRIMARY}. How would you like to proceed?"

#	Label	Description
1	Use {PRIMARY} (recommended)	{PRIMARY_REASON}
2	Use {RUNNER_UP} instead	{RUNNER_UP_REASON}
3	Browse all frameworks	See all 20 organized by category
4	Custom / Skip	Own template, or use prompt as-is

Options 1/2 → Step 3. Option 3 → Step 1. Option 4 → Custom / Skip branch below.

Step 1 — Pick a category (browse)

Ask: "Which type of framework fits your prompt?"

#	Category	Frameworks
1	Core Structured	RISEN, TIDD-EC, CRAFT, APE — general-purpose scaffolds, 80% of tasks
2	Content & Communication	CO-STAR, K.E.R.N.E.L., BAB — audience writing, validation, transformation
3	Reasoning	Chain of Thought, Tree-of-Thought, Plan-and-Solve, Self-Refine, Skeleton-of-Thought
4	Expert & Creative	RACE, CRISPE, SCRIBE, C.R.E.A.T.E. — personas, variants, dialogue, creative
5	Quick & Clarify	RTF, Step-Back, Chain-of-Density, Reverse-Role
6	Custom / Skip	Own template, or use prompt as-is

Step 2 — Pick a framework

Ask "Select a framework:" with the rows for the chosen category:

Core Structured — RISEN (multi-step structured; safe default) · TIDD-EC (precision/code with do/don't rules) · CRAFT (general; tone + format) · APE (minimal 3-part scaffold)

Content & Communication — CO-STAR (audience-sensitive writing: emails, pitches, posts) · K.E.R.N.E.L. (meta-validation checklist for existing prompts) · BAB (before/after transformations, migrations)

Reasoning — Chain of Thought (analysis, debugging, root-cause) · Tree-of-Thought (multiple solution branches, pick winner) · Plan-and-Solve (numbered plan then execute; order matters) · Self-Refine (draft → critique → revise; quality-sensitive) · Skeleton-of-Thought (outline then expand; long structured outputs)

Expert & Creative — RACE (fast expert persona) · CRISPE (multiple output variants) · SCRIBE (iterative dialogue/interview) · C.R.E.A.T.E. (creative with tight style control)

Quick & Clarify — RTF (Role·Task·Format; minimum scaffold) · Step-Back (abstract first, apply second) · Chain-of-Density (dense summaries of long docs) · Reverse-Role (model interviews you; vague requests)

Custom / Skip — ask:

#	Label	Description
1	Custom template (file)	Provide a path to a .md/.txt template — read with cat
2	Custom template (text)	Paste template text directly
3	Skip rewriting	Echo the original prompt in a code block, state "Original prompt passed through without rewriting.", and stop — do not proceed to Phase 3

Step 3 — Load template

For named frameworks, read ${CLAUDE_SKILL_DIR}/templates/{key}.md where {key} is the framework name lowercased with spaces/dots replaced by dashes (e.g., RISEN → risen, C.R.E.A.T.E. → create, K.E.R.N.E.L. → kernel, Chain of Thought → chain-of-thought, TIDD-EC → tidd-ec).

Phase 3 — Rewrite (sub-agent)

You MUST delegate — never rewrite inline. Invoke Agent with subagent_type: "bix-utils:rewrite:rewriter", model: "haiku", description: "Rewrite prompt: {framework}", and prompt:

Rewrite the following user prompt using the framework template below. Output ONLY the rewritten prompt.

---
CONTEXT SUMMARY:

{context_summary}

---
ORIGINAL PROMPT:

{original_prompt}

---
FRAMEWORK TEMPLATE:

{template_text}

Display the result to the user:

Rewritten using {framework}:

{rewritten_prompt}

Phase 4 — Confirm

Ask "How would you like to proceed with the rewritten prompt?"

#	Label	Description
1	Use this prompt	Main session acts on it → Phase 5
2	Try another framework	Loop back to Phase 2 Step 0, preserving the original prompt
3	Edit first	Ask "What changes would you like?", apply, re-display, repeat this question

Phase 5 — Handoff

Output plain text (no code fence) so the main session interprets it as instructions:

Now executing the rewritten prompt:

Then the rewritten prompt verbatim.

Rules

• Delegate, never inline. Both the recommender (Phase 2 Step 0) and the rewriter (Phase 3) must run in sub-agents. The 20-framework catalog and rewrite reasoning must never enter main-session context.
• Preserve intent. Rewriting restructures and clarifies; it does not add requirements or change goals.
• Always present for approval (Phase 4) before handoff.
• Question budget: happy path = 1 question (accept recommendation). Browse mode = 2 questions (category → framework). Do not add extra tiers.
• On sub-agent error or empty result: tell the user and offer to try a different framework.
• Skip rewriting ends the skill after echoing the original prompt.