article-level3

name: article-level3 description: Synthesize 3 recent L2 explanation articles into one L3 analysis (insight) article in Japanese via inductive reasoning — the agent-workforce equivalent of the GAS L3_BATCH/handleL3Create pipeline. Use when an editorial-stream agent must find the deeper principle connecting several ostensibly-unrelated explanations and project its implications. Published to the unified Notion Articles DB as Type=analysis, Author={agent_slug}, Status=ready.

article-level3

Synthesize 3 recent L2 explanation articles into one L3 analysis (insight) article in Japanese, by inductive reasoning — find the deeper principle that connects ostensibly-unrelated explanations and project its implications.

This is the agent-workforce counterpart of the GAS L3_BATCH / handleL3Create pipeline (newsletter/gas/src/Code.gs). The GAS path samples recent explanation rows, prompts Azure to induce a unifying principle, and writes an analysis-type row into the unified Notion Articles DB. This skill produces the same deliverable — a cross-source synthesis — attributed to the running agent so the analysis carries a byline on kohuehara.xyz.

It runs on the CCR execution model (the same pattern as article-level2 and Dario's feed-post): the binding is executor=claude-code-routine + scheduler=external/api, fired by wf-orchestrator-tick into the generic agent-runner routine (workforce/docs/routines/agent-runner.md). The routine composes your persona + this skill body; you generate the synthesis; then a bundled write script owns the Notion write — you do not hand-edit any file and do not open a PR.

One DB, one credential (apiKey only)

Both input and output live in the unified Articles DB, distinguished by Type (explanation = the L2 input pool, analysis = what you write). Only the Notion apiKey is a secret; the DB id is not (it's a constant in the scripts, mirroring newsletter/pipeline/normalize-categories.mjs). You need just one injected credential:

Instructions

1. Pick the L2 sample — run the picker, don't guess

Run pick-l2-sources.mjs. It reconstructs the GAS handleL3Batch sampling rule from Notion alone (no extra state store): a 14-day recent window, a fresh-entry gate (≥1 L2 created since the most recent analysis row), and reuse-avoidance derived from the source URLs stamped on recent L3 rows. It returns 3 L2 explanations to synthesise, the majority canonical A–E bucket, and the comma-joined source URLs to record.

NOTION_API_KEY="<credentials['notion.integration_token'].apiKey>" \
  node workforce/skills/article-level3/pick-l2-sources.mjs

• If it prints {"skip": true, …} — stop. Produce nothing this fire. No new L2 has arrived since the last analysis, or there are fewer than 3 recent L2s. Skipping (not calling the write script) is the correct behaviour, not an error.
• Otherwise you get { sources: [ {l2PageId, title, abstract, sourceUrl, category} ×3 ], canonicalCategory, sourceUrls }. Use the three {title, abstract} pairs as your evidence base. Ground every claim in those three sources — quote their figures, names, and findings; do not import facts they don't contain. (The abstract is each L2's lead; when a sourceUrl is reachable and you need a specific figure, you may fetch it for verification, but the synthesis is over the three L2s, not fresh reporting.)

2. Synthesise one Japanese analysis (3000–4000 字)

This is the GAS handleL3Create contract. Observe the concrete facts the three L2s report, then induce the single most-probable deep principle that unifies them, and build the article on that hypothesis. Not a summary of three articles — a new cross-cutting insight.

• Line 1: a # H1 — a concrete Japanese insight title (15–25 字) that names the underlying principle, not the topic. Name the shift, not the field ("担い手の交代" not "AIの影響"; "変化への対応" and "AIの可能性" are banned).
• 導入 (100–200 字): present the seemingly-unrelated facts from the three L2s, then declare "実はこれらは同じ原理で説明できる". Open with a ~50-字 thesis.
• 分析セクション (## × 2–4): each deep-dives the facts from a different angle, carrying concrete numbers, quotations, and examples drawn from the L2s — and naming which L2 (by title) each claim comes from. Use ### for sub-points.
• 共通原理の提示 (##, 500–800 字): "Why → So What". Develop, logically, the root principle that unifies the observed facts. Make the inductive step explicit ("これらの事実は〜を示唆している").
• 未来予測と示唆 (##, 500–800 字): the situational changes this principle implies, with their probability and grounds, and a practical takeaway for the reader.
• Do not append a bias-disclosure footer (or any byline boilerplate) to the body. Disclosure is carried by the Author metadata — rendered as the AuthorChip byline on kohuehara.xyz — the same policy feed-post already follows (Epic-011 §7 / Q9). In-body boilerplate duplicates that metadata, freezes a model id in prose, and was the trigger for the ML-006 deploy-gate false positive.

Hard rules (editorial integrity — C-1, fail loud — C-4)

• Synthesis, not summary. The job is to make three ostensibly-unrelated explanations cohere under one principle. If they're obviously about the same topic, find the non-obvious bridge; if you can't bridge them honestly, prefer a thinner-but-true principle over a forced one.
• Claim every insight back to a source. Each concrete claim must cite one of the three L2s by title (or its category). Don't introduce figures, companies, people, or dates absent from the three sources.
• Make it falsifiable. State the principle so a reader could disagree with it. An unfalsifiable "AIは世界を変える"-class platitude is a hollow analysis — rewrite it into a claim with an edge.
• No generic titles, no reviewer-voice hedges ("重要だ", "今後注目される"), no throat-clearing preambles. Objective, incisive tone.
• Output全文を日本語で.

Write the analysis — run the script, do NOT hand-edit any file

The page is written by a deterministic script, not by you editing JSON. You generate the judgment (the synthesis markdown); publish-notion.mjs owns the write (correct schema, properties, block conversion) by POSTing a new analysis page into the unified Articles DB with the injected integration token.

1. Write the full analysis markdown to a temp file (e.g. /tmp/l3-article.md) — a file, not a shell arg. The first line must be the # Title H1.

2. Write the abstract — a faithful 2–3 sentence lead (your opening thesis) — to a second temp file (e.g. /tmp/l3-abstract.txt). This populates the Abstract column.

3. Run (the script writes to the unified Articles DB — its id is a built-in constant, so only NOTION_API_KEY is needed). Pass the picker's sourceUrls and canonicalCategory verbatim, plus the free-form テーマ1 × テーマ2 theme you chose for the synthesis:

   NOTION_API_KEY="<credentials['notion.integration_token'].apiKey from your task>" \
     node workforce/skills/article-level3/publish-notion.mjs \
       --author "<agent_slug>" \
       --type analysis \
       --status ready \
       --body-file /tmp/l3-article.md \
       --abstract-file /tmp/l3-abstract.txt \
       --source-urls "<sourceUrls from the picker>" \
       --category "<canonicalCategory from the picker, e.g. B>" \
       --theme "<テーマ1 × テーマ2>"   # omit if you have none
   

4. Report the script's exit code:

   • 0 — page created. The row carries Author={agent_slug}, Type=analysis, Status=ready, plus Abstract, SourceURLs (the three source L2 URLs), and Category/CategoriesMulti (the canonical bucket + your theme tag). The GAS L4 batch generates the hero image and flips Status to published. Done.
   • 2 — W-1 editorial guard failed (empty/short body, LLM-artefact prelude, or a last line that looks cut off mid-content — the shared scripts/lib/truncation.mjs heuristic), or 401/403 auth (project credential bag misconfigured). Read stderr; do not retry blindly. A sentence ending wrapped in emphasis (*…。*) is a valid ending — if the guard trips, the body really is cut off; regenerate the ending.
   • 1 / 3 — bad args / missing H1 title, or Notion API / network error.

NOTION_API_KEY comes from your task's injected credentials["notion.integration_token"].apiKey — never read it from anywhere else, never hard-code it. The script re-runs the W-1 guards before writing, so a degraded body fails loudly rather than landing on the site.

The page lands directly in Notion. No PR, no human-approval gate. The existing GAS L4 batch picks up the row and publishes it to kohuehara.xyz; newsletter/pipeline/fetch-notion.mjs surfaces Author + Type into the front-end manifest so AuthorChip renders the byline. Recording SourceURLs is also what lets the next fire's picker avoid re-synthesising the same sample.

When NOT to use

• The picker returned {skip:true} — no fresh L2 input, or fewer than 3 recent L2s. Synthesising on a stale or too-thin sample produces a hollow analysis.
• You cannot find an honest bridge between the three sampled L2s — escalate or skip rather than forcing a fake unifying principle (C-1).
• A single source needs explaining faithfully — that's article-level2 (explanation), not this skill.