Korean Analyst Report Insights

Given the extracted-text JSON from kr-analyst-report-fetch, produce a standalone Markdown digest of sell-side coverage. The digest is the final deliverable of the kr-analyst-report-* chain.

When To Use

• After kr-analyst-report-fetch has produced an extracted-text JSON for a ticker.
• When the user asked for a sell-side coverage summary, TP consensus, or a street-view brief.

Do not use this skill to:

• Produce a memo thesis — belongs in kr-stock-analysis.
• Cite analyst bullets as primary facts — they're inference on top of filings. Treat them as the sell-side take, not the filing itself.
• Re-synthesize data already covered by kr-naver-insight (retail / independent views are a separate source role).

Pipeline

1. Read the extracted-text JSON from --input.
2. Compute consensus:
   • TP median / mean / min / max over reports with targetPrice != null.
   • Rating distribution (BUY / HOLD / SELL / null).
   • Count of reports in window, most recent publish date.
   • Keep only the latest per broker for the broker table (by publishedDate desc, tie-break on reportId).
3. Build broker table: broker | analyst | date | rating | TP | TP Δ vs broker's prior | 1-line summary (taken from the first bullet pulled out of the extracted text, or the report title if extraction is empty).
4. Pick recent reports: top --max-reports (default 10) by publishedDate desc.
5. For each recent report, extract 5–8 key-point bullets:
   • Split the .txt file on blank lines and newlines.
   • Score candidate lines by: contains '→', contains a numeric token, matches 목표주가 / 투자의견 / 실적 / 매출 / 영업이익 / 가이던스 / BUY / HOLD.
   • Take the top-scored lines up to 8, truncate each to 220 chars, quote verbatim. Do not rewrite.
   • Cap the total snippet length at --snippet-chars (default 1500).
6. Divergences:
   • Brokers whose TP is > 15% above or below the consensus median.
   • Brokers whose rating differs from the modal rating.
7. TP trajectory: if ≥ 6 dated reports with non-null TPs, bucket by calendar month, compute median per month, render a two-column list (YYYY-MM | consensus TP) and an ASCII sparkline.
8. Source-quality footer: extractionOk counts, skipped (auth / no PDF), warnings from the input meta.warnings.
9. Write the Markdown to --output.

Read references/output-format.md for the exact section order and the ASCII sparkline rule.

Operating Rules

1. Quote, don't paraphrase. Key-point bullets must be lifted verbatim from the extracted .txt. Synthesis belongs in the downstream memo.
2. Always include landing URLs so the memo writer can click through to the original report.
3. Never invent target prices or ratings. If the discover / extract index lacked them, the digest leaves the cell empty (—) and flags the broker in the source-quality footer.
4. Dates are ISO throughout (YYYY-MM-DD).
5. TP Δ is computed within broker (current minus the broker's immediately-prior TP in the window). If the broker has no prior, show —.
6. Stay deterministic: same input JSON produces the same Markdown.
7. Do not deduplicate across brokers for the recent-reports section. If the same broker published twice in a month, both rows should appear. The broker table deduplicates (latest only); the recent-reports section does not.

Scripts

summarize-reports.js

node skills/kr-analyst-report-insight/scripts/summarize-reports.js \
  --input .tmp/analyst-report-cache/extracted/066970/2026-04-18.json \
  --output analysis-example/kr/엘앤에프/analyst-report-insight.md

Optional flags: --max-reports N (default 10), --snippet-chars N (default 1500), --verbose.