Summarize Paper

Phase 2: Write

1. Read only the sections identified in Phase 1. Use pdf_search with the queries from Phase 1.
2. Follow template.md exactly. Include only the mechanism blocks matching the paper type. For General papers, invent domain-appropriate blocks that parallel the structure and granularity of the Learning/Physics blocks (see template for guidance).
3. For the logical flow, map every paper to: Problem → Bottleneck → Mechanism → Steps → Evidence → Boundary.

Style

Intuitive first, formal second — every section (except Results) should read like explaining the paper to a smart colleague, not like paraphrasing the abstract:

• Lead with why something is needed or why it works, then give the formal name/equation.
• Use analogies, concrete mental images, and cause-effect reasoning (e.g. "triangular solves are sequential because each row depends on the previous — that kills GPU parallelism" rather than "triangular solves are poorly parallelizable").
• The High-Level Idea is the single most important paragraph: the reader should think "ah, I get it" before seeing any math.
• Rephrase jargon in plain words first, formal name in parentheses after.
• Keep each bullet to one core idea. If a bullet needs two sentences, split it or cut.
• Exception: Results (Section 5). Rigorous only — exact numbers, metrics, comparisons as stated. No analogies.

Conciseness:

• For long derivations: state start + result, then *(see Section X)*.
• Only include equations that appear in "How it works" or Pseudocode.
• For theorems/proofs: one-sentence claim; skip proof.

Output Rules

• Append *(Source: Section X Title)* after each item/bullet.
• Hard limit: 680 words (excluding pseudocode/algorithm block). Aim for the length in example.md.
• Save output using the save_summary MCP tool. Pass the source PDF path and the full Markdown content. The tool derives the filename (<pdf_stem>_summary.md), creates Summaries/ if needed, and writes the file. Do NOT use any other method to write the summary.
• After saving, call the convert_md_to_pdf MCP tool with the path returned by save_summary. This produces a matching PDF in the same directory (Summaries/<pdf_stem>_summary.pdf). Both files share the same base name — only the extension differs.