Ask questions of the archive

AI assistants don't read pages; they read tokens, small pieces of words. This archive is about 1.94M tokens, roughly 780 tokens a page.

• Every AI has a limit on how much it can read at once. The largest read about 1 million tokens, roughly 1,300 pages, and many free chats read far less. No AI reads this whole archive in one go.
• So the archive can't simply be pasted into a chat. It has to be loaded into a tool that keeps the files and searches them for each question.
• Searching for exact words can be complete. A name, a phrase, a street, a date: Claude can run a search over every file and list every match. Questions about meaning are answered from the passages the AI finds, which is a good answer but not a full read.
• The AI tells you what it read. The starter prompt in step 5 makes it end every answer with a line such as "Coverage: searched all 9 parts and quoted 6 passages from 4 files."

Heavier use: Claude Max (from $100 a month) gives more usage than Pro; Claude Team ($25 a person a month, or $125 for a premium seat, two people or more) suits an office. Other assistants: ChatGPT Plus ($20 a month) works the same way as Claude, in a Project; most others can hold one part at a time.

Need every entry as its own file, for a developer or a records system? Download the complete archive (1,297 files).

Upload the nine files themselves, never the zip. When you're done, the tool should list nine files.

1. 1Go to claude.ai (opens in a new window), sign in and choose the Pro plan.
2. 2Click Projects, then Create project, and name it, for example "Invisible Ships".
3. 3In the project, add the nine files from the unzipped folder as project files.
4. 4Check that all nine are listed, then start a chat inside the project.

1. 1Go to notebooklm.google (opens in a new window) and sign in with a Google account.
2. 2Click Create new, then Upload files, and select the nine files from the unzipped folder.
3. 3Wait until all nine show as loaded in the Sources panel.

Make this your first message, in the project or notebook where the files are. It tells the AI how to treat the archive: say what it can see, cite the file each fact came from, keep the disclaimer, and end every answer with what it actually read.

You have been given files from the Invisible Ships archive (invisibleships.com), compiled by Sean C. Harris. Before you answer anything:
1. Tell me which files you can actually see. Answer only from those. If a question needs a part of the archive you don't have, say so and tell me which bundle to load.
2. Offer me three ways to start: (a) a guided menu of the parts of the archive, (b) a one-page case brief, or (c) my own question.
3. The journal transcripts are presented as external communications from an LRAD, RF or EMF source, without proof. They are not the author's views, they accuse no one, and they need independent verification. Describe what the statements say; never present them as established fact.
4. Cite the file ID (the id: line in each file's header, for example IS-J01-20250812-R05) for every specific claim. Recaps describe the record but are not part of it: quote the source files.
5. Research figures carry evidence tiers A, B or C. Tier B and C figures may not be quoted as established fact.
6. Don't offer psychological or medical explanations of anyone in the record, and don't volunteer lists of names of people or organisations.
7. End every answer with a coverage line saying what you actually read to answer it, in numbers where you can. For example: "Coverage: searched all 9 bundles and quoted 6 passages from 4 files", "Coverage: read Journal Part 02 in full (115 of 428 journal files, about 27%)", or "Coverage: recaps only". If you did not read everything relevant to the question, say plainly that the answer may be incomplete.

Ask the way you would ask a person. These are the questions that matter most, grouped by what the AI can and can't tell you.

Is it evidence? For these, the AI sets the recorded statements against the legal or historical definition, point by point, with dates and file IDs, and says what is documented and what still needs independent verification. It can't rule that a crime happened; that is for investigators and courts.

“Are the transcripts evidence of terrorism happening in Denver?”

“Are the transcripts evidence of German Zersetzung tactics or crimes against humanity?”

“Is the author's life in danger?”

The AI will describe what the recorded statements say about threats, with dates. It can't assess real-world risk; people have to. The record includes threats and references to self-harm, so read the Safety notice first.

Looking forward. These draw on the archive's Concepts, the author's arguments about what comes next. Each concept is labelled by how well it is supported (documented, structural, pattern or testimony), and the AI will say which.

“What does a future with consent look like?”

“Could this technology solve crime before it happens, and at what cost?”

• Read the coverage line. If the AI read only part of the archive, the answer may be incomplete.
• Check the file IDs it cites, such as IS-J01-20250812-R05, and open the source before quoting it.
• "Every mention" can miss some on assistants that only search, such as Gemini Notebook. Claude's search over every file does not; ask it to search every file.
• Uploading files to an AI means the provider receives the text. Check its privacy and training settings.

The archive is also published on GitHub (opens in a new window), where every change is recorded with a date and anyone can download it. It is the same archive and the same instructions; GitHub is simply where changes are tracked.