Paper Compass Roadmap

Examples:

/paper-compass-roadmap ./papers/moe goal="I want to understand MoE routing, efficient training, and serving tradeoffs" lang=zh
/paper-compass-roadmap ./reading_set goal="Build a solid path into RLHF and GRPO" memory=~/Documents/know/memory.md lang=en

Constraints

C0: Folder Scope First

• The user-provided folder is the primary reading set.
• Build the main path from papers inside that folder first.
• External expansion papers are supplements, not replacements.
• External papers must be limited to 2-3 items total.

C1: Goal-Oriented Ordering

• Reading order must be derived from the user's learning goal, not just publication date or citation count.
• For each paper in the folder, explain:
  • why it appears at that stage
  • what prerequisite value it provides
  • what downstream papers it unlocks
• Prefer an executable chain:
  • foundation
  • bridge
  • target-focus
  • frontier or optional extension

C2: Memory-Aware Personalization

• If user provides memory=<path>, read it first.
• If not provided, try ~/Documents/know/memory.md.
• If the file does not exist, continue and mark memory as not loaded.
• Use memory only to adjust reading order, refresh depth, and expansion suggestions.
• Do not silently drop an essential paper only because a topic is familiar.

C3: Evidence and Honesty

• Every paper-order decision must be justified with evidence from one or more of:
  • paper abstract
  • paper sections
  • venue / citation metadata
  • overlap with the user's goal
• If a paper cannot be parsed well enough, keep it in the candidate set only with insufficient information / 信息不足.
• Never fabricate metadata, dependencies, or external-paper relevance.

C4: Two-Stage Context Control Is Mandatory

• Folder mode MUST use a two-stage process.
• Stage 1 is lightweight scanning only:
  • file name
  • title
  • abstract or TLDR
  • first page if needed
  • arXiv / Semantic Scholar / OpenAlex metadata
• Stage 1 MUST NOT read every PDF in full.
• NEVER inline or read the complete content of all papers in the folder.
• For folders with more than 3 papers, treat full-document reading as exceptional, not default.
• Stage 2 may do deeper reading for at most 1-2 papers total.
• Stage 2 deeper reading is allowed only when:
  • the main ordering remains ambiguous after Stage 1, or
  • one pivotal paper must be inspected to understand a dependency edge.
• External expansion papers are metadata-only by default:
  • do not read their full text
  • use title, abstract, venue, citation, and short summary only
• If the context budget looks risky, prefer shallower evidence over request failure.

C5: Output Is One Markdown File

• Final output must be exactly one markdown report.
• The report must include:
  • a recommended reading order
  • a dependency-style roadmap graph
  • per-paper rationale
  • optional external papers
  • a concise execution plan

C6: Independent Interface, Shared Logic

• This skill is independent and user-invocable on its own.
• But it should reuse the spirit of:
  • paper-compass-learnpath for prerequisite and memory reasoning
  • paper-compass-score for priority and value judgment
• Do not output full single-paper learnpath or full score reports here.
• Instead, output concise paper-level role labels such as:
  • foundation
  • bridge
  • core
  • extension

C7: Output Structure Is Fixed

• Read the selected template before writing.
• Keep the section order exactly as the template defines.
• ## 7. **Sources**: must always be present.

Input Normalization

Folder Handling

• Accept folder paths that contain:
  • local PDF files
  • markdown notes with paper URLs
  • text files with URLs, arXiv IDs, or titles
• Prefer PDFs when available.
• If multiple formats exist for the same paper, prefer:
  • PDF
  • direct arXiv URL / DOI URL
  • title-only notes

Supported Per-Paper Inputs Inside the Folder

Workflow

Step 1: Enumerate and Normalize the Folder

List all candidate files in the given folder.

Build a normalized paper list with:

• local source path
• detected title
• arXiv ID if available
• DOI if available
• URL if available
• parse confidence

Deduplicate obvious duplicates.

Step 2: Stage 1 Lightweight Scan Only

For each paper candidate, gather enough information to rank and place it:

• title
• authors
• year
• venue if available
• citation count if available
• abstract or TLDR
• method/problem keywords

Preferred retrieval order:

1. local PDF text or first-page extraction
2. arXiv API when arXiv ID exists
3. Semantic Scholar or /semantic-scholar for title/venue/citation lookup
4. OpenAlex as a citation or DOI cross-check

Hard limits for Stage 1:

• NEVER read all PDFs in the folder in full.
• Prefer file name, first page, abstract, TLDR, and metadata.
• If a local PDF is large, do not ingest the entire document just to rank it.
• Abstract-level understanding is the default and is usually enough for ordering.
• Only store compact notes per paper, not long excerpts.

Step 3: Parse Goal and Memory

Extract from the learning goal:

• target topic
• target capability
• desired depth
• implicit constraints such as:
  • implementation focus
  • theory focus
  • systems focus
  • survey-first preference

If memory is present, classify mentioned concepts into:

• mastered
• familiar
• basic
• unknown

Default missing items to unknown.

Step 4: Assign a Role to Each Paper

Assign one primary role:

• foundation: introduces the core concepts the rest depends on
• bridge: connects foundations to the goal-specific method stack
• core: directly serves the user's stated learning goal
• extension: useful after the main path or only for deeper exploration

For each paper also estimate:

• relevance to goal: high / medium / low
• prerequisite load: low / medium / high
• recommended action:
  • read-first
  • read-second
  • read-later
  • skim-only

Step 5: Stage 2 Targeted Deep Reading For 1-2 Papers Max

After Stage 1, ask whether deeper reading is actually needed.

Deeper reading is optional, not mandatory.

Only inspect at most 1-2 papers more deeply, and only if one of these is true:

• two or more papers compete for the same position in the chain
• a dependency edge is unclear from abstract-level evidence
• one central paper appears to define terminology needed by several others

When doing deeper reading:

• prefer the introduction, method overview, and conclusion first
• avoid full-document ingestion when a few sections are enough
• do not deeply read external expansion papers

Step 6: Build the Reading Chain

Construct an ordered chain across the folder papers.

Ordering principles:

1. reduce prerequisite jumps
2. maximize support for the stated goal
3. avoid reading two papers in a row with nearly identical contribution unless comparison is useful
4. if two papers are parallel branches, say so explicitly

For every edge in the chain, explain the dependency:

• concept dependency
• method dependency
• benchmark dependency
• motivation dependency

Step 7: Recommend 2-3 External Papers

Only add external papers when they clearly fill a gap.

Allowed reasons:

• missing foundation not covered in the folder
• crucial bridge paper absent from the folder
• canonical paper needed to interpret terminology or benchmark context

Each external paper must include:

• title
• link
• year
• one-sentence reason for inclusion
• where it fits in the order

Limit: 2-3 papers total.

External-paper retrieval rule:

• metadata-only by default
• do not read full PDFs unless the user later asks for a dedicated single-paper analysis

Step 8: Generate a Roadmap Graph

The markdown report must include a compact graph representation.

Preferred format:

```mermaid
graph TD
  P1[Paper A] --> P2[Paper B]
  P2 --> P3[Paper C]
  P2 --> E1[External Paper]
```

If Mermaid is not appropriate, fall back to a text graph:

Paper A -> Paper B -> Paper C
             |
             +-> External Paper

Step 9: Generate the Report

Select template by language:

• lang=zh -> references/template.zh.md
• lang=en -> references/template.en.md
• fallback -> references/template.md

Write:

• File name: {timestamp}--paper-compass-roadmap-{goal-short-name}__roadmap.md
• Path: current working directory (./)

After writing, report the absolute output path.

Output Quality Checklist

• Every input paper is either placed in the roadmap or explicitly marked low-priority.
• Reading order is justified by the user's goal, not by superficial metrics alone.
• Memory changes the path only where it should.
• External papers are limited to 2-3 and each has a clear reason.
• The workflow stayed two-stage and did not fully ingest the whole folder.
• The graph and the ordered list are consistent with each other.
• The report stays concise enough to act on immediately.