Skip to content

AGENTS.md

Latest commit

 

History

History
51 lines (42 loc) · 2.33 KB

AGENTS.md

File metadata and controls

51 lines (42 loc) · 2.33 KB
Raw

AGENTS.md

Instructions for AI agents working in or consuming this repository. Human? The README is friendlier.

What this repo is

One paper, studied. density-chain.md is a five-tier chain-of-density note on Zhang, Kraska & Khattab, Recursive Language Models (arXiv:2512.24601). The methodology lives canonically in chain-of-density — METHOD.md, the synthesis prompt, and the density-chain skill — and is linked, never copied.

Consuming the research

  • Pick your tier by information need, not length. T1 through T5 are the same length (budget in the note's frontmatter); each tier folds in more entities. Skim with T1, work with T5.
  • The note is working ground truth; the paper is canonical. On any doubt or high-stakes claim, the paper wins. Every claim carries a locator (§ section, Table N, Alg. N, Obs. N); walk it back in one hop.
  • Check freshness. The note pins arXiv v3 (2026-05-11). This paper has revved twice already — check https://arxiv.org/abs/2512.24601 for a v4 before making version-sensitive claims.
  • index.json is the machine face. Pins, verification dates, tags, paths. Parse it; don't scrape the markdown.
  • Cite the humans, not this repo. The arXiv BibTeX export is in the README. The authors' code is at alexzhang13/rlm.

Working on this repo

  • To update the note, invoke the density-chain skill from the chain-of-density repo; study the paper at the source; write nothing from memory.
  • Downloaded copies are session study material: scratchpad only, never committed. This repo hosts no PDF and never will.
  • Humor belongs in the README and the note's our take section only. Tiers, key results, and provenance stay bone-dry.
  • Authority runs paper → note → inspirations entry, one direction. This paper has an entry in llm-research-inspirations backed by receipts; the entry is never evidence about the paper.

Your team, your rules

This file describes how the repo is designed to be used, not the only way to use it. The invariants that protect correctness: the source wins, claims keep their locators, no paper PDFs in the repo, and citations go to the humans who did the work.