| name | paper-writing |
| description | Use when writing or polishing academic papers. Use when the user mentions writing a paper, system paper, AI paper, polishing a draft, academic writing, or needs hierarchical logic flow for consistency. |
Paper Writing — Hierarchical Logic Stream (HLS)
When to Use
| Scenario | Mode | Path |
|---|
| Starting a paper from scratch | Write | Full HLS: L0 → L1 → L2 → L3 |
| Polishing an existing draft (no HLS) | Polish | Extract L0/L1 from draft → critical review → L2/L3 |
| Polishing an existing draft (HLS exists) | Polish | Critical review L0/L1 → L2/L3 |
| Language polish only (draft is structurally sound) | Polish-lite | Skip L0/L1. Use existing structure, run L2/L3 for prose |
Core principle: Write mode builds from ideas. Polish mode extracts structure from a draft. Both use the same HLS levels — only the starting point differs.
Overview
4-level refinement. Each level validates before descending. No skipping. No merging.
Venue-driven discovery: Target venue determined at L0. All resources (skeleton, blueprint, writing guide) auto-discovered by exploring templates/ — never hardcoded.
Every level commits to git.
Universal Interaction Protocol
All L0/L1 confirmations use clickable options — never free-text. The user clicks, does not type.
| Interaction | Option Pattern |
|---|
| Venue selection | List discovered venues as clickable options (e.g., "NeurIPS (9pg)", "ICML (8pg)", "AAAI (7pg)", "OSDI (12pg)"), mark recommendation |
| Point judgment | "PASS — [reason]" / "WEAK — [what's missing]" / "REJECT — [why]" as selectable options |
| Section name | 2-3 name options with purpose, mark recommendation |
| Flow chain | Accept / Reorder / Add step / Remove step / Revise |
| Figure placement | Accept / Move X to step Y / Add at step Z |
| Polish issue | "Accept suggestion" / "Revise — [counter-proposal]" / "Skip — not relevant" |
| Proceed | "Confirmed → next" / "Let me revise this" |
Every prompt to the user MUST end with a set of clickable options. L0 and L1 each specify which options apply at which step.
Polish Mode
When polishing an existing draft, HLS documents MUST be generated — not just discussed verbally.
- If no L0/L1 exists — read draft → extract implicit HLS → critical think silently (identify issues, form 2-3 suggestions) → present to user → discuss → write
docs/systematic-research/plans/stream-L0.md and stream-L1.md → commit
- If L0/L1 exist — critical review: points still valid? chains match draft? → critical think (2-3 issues + suggestions) → present → discuss → update L0/L1
- Invoke L2 — parallel revision following (revised) flow chains + writing guide
- Invoke L3 — consistency check (7 structural + prose checks). Polish complete when all checks pass.
Polish-lite (language only): User says "just polish the language" → use writing guide for voice/pitfalls → L2 revise prose → L3 consistency. Do NOT restructure.
Write Mode (Full HLS)
Standard flow: L0 → L1 → L2 → L3. Start from ideas, end with paper.
Hard Gates
Do NOT proceed to L1 until: target venue determined, 6 core points written to `docs/systematic-research/plans/stream-L0.md`, each point discussed ONE AT A TIME and judged PASS, NOT rejected, user confirmed each point individually.
Do NOT batch all 6 points into a single message. Ask one → judge → user confirms → next.
**Every challenge MUST pass the causal test:** "would solving the root cause eliminate this?" Weak links = reject or remove the challenge.
Do NOT proceed to L2 until: each Section proposed ONE AT A TIME, name confirmed, A→B→C flow chain defined and confirmed, figures placed, L1 document accumulated incrementally and user approved.
Do NOT present all Sections at once. Propose Section N → discuss → confirm name → define chain → confirm chain → next Section.
Do NOT proceed to L3 until: skeleton copied to `paper/`, every Section drafted by its OWN parallel agent (all dispatched simultaneously, one per Section), each Section follows its L1 flow chain, abstract written (5-sentence formula, consistent with all Sections), user reviewed each Section.
Do NOT write Sections sequentially. Dispatch all Section agents at once.
Paper NOT complete until: all 7 consistency checks pass (6 structural + 1 fine-grained line-by-line prose polish), report shows zero issues.
Process Flow
digraph hls {
rankdir=TB;
"Start" [shape=oval];
"L0: Judge + Venue\n[l0-core-idea.md]" [shape=box];
"User OK?" [shape=diamond];
"Commit L0" [shape=box];
"L1: Architect\n[l1-chapter-structure.md]" [shape=box];
"User OK?" [shape=diamond];
"Commit L1" [shape=box];
"L2: Build\n[l2-parallel-writing.md]" [shape=box];
"User OK?" [shape=diamond];
"Commit L2" [shape=box];
"L3: Audit\n[l3-consistency-check.md]" [shape=box];
"7 checks pass?" [shape=diamond];
"Commit final" [shape=box];
"Done" [shape=oval];
"Start" -> "L0: Judge + Venue\n[l0-core-idea.md]";
"L0: Judge + Venue\n[l0-core-idea.md]" -> "User OK?"; "User OK?" -> "L0: Judge + Venue\n[l0-core-idea.md]" [label="no"];
"User OK?" -> "Commit L0" [label="yes"]; "Commit L0" -> "L1: Architect\n[l1-chapter-structure.md]";
"L1: Architect\n[l1-chapter-structure.md]" -> "User OK?"; "User OK?" -> "L1: Architect\n[l1-chapter-structure.md]" [label="no"];
"User OK?" -> "Commit L1" [label="yes"]; "Commit L1" -> "L2: Build\n[l2-parallel-writing.md]";
"L2: Build\n[l2-parallel-writing.md]" -> "User OK?"; "User OK?" -> "L2: Build\n[l2-parallel-writing.md]" [label="no"];
"User OK?" -> "Commit L2" [label="yes"]; "Commit L2" -> "L3: Audit\n[l3-consistency-check.md]";
"L3: Audit\n[l3-consistency-check.md]" -> "7 checks pass?"; "7 checks pass?" -> "L2: Build\n[l2-parallel-writing.md]" [label="no"];
"7 checks pass?" -> "Commit final" [label="yes"]; "Commit final" -> "Done";
}
Quick Reference
| Level | Output | Key Action | Ref |
|---|
| L0 | docs/systematic-research/plans/stream-L0.md | Venue + judge 6 points, reject until clear | l0 |
| L1 | docs/systematic-research/plans/stream-L1.md | Propose 2-3 custom structures, define A→B→C chains | l1 |
| L2 | paper/ dir | Write: draft Sections + abstract. Polish: revise Sections + abstract following writing guide | l2 |
| L3 | Consistency report | Run 7 checks (6 structural + 1 prose), fix root causes. Polish-lite also runs L3 | l3 |
Git
git commit -m "L0: core idea for <topic>"
git commit -m "L1: structure for <topic>"
git commit -m "L2: draft for <topic>"
git commit -m "L3: final for <topic>"
Red Flags
- "The core idea is clear enough, let's start writing" → return to L0
- "We can figure out the details while writing" → return to L0
- "This point is obvious, no need to write it down" → return to L0
- "We can't write anything until experiments are finished" → return to L0 (draft mode exists)
- "We can just make up placeholder numbers without marking them" → stop
- "One commit at the end is fine" → return to current level
- "I'll check consistency after the paper is accepted" → return to L3
- "We need three challenges to match three designs" → return to L0 (challenges are determined by evidence, not symmetry)
- "This challenge is related to the root cause somehow" → return to L0 (causal link must be direct, strong, and testable)
Anti-Patterns
- Writing before clarifying → L0 must be committed before any prose
- Soft judgement → use rejection criteria mechanically
- Editing skeleton in-place → copy to
paper/ first; templates are read-only
- Big bang commit → four commits, one per level
- Hardcoding paths → discover from venue, never assume skeleton location
