WED, SEPTEMBER 23, 2026
Independent · In‑Depth · Practitioner‑Tested
ChatGPT Productivity

GPT-6 Luna Prompts for High-Volume Work at $0.10 per Million Tokens

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output. It scores 66.6% on DeepSWE v1.1, only 2.2 points behind Sol, but drops to 20.7% on AutomationBench against Sol's 33.2% and carries a 7.6% factual error rate. The line is clean: Luna executes well and plans badly. These seven prompts are built to stay on the right side of that line - fully specified tasks, no decisions left to the model, and a cheap self-check that costs less than a single Sol call.

⌨️ 7 prompts 🕐 Updated Sep 23, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Closed-Set Classification
Closed sets and an explicit escape hatch keep Luna inside the task profile it handles well.
Classify each item below into exactly one of these categories: [list your categories].<br/><br/>Rules:<br/>1. Use only the categories listed. Never invent a new one.<br/>2. If an item does not fit any category, output UNCLASSIFIED. Do not guess.<br/>3. Output one line per item in the format: item_id | category<br/>4. No explanation. No preamble. No summary at the end.<br/><br/>ITEMS:<br/>[paste]
2
Field Extraction With a Fixed Schema
Rule 2 is the important one. Luna's 7.6% error rate mostly shows up as confident invention.
Extract these exact fields from the text below: [list field names].<br/><br/>Output valid JSON only. No markdown fences. No commentary.<br/><br/>Rules:<br/>1. Every field in the list must appear in the output.<br/>2. If a value is not present in the text, use null. Never infer, estimate or fill from general knowledge.<br/>3. Copy values verbatim from the source. Do not reformat dates, numbers or names.<br/>4. Add no fields beyond the ones listed.<br/><br/>TEXT:<br/>[paste]
3
Two-Pass Self Check
Two Luna passes cost a fraction of one Sol call and catch a large share of its error rate.
You will do this in two passes.<br/><br/>PASS 1: Complete the task below.<br/><br/>PASS 2: Re-read the source material and your own Pass 1 output. List every claim in your output that you cannot point to a specific line in the source for. For each one, either correct it or replace it with null.<br/><br/>Output both passes, labelled.<br/><br/>TASK:<br/>[paste]<br/><br/>SOURCE:<br/>[paste]
4
Bulk Rewrite to a Fixed Format
Rules 2 and 3 make the output mechanically checkable, which matters at volume.
Rewrite each item below to match this exact format: [describe format, or paste one example].<br/><br/>Rules:<br/>1. Change wording and structure only. Never add facts not present in the original.<br/>2. Never remove a fact that is present in the original.<br/>3. Keep the same number of output items as input items, in the same order.<br/>4. Output one item per line, separated by a blank line. Nothing else.<br/><br/>ITEMS:<br/>[paste]
5
First-Pass Code With an Explicit Contract
Luna scores 66.6% on DeepSWE - strong for a first draft when every case is stated up front.
Write an implementation for exactly this signature. Do not change the signature.<br/><br/>SIGNATURE:<br/>[paste]<br/><br/>Requirements:<br/>1. Handle these cases: [list them explicitly - do not leave any to inference]<br/>2. Use only the standard library. Add no dependencies.<br/>3. No comments. No docstring. Code only.<br/>4. If any requirement is ambiguous, stop and list the ambiguity instead of guessing.<br/><br/>This is a first draft that will be reviewed. Optimise for correctness on the listed cases, not for elegance.
6
Disagreement Check Between Two Runs
The second half of the run-it-twice pattern. Only the DISAGREE rows need a stronger model.
Below are two independent answers to the same question, produced by the same model on separate runs.<br/><br/>Compare them field by field.<br/><br/>Output three sections:<br/>AGREE - values identical in both<br/>DISAGREE - values that differ, showing both<br/>MISSING - present in one but not the other<br/><br/>Do not decide which is correct. Only report the differences.<br/><br/>RUN A:<br/>[paste]<br/><br/>RUN B:<br/>[paste]
7
Escalation Triage
A $0.0001 router. Luna decides whether Luna should handle it, and sends the rest to Sol.
Read the task below and answer one question only: can this be completed in a single pass with no tool calls and no decisions about what to do next?<br/><br/>Answer SINGLE_PASS or NEEDS_PLANNING. One word. Nothing else.<br/><br/>Answer NEEDS_PLANNING if the task requires any of: calling a tool and reacting to its output, choosing between approaches, breaking the work into steps, or looking something up.<br/><br/>TASK:<br/>[paste]