paper-catalog-update

name	paper-catalog-update
description	Monthly prompt engineering paper catalog update with scoring rules and safe add/retire operations. Triggers: update paper catalog, refresh paper DB, paper database update.

paper-catalog-update

Usage

Invocation name: paper-catalog-update

Use this skill when updating the prompt engineering paper catalog in:

• documents/reference/papers/prompt-engineering-papers.md
• documents/reference/papers/categories/*.md

Typical triggers:

• (today - Last_Updated) > 30 days
• User explicitly asks for paper DB/catalog update
• Master-prompt-writer is invoked and database freshness must be checked

Responsibility

This skill owns catalog update procedure only:

• Scope: stale-check, category selection, search, scoring, add/retire, metadata update
• Out of scope: prompt generation, prompt optimization, model choice recommendations

Inputs

• today date
• Last Updated value from master index
• Optional: target categories requested by user

Required Tools

Tool	Purpose	Fallback
context7/*	Library/framework doc lookup	Skip if unavailable; note in report
Web search (if available)	arXiv, Semantic Scholar, ACL Anthology queries	Manual user search; tag [TOOL LIMITED]
read / search	Read existing catalog files	Required; no fallback
memory/*	Store update decisions	Required; no fallback
edit / replace_string_in_file	Write updated catalog files	Required; no fallback

If web search tools are unavailable, this skill MUST:

1. Report [TOOL LIMITED] in the output summary.
2. Limit updates to papers the user provides directly.
3. Not fabricate citation metadata.

Protocol

1. Check freshness

• Read Last Updated from documents/reference/papers/prompt-engineering-papers.md.
• If within 30 days and no explicit update request, skip update and report "fresh".

2. Select target categories

• If user specified categories, use them.
• Otherwise prioritize: lowest current count first, then oldest category update date.

3. Collect candidates

• Search authoritative sources (arXiv, Semantic Scholar, ACL Anthology).
• Prefer papers published since the last update date.

4. Score candidates

• Use the rubric from the master index:
  • novelty: 1-100
  • impact: 1-100
  • score = round((novelty + impact) / 2)

5. Apply add/retire policy

• If category count < 100: add candidates with score >= 50.
• If category count >= 100: add only if score is above current minimum.
• Retire lowest-scoring entries until count returns to 100.
• Never retire score >= 80 papers automatically.

6. Update metadata

• Update category current counts in the master index.
• Update Last Updated in modified files.
• Keep tag/category format unchanged.

7. Report summary

• Return updated categories, added count, retired count, skipped reason (if any), and unresolved uncertainties.

Output Contract

Every run MUST return a structured report with these fields:

Field	Type	Required	Description
freshness_decision	"update" | "skip"	✅	Whether update was needed
skip_reason	string	if skip	Why update was skipped
affected_files	string[]	✅	Paths of modified files
added	number	✅	Papers added
retired	number	✅	Papers retired
categories_touched	string[]	✅	Category names updated
count_changes	{category: before → after}	✅	Per-category count delta
risk_notes	string[]	✅	Each entry: one specific concern (e.g., "arXiv unreachable — 3 candidates scored from abstract only")
tool_limitations	string[]	if any	Tools unavailable during execution

Guardrails

• Preserve existing markdown schema and table structure.
• Do not modify unrelated documents.
• If source confidence is low, prefer skip with explicit note over speculative additions.

Failure Handling & Escalation

Condition	Action
All external sources unreachable	Tag [TOOL LIMITED]. Skip steps 3–4. Report with added: 0 and risk_notes: ["No external sources reachable"]. Return to caller.
Category file schema broken (parse error)	Tag [EXECUTION BLOCKED]. Do NOT write to the file. Report broken file path. Escalate to user for manual fix.
Scoring ambiguity (novelty/impact uncertain for >50% of candidates)	Proceed but flag each ambiguous entry in risk_notes. Do not auto-retire if confidence is low.
Catalog count exceeds 100 after add but retire policy cannot select a victim (all ≥80)	Skip retirement. Report count overflow in risk_notes. Escalate to user for manual curation.