Turn NASA’s Apollo 13 Transcript Into 5 Slides Without Fake Quotes

Across four runs, quote locking raised usable slides from 1/5 to 4/5. The source audit still caught a false quote.A faithful summary can easily become a false quote.At 30 seconds, ChatGPT put one word in Mission Control’s mouth: “GO.” It wasn’t in NASA’s transcript that way.Clean copy.…

Across four runs, quote locking raised usable slides from 1/5 to 4/5. The source audit still caught a false quote.A faithful summary can easily become a false quote.At 30 seconds, ChatGPT put one word in Mission Control’s mouth: “GO.” It wasn’t in NASA’s transcript that way.Clean copy. Wrong quote.NASA’s Apollo 13 transcript actually says, 13, Houston. GO at 30 seconds. The output shortened the call and kept the quotation marks. A faithful summary had become a false quote.I ran the same excerpt through four fresh ChatGPT Free chats. Stricter quote rules made four of five slides usable, but they did not replace the source check.A faithful summary had become a false quote.One source, one narrow sceneThe source boundary was ten seconds before liftoff through the four-minute mark.I used four viewer pages from NASA’s Apollo 13 command-module transcript, reached through NASA’s mission transcript index. The excerpt includes the countdown, tower clearance, early trajectory calls, staging, and a few brief crew observations.NASA transcript page showing liftoff and early trajectory calls.That narrow boundary mattered. I didn’t let general Apollo knowledge fill gaps in the copy. If a detail wasn’t inside the excerpt, it couldn’t support the carousel.The fixed five-slide briefThe production brief was also fixed before the runs:Set the launch scene without a required quote.Cover countdown and liftoff with no more than one exact quote.Give one verifiable early-trajectory fact.Cover staging or a crew observation with no more than one exact quote.End with an author takeaway, clearly separated from astronaut speech.Every slide had to stay under 55 words. Quotations, paraphrases, facts, and interpretation needed visible labels.The four-run setupI pasted the same excerpt into four fresh chats on the first-party ChatGPT website. The live account showed the Free plan.The interface didn’t display a model name, so I’m not assigning one after the fact. No paid feature, upgrade, or billing flow was used. The total spend was $0.Runs A1 and A2 used the same baseline prompt. It asked for a punchy five-slide carousel with compelling quotations but gave no tracing rules.Runs B1 and B2 used the same quote-locked prompt. It required every direct quotation to match the excerpt verbatim, include a speaker and mission timestamp, and label supported paraphrases. It also told the model to omit a quote instead of inventing or repairing one.Here’s the compact result:https://medium.com/media/798d014e99218a3ff74f915bfaba30ce/hrefThose scores describe this excerpt and these four outputs. They aren’t a reliability rating for ChatGPT as a whole.What the baseline polished awayA1 produced energetic copy, but it treated source language as design material. Its second slide used both “LIFT-OFF.” and “GO at 30 seconds.” That broke my one-quote limit, and neither line included a speaker or timestamp.Another slide quoted “Look at the horizon out there.”, even though the speaker code was illegible in the scanned source.A2 went further. It turned the 30-second call into Mission Control calls: “GO.” It also wrote that Apollo 13 launched with “three astronauts aboard,” a true fact about the mission but not a fact supplied by the locked excerpt.On Slide 4, it placed “Beautiful.” after the horizon observation. In the transcript, the Beautiful. call occurs nine seconds before that line.ChatGPT baseline outputNone of these errors sounded absurd. That was the real failure pattern: compression, smoothing, and scene rearrangement in service of better copy.The baseline runs each left only one slide I could use without substantive repair. Both final takeaway slides were clearly presented as interpretation. The rest needed attribution, source anchors, deleted quotation marks, corrected chronology, or unsupported details removed.What the quote lock fixedThe locked prompt changed the shape of the answer immediately. B1 and B2 labeled quotations and paraphrases, added timestamps, and attached source ranges. Each produced four slides that were usable without substantive repair.B2, for example, returned this clean block:Quotation: “The clock is running....” — CDR, 00 00 00 02That wording, punctuation, speaker code, and timestamp all matched the excerpt. Its staging slide did the same with:“Got a little flash out the window.” — CMP, 00 00 02 45ChatGPT output with quotation labels, speaker codes, timestamps, and source ranges.The constraint also pushed broad scene summaries into labeled paraphrases instead of quotation marks. That made the editorial task faster because I could see what kind of claim I was checking before I checked it.But both quote-locked runs still mishandled the 30-second GO call. B1 wrote At 30 seconds, Houston reports: “GO.” B2 put quotation marks around GO inside a paraphrase. Both were recognizable reductions of the source, and neither was an exact quotation.The prompt improved traceability. It didn’t perform the final audit.The finished five slidesI used B2 as the closest structural base, then made six substantive changes: one pass on each slide and one global labeling pass.Slide 1: Launch in Ten SecondsParaphrase · source: -00 00 00 10–00 00 00 34The final countdown moves through ignition to liftoff. Apollo 13 clears the tower, completes yaw and roll, and begins pitching into the climb.Slide 2: The Clock StartsQuotation · CDR · 00 00 00 02“The clock is running….”Paraphrase · source: 00 00 00 01–00 00 00 16After liftoff, the crew calls the yaw and roll program while Houston confirms the roll.Slide 3: Right on TrajectoryVerifiable fact · source: 00 00 00 30–00 00 00 34At 30 seconds Houston reports GO; two seconds later the command-module pilot says the vehicle is right on trajectory. Roll completes and the pitch begins.Slide 4: Flash at StagingQuotation · CMP · 00 00 02 45“Got a little flash out the window.”Paraphrase · source: 00 00 02 42–00 00 02 52The crew is approaching four g before the flash. S-II ignition follows three seconds later, and the crew reports half a g.Slide 5: Keep the LabelsAuthor takeaway · interpretationA transcript can supply pace without supplying a clean narrative. Keep quotations exact and attributed, mark paraphrases, attach source lines to material facts, and reserve the final slide for interpretation.The human changes that matteredThe most important edit was removing quotation marks from GO and treating the moment as a sourced fact. That preserved the scene without pretending the shortened wording was speech.Then I replaced vague phrases with checkable events. “The launch sequence is underway” became yaw, roll, and Houston’s confirmation. The staging slide gained the calls immediately before and after the window flash, which kept the chronology intact.Finally, I rewrote the generic creator takeaway. The generated versions praised precision, pressure, and human perception. Those ideas were reasonable, but they did not help someone repeat the method. The final slide now explains the operating rule: exact quotes, labeled paraphrases, sourced facts, separate interpretation.The finished carousel contains two direct quotations. Both match the excerpt, both have legible speaker codes and timestamps, and neither slide contains a second quote. The final audit found zero untraceable quotations.A quote-lock checklist worth reusingBefore turning any public transcript into a social asset, I would repeat this sequence:Lock the exact source boundary before prompting.Keep every run in a fresh chat with the same input.Require verbatim quotes, speaker names or codes, and timestamps.Label paraphrases and attach source lines to material claims.Delete quotation marks when the model compresses or repairs speech.Audit chronology, not just wording.Keep author interpretation visibly separate from source speech.The key lesson from these four runs is small but useful. Quote instructions are a strong first filter. They aren’t proof. The publication-safe asset only appeared after the generated copy, the original lines, and the human edit ledger were checked together.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!Turn NASA’s Apollo 13 Transcript Into 5 Slides Without Fake Quotes was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Generative AI Pub — Published August 17, 2026 at 02:35 UTC — Category: Image AI