Rp Oracle Export V2

ChatGPT Prompt Export

Raw request: $ARGUMENTS

Your job: select the right files and export a prompt file that another model can act on directly.

Before you do anything else, extract the real task from the raw request above. Users often phrase this as "export a prompt for X" or "write a prompt about Y" — strip away any meta-framing about exporting/prompting and identify the underlying problem. For example:

• "export a prompt to evaluate the auth refresh logic" → the task is "evaluate the auth refresh logic"
• "write a ChatGPT prompt about the token caching bug" → the task is "investigate the token caching bug"
• "review the last 3 commits" → the task is already clean

Use the extracted task (not the raw request) for all downstream steps — intent classification, context_builder instructions, and the final exported prompt.

Rules

• Infer Question / Plan / Review when obvious. Ask only if unclear.
• For vague requests, use repo evidence before asking questions.
• Use the fast path only when the scope is already small, concrete, and obviously file-local.
• For broad Question/Plan exports, context_builder is the default path.
• For review exports, context_builder is the default path.
• Do not spend exploratory tool calls proving that a broad request is complex enough for context_builder.
• When you do use context_builder here, keep response_type: "clarify".
• If you used the fast path, review the selection and prompt text before exporting.
• If you used context_builder, trust its curated selection, budget, and generated prompt by default; only re-check or adjust prompt/selection/tokens if you noticed a concrete issue.
• Export to a unique repo-local file, usually in prompt-exports/.
• Derive a short slug from the user's request and use it in the filename.
• Use a relative repo-local path by default; do not use an absolute path or another folder unless the user explicitly asks for it.

Workflow

0: Workspace Verification (REQUIRED)

Before any building context, bind to the target codebase using its working directory:

{"tool":"bind_context","args":{"op":"bind","working_dirs":["/absolute/path/to/project"]}}

This auto-resolves to the window containing your project. No need to list windows first.

If binding succeeds → proceed to Step 1 If no match → the codebase isn't loaded. Find and open the workspace:

{"tool":"manage_workspaces","args":{"action":"list"}}
{"tool":"manage_workspaces","args":{"action":"switch","workspace":"<workspace_name>","open_in_new_window":true}}

Then retry the working_dirs bind.

1. Determine intent and scope

Infer the prompt type from the request:

• Review for git diff / PR / branch comparison requests — i.e. the user wants to inspect changes
• Plan for design / approach / implementation-plan / architectural audit / code evaluation requests — even if the user says "review" or "audit", if there are no diffs involved, this is a Plan
• Question only when the user is asking a specific, bounded question with a clear answer
• When in doubt, default to Plan. Generic or open-ended requests ("look into X", "help me with Y", "figure out Z") produce better results with the Plan preset, which gives the receiving model structured guidance.

If the request is vague:

• for Review: inspect git state first
• for Question/Plan: if it sounds broad, architectural, evaluative, redesign-oriented, or likely multi-file, skip manual exploration and go straight to context_builder

Ask one specific question only if needed, and base it on the repo state you found. Good question shapes:

• “I see changes in A and B. Do you want review of these current uncommitted changes, or against main?”
• “I found likely touchpoints in X and Y. Is the fix plan for X only, or this broader flow?”

If the scope is still unclear, STOP and ask the user. Do not ask generic workflow questions when you could ask a concrete scope question instead.

2. Choose context path

Because this prompt does not expose the workflow export budget directly, prefer context_builder unless the review scope is obviously tiny.

Review

Start by checking git state:

{"tool":"git","args":{"op":"status"}}
{"tool":"git","args":{"op":"diff","detail":"files"}}

Review Scope Confirmation

Determine the comparison scope from the user's request and git state.

If the user already specified a clear comparison target (e.g., "review against main", "compare with develop", "review last 3 commits"), skip confirmation and proceed using the scope they specified.

If the scope is ambiguous or not specified, ask the user to clarify:

• Current branch: What branch are you on? (from git status)
• Comparison target: What should changes be compared against?
  • uncommitted – All uncommitted changes vs HEAD (default)
  • staged – Only staged changes vs HEAD
  • back:N – Last N commits
  • main or master – Compare current branch against trunk
  • <branch_name> – Compare against specific branch

Example prompt to user (only if scope is unclear):

"You're on branch feature/xyz. What should I compare against?

• uncommitted (default) - review all uncommitted changes
• main - review all changes on this branch vs main
• Other branch name?"

If you need to ask, STOP and wait for user confirmation before proceeding.

This prompt does not expose the workflow export-mode budget directly. Lean on context_builder unless the uncommitted review scope is clearly tiny, obviously bounded, and safe to include in full. For Review, the fast path is the exception, not the default. It is allowed only when the confirmed scope is uncommitted changes and the full changed-file review scope is obviously tiny and safe to include in full. Otherwise require context_builder. For Review, this is the default path. If the review is not a tiny uncommitted-change export that is obviously safe to include in full, context_builder is required. For review exports, explicitly reference the diff / changed files in the context you build.

Always include the phrase "code review" in your context_builder instructions for Review exports. This phrase activates diff analysis in the discovery agent. Without it, the builder treats the request as a general exploration.

Question / Plan

Default to context_builder for any request that is broad, architectural, evaluative, redesign-oriented, or likely to touch multiple files.

Do not spend tool calls proving that these requests are complex. If the user is asking you to evaluate logic, assess a design, rethink a flow, or reason about behavior across a system, call context_builder immediately.

Use the fast path only when the request is already small and obvious:

{"tool":"file_search","args":{"pattern":"<key term>","mode":"both"}}

{"tool":"manage_selection","args":{"op":"add","paths":["RootName/path/to/FileA.swift","RootName/path/to/FileB.swift"]}}

If there is any real doubt that the fast path will fully cover the task, use context_builder.

Otherwise use context_builder:

{"tool":"context_builder","args":{
  "instructions":"<task>The actual problem to solve — not about exporting or prompting</task>\n<context>Scope: <what you found>.</context>",
  "response_type":"clarify"
}}

{"tool":"context_builder","args":{
	"instructions":"<task>Code review of changes against <confirmed_scope>.</task>\n<context>Intent: code review. Branch: <branch_name>.</context>",
  "response_type":"clarify"
}}

3. Final check (fast path only — skip after context_builder)

If you used context_builder, skip this step entirely and go straight to Step 4. The builder already curated the selection, managed the token budget, and wrote the prompt. Do not read the prompt back, do not inspect the selection, do not check token counts, and do not critique, rewrite, or "improve" the generated prompt text. Treat the builder's output as the final payload for export.

If you used the fast path, check the selection and prompt text before exporting:

{"tool":"manage_selection","args":{"op":"get","view":"summary"}}

{"tool":"prompt","args":{"op":"get"}}

If available in this surface, the fast path may also inspect token state:

{"tool":"workspace_context","args":{"include":["selection","tokens"]}}

If the prompt wording or selection is off, fix it before exporting.

4. Export

Use a unique repo-local relative path such as:

• prompt-exports/<yyyy-mm-dd>-<hhmmss>-question-<slug-from-request>.md
• prompt-exports/<yyyy-mm-dd>-<hhmmss>-plan-<slug-from-request>.md
• prompt-exports/<yyyy-mm-dd>-<hhmmss>-review-<slug-from-request>.md

Choose <slug-from-request> by summarizing the user's request into a short filesystem-safe phrase. Prefer descriptive slugs like collapsing-tool-logic or agent-transcript-redesign, not generic names like export or question.

Unless the user explicitly asks for another destination, keep the export path relative and repo-local under prompt-exports/.

Preset mapping:

• Question → standard (only for specific, bounded questions)
• Plan → plan (default for generic, open-ended, or ambiguous requests)
• Review → codeReview

{"tool":"prompt","args":{"op":"export","path":"prompt-exports/<unique filename>.md","copy_preset":"<standard|plan|codeReview>"}}

Anti-patterns

• Asking generic workflow questions before checking repo state
• Skipping context_builder for branch / PR / large review exports
• Doing exploratory searches or file reads before context_builder for a broad Question/Plan export just to prove the task is complex
• Treating requests like "evaluate this logic", "assess this design", or "rethink this flow" as fast-path exports
• Using the fast path when scope is still vague
• Exporting from the fast path without checking the selection and prompt text
• Re-checking selection, prompt text, or tokens after context_builder — the builder already finalized the payload
• Reading the prompt back after context_builder to review, critique, rewrite, or "improve" it — export it as-is
• Calling prompt get, manage_selection get, or workspace_context after context_builder completed — go straight to export
• Reusing generic filenames like oracle-prompt.md by default
• Using generic slugs like export, question, or plan when the request gives you enough detail for a better filename
• Writing to an absolute path or outside the repo by default when the user did not ask for that
• Passing export/prompt meta-framing to context_builder — instructions like "export a prompt for X" or "build context for a ChatGPT prompt about Y" cause the builder to write a prompt about prompting instead of a prompt that solves X. Always pass the extracted task directly.

Report the final export path, prompt type, whether you used the fast path or context_builder, and token count if available.