| name | distill-paper |
| description | Read a paper or product writeup (and its key references) and distill it into a structured, implementable spec: the core mechanism, the claimed metric, a minimal reproduction, and the constraints. Use at the start of adopting a community idea, before reproducing or grafting it. |
Status: v0 scaffold — unverified. This runbook's end-to-end behavior has not
been validated end-to-end in this repo; it will be hardened against real papers.
See ../../STATUS.md.
distill-paper
Turn "I saw an idea in a paper/product" into a spec you can act on — the analog of
program.md.
Steps
- Read with a subagent so heavy reading stays out of the main context.
- Follow the lineage. Pull the most relevant cited references (and their repos)
to recover implicit conventions the paper omits (the AutoReproduce idea).
- Extract:
- the core mechanism (what actually changes and why it should help);
- the claimed metric and dataset/benchmark;
- a minimal reproduction (smallest setting that should show the effect);
- the constraints / assumptions (compute, data, preconditions);
- risks (what the authors gloss over; overfitting, cherry-picked baselines).
- Emit a structured spec (Markdown) with those sections.
Output
A spec file that downstream skills consume: repro-harness (to confirm) and
graft-change (to integrate). Keep claims sourced (links/arXiv ids) so they stay
checkable.
Notes
- Distinguish the idea from the paper's framing; you want the transferable
mechanism, not the narrative.
- Flag any number you could not verify rather than restating it as fact.
