← Blog

What’s inside a Claude project ZIP?

· AI Chat Archive team

Start with the indexes An AI Chat Archive Pro project export puts instructions in instructions.md, knowledge files in files/ and chats in conversations/. Open the root _INDEX.md and files/_INDEX.md first, then check any omission reports. The ZIP is a readable archive you can reuse manually.

This guide describes AI Chat Archive’s project ZIP layout as of 5 November 2026. It covers classic Claude projects and Beta Projects at claude.ai/code/project/…. Whole-project export is a Pro feature. You can test an individual chat free before buying; that does not include a free project export.

A project ZIP, opened

A small export downloads as <Project Name>.zip and contains a folder with that sanitized project name. This example uses Markdown transcripts. Sample titles and documents are illustrative; the filenames for indexes, metadata, instructions and reports are the exporter’s actual names. Reports and attachment folders appear only when relevant.

Research project/
  README.md
  metadata.json
  instructions.md
  files/
    _INDEX.md
    research-notes.md
    source-report.pdf
    reference/                    (Beta Library folders are preserved)
      diagram.png
    _SKIPPED.md                   (failed or cancelled project files)
  conversations/
    2026-11-01_Research plan/
      Research plan.md
      attachments/                (available classic chat files, if enabled)
      _SKIPPED.md                 (files that could not be included)
      _INCOMPLETE.md              (conversation warnings)
    2026-11-02_Summarize findings/
      Summarize findings.md
      outputs/                    (available task or thread outputs, if enabled)
        findings.csv
        _PROJECT_FILES.md         (links to Library files already in files/)
  _INDEX.md
  _INCOMPLETE.md                   (project inventory or content warnings)

You will not necessarily see every optional entry above. An empty or skipped conversation may have no folder. The transcript directory is named conversations/; the exporter does not create a top-level chats/ directory or a separate manifest.json.

README, metadata and instructions

README.md describes the project, records dates and counts, explains the archive and calls out excluded content. metadata.json is its machine-readable companion: project information plus lists of knowledge documents, files and conversations. It keeps text knowledge and binary uploads in separate lists even though both live under files/. It does not duplicate document contents, asset download URLs or the instructions text.

instructions.md starts with an Instructions heading and stores the project’s instruction text inside a fenced block. If you reuse it, copy the text inside the fence. If no instructions were set, the file says so. The wrapper is for reading the saved prompt; pasting the whole wrapper into a new project would add text that was not part of your original instructions.

Knowledge files: original bytes, previews and extracted text

Project knowledge is exported whether or not Include attachments in the ZIP is enabled. That setting controls files attached to conversations. Choosing Markdown for transcripts does not convert all knowledge files to Markdown.

For classic projects, the exporter saves text-knowledge entries from the text Claude exposes, using each entry’s sanitized filename. These are text copies, often Markdown, rather than reconstructed Word or PDF documents. If Claude exposes extracted text instead of an original document, that text can preserve useful content without preserving the original file’s layout or binary format. A filename alone does not prove that you recovered an original binary.

For classic uploaded-file entries, the exporter chooses the first available asset in this order: the document, then a preview, then a thumbnail. A preview or thumbnail can differ from your original upload. This is selection from the assets Claude exposes; a failed document download is reported as a failure rather than triggering a guaranteed attempt at every other asset or a new text extraction. Keep your own originals when exact document fidelity matters.

For Beta Projects, accessible Library files are downloaded as file bytes and retain their relative folders under files/, with names adjusted when needed for safe paths or collisions. files/_INDEX.md lists the filenames actually included. Failed or cancelled project-file downloads appear with reasons in files/_SKIPPED.md.

Chats, threads and task outputs

Each exported conversation has a folder under conversations/, even if attachment bundling is off. Choose one transcript format for the run: Markdown (.md), HTML (.html), JSON (.json) or TXT (.txt). The exporter writes that format, not all four copies of every chat.

There are no PDF transcripts in a project export. An uploaded PDF can still be a knowledge file or attachment. For a PDF transcript, open the individual chat or thread and use free single-chat PDF export, then save through your browser’s print dialog.

Beta archives include the main project conversation, threads and remaining imported chats or Cowork tasks. Continued threads include the original context when accessible, followed by their new replies. An imported source already covered by a continuation is not separately exported again.

With attachment bundling enabled, available classic chat files go into that chat’s attachments/ directory. A continuation’s original classic attachments can appear under attachments/original/. Cowork and Beta output files can appear under outputs/; outputs from multiple task sessions are separated by session ID. Task images can be under attachments/.

If a thread wrote or attached a Library file already saved in the project’s files/ directory, outputs/_PROJECT_FILES.md links to that shared copy instead of saving identical bytes twice. If that note’s name collides with a user file, it can receive a numbered suffix. Library copies reflect the files available at export time. An expired task sandbox may yield a delivered copy or a supported reconstruction from the task log; a potentially stale delivered version carries a warning when its latest state cannot be verified.

How filenames are chosen

Conversation folders use YYYY-MM-DD_<safe title>/, based on the listed conversation creation date in local time. Inside is <safe title>.<ext>. Spaces can remain: Research plan.md is a valid transcript name. Missing or invalid dates can leave the date prefix empty. A title that sanitizes to nothing becomes Untitled.

Unsafe filename punctuation is replaced, whitespace is collapsed and names are capped at 120 characters before further collision handling. Leading dots are removed from project and conversation names. If the project name becomes empty, its folder is Project/. Two chats with the same date and safe title get distinct folders, such as 2026-11-01_Research plan (abcd1234)/, using the first eight identifier characters; further collisions add a counter.

Classic knowledge files use numbered suffixes such as notes (2).md for duplicates. The index/report names are reserved, so a knowledge document called _INDEX.md is renamed rather than overwriting the index. A text entry with no usable filename falls back to doc-1.md, with the number based on its position; binary entries use file-1 and similar names. Beta Library path collisions are also numbered, accounting for case-insensitive filesystems.

Read the reports before relying on the archive

Automatic project Memory is outside the archive. Beta exports also exclude scheduled-task definitions, repository clones and connector configuration. Shell-created files without a visible output path or delivery record may not be discoverable. An OK row is the exporter’s outcome for that conversation, not a guarantee that every part of Claude’s workspace can be restored.

Troubleshooting

I received several ZIPs

Large exports are saved in parts, normally around 500 MiB each. Their filenames include a run timestamp and (part 1), (part 2) and so on. A large item can span parts. Extract every part into the same destination, in order. Part 1 includes How to combine the parts.txt inside the project folder; the final part’s root index includes a Part column. Keep every part from the same run.

The export stopped or I cancelled it

Keep the saved parts. If the extension offers Resume export, use it to continue an interrupted export that already saved a part. It keeps the original format and attachment setting and can retry failed conversations or attachments. A run that never saved a part has no saved progress to resume. Follow the combine note and check the latest reports after extracting all parts.

A file is missing or will not open

Check both skipped-file reports and the root incomplete report. Confirm that you can still open the project in the signed-in Claude account, then retry when the session and service are working. For a classic knowledge file, inspect whether Claude supplied a preview or text copy. Compare against an original you already have; a text export cannot recreate a missing original document.

I expected PDFs or chat attachments

Use single-chat export for PDF transcripts. Enable attachment bundling before project export for available conversation files; project knowledge files are attempted either way. Read the project-export FAQ for the controls and setup.

Exports use your own signed-in Claude session and are processed in your browser. Chat content is not uploaded to our servers. Pro license checks send Polar only your key, never chat content. See the privacy policy.

For setup and a demo, visit Claude project download. To reuse the archive in a second account, follow the manual project rebuild guide.

Know where every part of your project belongs

Whole-project ZIP export requires Pro. Single chats are free and unlimited in PDF, Markdown, HTML, JSON and TXT.

See Claude project export See Pro plans