curatorbot / project-kit /README.md
ignatius's picture
Partner context: pointers to USHMM's intake workflow and the Accessioning Rationale Form
2e7e36c
|
Raw History Blame Contribute Delete
2.35 kB

Setting up the Claude and ChatGPT projects

Both projects get the same instructions and the same files, so when their answers differ you are comparing models, not setups. Use them for planning, prompt design, writing and code review; the repository and the Hugging Face Space are where the demo is actually built and run.

Steps (about 10 minutes for each project)

  1. Create the project. Name it CuratorBot. Description: "Evidence-grounded assistant for oral history intake review; hosted demo for USHMM and other partners."

  2. Paste the instructions. Copy all of instructions.md into the project's instructions (Claude: project instructions; ChatGPT: project instructions).

  3. Add the knowledge files from knowledge/. Files 01, 02 and 07 are internal and are not in the public repository; they come in the separate project-kit download.

    File What it gives the model
    01_curatorial_workflow.md The museum's intake workflow and value proposition
    02_design_plan.md The full design, implementation and deployment plan
    03_repository_overview.md README, coding rules, brief schema and data policy
    04_spatio_textual_README.md The library's features, CLI and output fields
    05_vocabulary_demo.txt The closed heading list used in the demo
    06_synthetic_transcript.txt A fictional interview to test prompts on
    07_ushmm_intake_workflow_and_arf.md USHMM's intake stages, the Accessioning Rationale Form section by section, and where CuratorBot fits (internal)

    If your Claude plan lets you connect GitHub to project knowledge, add SpaceTimeNarratives/curatorbot as well, so the model sees current code.

  4. Start one conversation per workstream with the starters in kickoff-prompts.md.

  5. Keep both in step. When a knowledge file changes in the repo, replace it in both projects the same day.

What not to put in either project

Recordings, transcripts or catalogue records you are not cleared to share, donor submissions under consideration, API keys. See docs/data-policy.md.

Coding in the repository

CLAUDE.md (Claude Code) and AGENTS.md (Codex and other coding agents) at the repository root carry the same rules, so a coding agent working on the repo follows the principles the chat projects follow.