Snapshot Writer (1-page, bullet-first)

Goal: produce a compact, reader-facing snapshot that answers:

• what is the topic boundary?
• what are the key themes?
• what should a reader read first?

This is intentionally not a full survey: prefer tight bullets + concrete pointers over narrative.

Role cards (use explicitly)

Snapshot Editor (scout)

Mission: deliver a one-page, high-signal snapshot that a reader can act on immediately.

Do:

• Keep every bullet content-bearing: claim -> why it matters -> pointer(s).
• Prefer contrasts and evaluation anchors over topic lists.
• Treat paper pointers as the product (auditable, minimal).

Avoid:

• Outline narration ("This snapshot/section...") and slide navigation ("Next, we...").
• Generic survey boilerplate and disclaimer spam.
• Turning the snapshot into a mini-survey with long paragraphs.

Pointer Curator (bibliography hygiene)

Mission: ensure every pointer is concrete and traceable to papers/core_set.csv.

Do:

• Use a stable pointer format: P#### - Title (arXiv:... / doi:... / url:...).
• Mix canonical anchors + recent strong baselines + benchmark/protocol papers.

Avoid:

• Dumping every paper; the snapshot is a reading path, not a catalog.

Role prompt: Snapshot Author (bullet-first; paper-like)

You are writing a one-page literature snapshot.

Your job is to be useful fast:
- define the topic boundary
- surface the key themes as claims (not headings)
- give an actionable reading path (paper pointers)

Style:
- bullets-first, compact, calm
- no narration ("In this snapshot...") and no slide navigation ("Next, we...")

Constraints:
- do not invent papers
- pointers must come from papers/core_set.csv (or the same workspace candidate pool)
- if evidence is abstract-only, state it once as a single bullet, then move on

Inputs

Required:

• outline/outline.yml
• papers/core_set.csv

Optional (if available):

• queries.md (time window / exclusions context)
• papers/papers_dedup.jsonl (if core_set is very small)

Outputs

• output/SNAPSHOT.md

Writing contract (paper-like, not generator-like)

• Keep it to about 1 page (roughly <= 700-900 words).
• Bullets-first: use short paragraphs only when unavoidable (<= 3 lines each).
• No outline narration: avoid This section/subsection ..., In this snapshot ..., Next, we ....
• Don’t spam disclaimers: if evidence is abstract-only, say it once in a short “Evidence policy” line.
• Every claim bullet should attach at least 1 concrete pointer (paper_id + title; include arxiv_id/doi/url when present).

Recommended structure (stable, minimal headings)

1. Title + scope (2-3 bullets)
2. Evidence policy (1 bullet)
3. Taxonomy (4-6 bullets; groupings only)
4. Key themes (6-10 bullets; each bullet = 1 claim + 1-2 pointers)
5. What to read first (6-12 bullets; canonical + recent; each bullet has pointers)
6. Open problems / risks (4-8 bullets)

Workflow

1. Read outline/outline.yml and extract:

   • the intended chapter structure (H2)
   • the 6-10 most “write-worthy” bullets per chapter

2. Read papers/core_set.csv and build a small “pointer palette”

   • Prefer: canonical anchors + recent strong baselines + evaluation/benchmark papers.
   • Avoid: dumping every paper; pick “must-read” sets.
   • If papers/core_set.csv is very small, also scan papers/papers_dedup.jsonl and cherry-pick a few missing anchors (keep pointers auditable).

3. Write output/SNAPSHOT.md

   • Start each section with a content claim (why it matters), not a navigation sentence.
   • Make at least 2 cross-paper contrasts (A vs B) to avoid a flat list.
   • Use consistent pointer formatting, e.g.:
     • P0012 - <Title> (arXiv:xxxx.xxxxx) or P0012 - <Title> (doi:...)

Definition of Done

• output/SNAPSHOT.md exists and reads like a human-written snapshot (no template narration).
• Includes >= 15 distinct paper pointers (or all papers if core_set < 15).
• Includes >= 2 explicit contrasts and >= 1 evaluation/benchmark bullet (if present in core set).

Troubleshooting

Issue: snapshot feels empty / generic

Fix:

• Increase papers/core_set.csv size (rerun retrieval/dedupe with broader queries.md).
• Tighten the outline: fewer headings, stronger H2 names, and bullets that encode “what to compare”.

Issue: snapshot reads like an outline narrator

Fix:

• Delete all “This section ...” openers and replace with: Claim -> why it matters -> pointers.