PaperForge

This skill turns a Zotero library into a local research memory that OpenClaw or Codex can reason over.

PaperForge supports optional built-in acquisition tasks for OpenClaw or Codex, but still treats Zotero as the source boundary for the core memory loop.

This skill is self-contained. If installed by itself, initialize local config from assets/paperforge.config.example.json via scripts/init_config.py.

Use This Skill For

• importing Zotero exports and attachment paths into a local workspace
• watching Zotero for changes and triggering incremental sync
• creating executor-facing acquisition tasks for finding and importing papers into Zotero
• working after papers were added to Zotero either manually or through a PaperForge acquisition task
• detecting papers that still need AI summary
• generating executor task files for paper summaries
• storing structured paper memory after the executor reads a paper
• assembling evidence bundles for novelty, limitation, and idea-refinement questions

Do Not Use This Skill For

• replacing Zotero as the source of truth for raw papers
• embedding model logic into local scripts
• broad web search before checking local evidence

Core Principle

Split responsibilities cleanly:

• acquisition is optional: human download or PaperForge acquisition task
• Zotero stores papers and human corrections
• local scripts prepare, persist, retrieve, and package evidence
• OpenClaw or Codex performs the actual model reasoning

Minimal Workflow

1. Optional: create an acquisition task with python3 skills/paperforge/scripts/init_acquisition_task.py --config paperforge.config.json --title "paper title"
2. If acquisition happened, save the result with python3 skills/paperforge/scripts/save_acquisition_result.py --config paperforge.config.json --input /path/to/result.json
3. If paperforge.config.json does not exist yet, create it with python3 skills/paperforge/scripts/init_config.py
4. Run python3 skills/paperforge/scripts/watch_zotero.py --config paperforge.config.json --interval 30 for automatic detection, or run sync_zotero.py manually.
5. Inspect workspace/index/pending_summaries.json
6. For each pending paper:
   • run python3 skills/paperforge/scripts/init_summary_task.py --config paperforge.config.json --paper-id <paper_id>
   • read the generated task file and references/summary_prompt.md
   • read the linked PDF
   • produce structured memory following references/schemas.md
   • persist it with python3 skills/paperforge/scripts/save_memory.py --config paperforge.config.json --paper-id <paper_id> --input /path/to/memory.json
7. Before answering a research question, build a context bundle:
   • python3 skills/paperforge/scripts/build_context.py --config paperforge.config.json --question "..." --idea workspace/ideas/current_idea.md
8. Read references/qa_prompt.md and use the generated context bundle as the grounding package for the executor's final answer.

When To Read References

• Read references/schemas.md when creating or validating AI summary payloads.
• Read references/acquisition.md when using PaperForge's built-in acquisition flow.
• Read references/acquisition_prompt.md and references/acquisition_result_schema.md when completing an acquisition task.
• Read references/summary_prompt.md when preparing a single-paper AI summary.
• Read references/qa_prompt.md when answering a grounded idea question from a context bundle.
• Read references/workflow.md when you need the full end-to-end process for sync, summarization, and idea-centered Q&A.

File Rules

• source.json stores imported Zotero metadata and attachment paths.
• memory.json stores structured AI understanding of a paper.
• context/*.json stores evidence bundles for downstream reasoning.
• assets/ contains templates and example inputs required to use the skill standalone.
• Keep memory.json machine-readable and concise. Put the final long-form answer in executor output, not inside the memory file.

Retrieval Rules

• Always search local memory.json files before using the web.
• Use acquisition tasks only to get papers into Zotero.
• Use the web only when the local workspace lacks enough evidence or the user explicitly asks for newer external work.
• When web evidence is added, label it as external rather than Zotero-backed.