lazy-paper template turns YOUR research question into the outline the
pipeline runs on — instead of picking the closest hand-written template (the
single most load-bearing choice, see README), you generate one that matches
both the paper and your intent.
# From a new PDF (text-layer prescan, no OCR API):
uv run python -m cli template --idea "can this transfer to bipeds?" \
--pdf papers/mypaper.pdf --lang en
# From an existing run (richer: chapters, captions, context):
uv run python -m cli template --idea "能量项的代价与约束" --run mypaper
# Grounded in your knowledge library (adds cross-paper comparison questions):
uv run python -m cli template --idea "..." --run mypaper --use-libraryOutput: templates/auto-<idea-slug>.docx + .prompt.md / .response.json
audit sidecars. The console prints the full outline for review.
- The docx is plain Word — open it, edit titles/questions, delete or add sections. Human review is the intended last step, not an afterthought.
- It round-trips deterministically through stage s05: numbered lines
("1 Title") become section headings; every question is a plain paragraph
and lands in that section's
guidance. A self-check parses the file with the real s05 parser right after writing and hard-fails on any drift. - Then run the pipeline as usual:
… run --pdf <paper> --template <generated>.
| Flag | Effect |
|---|---|
--idea (required) |
Your lens; ≥ half of all questions must serve it |
--pdf / --run |
Prescan source: cheap text layer vs existing artifacts |
--use-library |
Inject library manifest + idea-relevant excerpts; forces ≥ 2 named cross-paper comparison questions |
--sections N |
Outline width (default 6) |
--lang zh|en |
Language of titles and questions |