report

name: report description: "Cross-vault topic synthesis: reads existing notes across investing + sibling vaults and writes a narrative brief to investing/Reports/. Read-mostly — never modifies entity notes; --deepdive first repairs weak anchor notes. Use when the user asks to brief them on a topic, says /report TOPIC, wants a pointed read on something already in the vault, or asks for a report with a deepdive/research component. NOT for daily wrap-ups (/newsletter)."

/report — Cross-Vault Topic Synthesis

Default read-mostly synthesis skill. Pulls existing notes across all vaults that have content on a topic and produces a narrative brief. In normal mode, never edits entity notes; only writes the report itself plus the daily-note log line. With --deepdive, first run a bounded deepdive component to create or expand weak anchor notes, then synthesize from the improved vault state.

Usage: /report <topic> [--lens neutral|allocator|contrarian|what-changed] [--since YYYY-MM-DD] [--full] [--force] [--deepdive]

What this skill is NOT

• Not /deepdive by default — no note expansion, no SEC filings, no DB lookups, no chart generation. Default reports may use a cold framing search for the report body, but they do not write fresh research back into entity notes. If the primary note is missing or a stub, stop and suggest /deepdive unless --deepdive is set.
• Not /newsletter — that's today's short daily wrap.
• Not /story — that's the exhaustive "what is the story" map for everything touched in one daily note.
• Not /ingest — that processes a single source. This synthesizes existing vault content.
• Not a note expansion unless --deepdive is explicit. Normal mode never touches the entity notes it reads. --deepdive may edit source notes under the full /deepdive gates.

Hard rules

• Output saved to investing/Reports/YYYY-MM-DD-<topic-slug>.md. Always in-vault, always dated.
• [[wikilinks]] preserved in body — reports are printable per the repo printing rule.
• No bold in body. Headers only. Same voice rules as /newsletter.
• All four sibling vaults (geopolitics, Brazil, history, technologies) consulted on every run; report integrates whatever they have.
• Default mode writes only the report and the daily-note report line. --deepdive may also write entity/concept notes, stubs, charts, DB rows, and daily-note edit-log entries exactly as /deepdive requires.
• Reports are disposable. Re-run anytime; old reports are not load-bearing.

Phase 0 — Resolve the topic

1. Parse the input shape — classify before anything else. The word joining the topics matters:

   Input pattern	Behavior
   Single topic (CUDA moat)	Proceed with steps 2-5 below.
   <A> vs <B> or <A> versus <B>	Comparison report — one file, dual anchors. Slug keeps vs (brazil-carry-vs-us-ai-capex). Resolve both anchors in step 3; Phase 1 gathers ~7-8 notes per side; Phase 2 arc is the tension between them.
   Comma-separated (<A>, <B>, ...)	Sequential reports — run the full skill once per topic. One file + one daily-note line per topic. Do NOT synthesize across them.
   Ambiguous connectives (and, &, plus, +)	Stop and ask — "Comparison report (one file, juxtaposition is the point) or separate reports (N files)?" Do not guess.

2. Slugify the topic: lowercase, hyphenate. CUDA moat → cuda-moat. Keep the original case for display. For comparison reports, slugify the full phrase including vs.

3. Resolve in investing vault (primary):

   python scripts/check_before_create.py "<Topic>"
   

   Use the same resolver vault uses for entity matching. Catches aliases. For comparison reports, resolve each anchor independently; if either fails, stop and report which one.

4. Branch:

   • Exact match → use that note as primary. (Comparison reports require exact match on both anchors.)
   • Multiple candidates → ask user to pick.
   • No match in investing → check sibling vaults (grep -r on ~/obsidian/{geopolitics,brazil,history,technologies}/). If found in a sibling, ask user: "Primary note is in <sibling> vault. Generate report from that lens?"
   • No match anywhere → if --deepdive, run Phase 0.5 to create the anchor note, then continue. Otherwise stop, tell user, suggest /deepdive.

5. Stub gate — count substantive lines in primary note (exclude frontmatter, Quick stats, Related). If under 30 lines:

   • If --deepdive, run Phase 0.5 to expand the weak anchor(s), then re-check the stub gate.
   • If no --deepdive, stop. Output: "<Topic> is a stub (N lines of substance). Run /deepdive <Topic> first, re-run /report <Topic> --deepdive, or re-run with --force to synthesize what's there."
   • For comparison reports, apply the stub gate to both anchors; if either is a stub, use the same branch.

Phase 0.5 — Optional deepdive component (--deepdive only)

Run this phase only when the flag is explicit or the user clearly asks for "a report with deepdive/research first." Do not infer it from a broad topic or from normal curiosity.

Purpose: make the report synthesize from a vault state that is good enough to trust. This is a pre-report research pass, not a separate final deliverable.

1. Scope the anchors — for single-topic reports, deepdive the resolved primary anchor. For comparison reports, deepdive both anchors that are missing, stubby, or stale enough to distort the comparison. For comma-separated sequential reports, run this phase separately per topic.

2. Follow /deepdive gates for source-note edits — use the /deepdive skill as the controlling procedure for the entity type:

   • check_before_create.py before any new note.
   • Vault-search first, then web/source research.
   • SEC filings, DB setup, price freshness, chart generation, and price verification for public companies.
   • Funding-round and valuation work for private companies.
   • Concept structure for concepts; no actor template leakage.
   • Actor-note compliance checks, daily-note edit logging, stale-reference scan, concept extraction, and cross-vault gate.

3. Keep it bounded — the deepdive component must make the reportable anchor(s) substantive, not recursively repair the whole vault. Create or expand first-order notes only when a missing/stub entity is essential to the report's central argument. Default cap: primary anchor(s) plus up to 3 essential first-order supporting notes. Put the rest in the report's ## Gaps.

4. Respect DB and chart hard gates — never DELETE or UPDATE DB rows without explicit authorization. Never overwrite existing chart files or remove chart embeds without asking. Verify chart file sizes before embedding.

5. Resume /report from the improved state — after the deepdive component, re-run Phase 0 resolution and stub checks, then continue into Phase 1. Include newly created/expanded notes in sources_read; set deepdive: true in report frontmatter.

Phase 1 — Gather

Read in this order, capping at ~15 notes total to keep the synthesis tractable:

1. Primary note — full read.
2. Cross-vault links — parse the ### Cross-vault subsection of the primary's Related section. Resolve each obsidian:// URI to a real path and read the counterpart. These are the user-curated cross-vault siblings.
3. Related (1 hop) — follow wikilinks in the primary's Related section. Cap at 8 within investing.
4. Sibling-vault grep — even if no Cross-vault subsection exists, run a grep for the topic across all four sibling vaults:

   grep -rli "<topic>" ~/obsidian/geopolitics ~/obsidian/brazil ~/obsidian/history ~/obsidian/technologies
   

   Read the top 1-2 hits per vault that aren't already covered by Cross-vault links.
5. Backlinks — get ranked backlinks within investing:

   "/c/Users/klein/AppData/Local/Programs/Obsidian/Obsidian.com" vault=investing backlinks file="<primary-note>"
   

   Rank by mention count × section depth. Use top 5 for context. Show all under "Other vault references" footer.
6. Cold research pass (mandatory) — before Phase 2, run WebSearch on the concept itself, not on whatever was most recently ingested. Use domain-independent queries ("X market structure 2026 analyst consensus," "X regulatory framework 2026 academic," "X structural analysis CSIS/IEA/Wood Mackenzie"). Pull authoritative framings from agencies, academic work, law-firm annual reviews, and research houses. In default mode, use this only to calibrate the report body and Gaps; do not edit source notes. With --deepdive, merge this discipline into the Phase 0.5 source-note work as needed. See docs/research-workflow.md#cold-research-pass for the full discipline, failure modes, scope, and the test question. Skipping this step is the single largest source of bias in report output.

If --since YYYY-MM-DD is set, use git log --since="YYYY-MM-DD" -- <note-path> to identify what's been added; bias the gather to those edits.

Even with --since, the synthesis weaves in the full arc. Recent edits get foreground attention, but the analytical paragraphs should still situate them inside the pre-existing structure documented in the notes — earlier events, prior framings, structural context. A report that narrates only what changed since the cutoff is a changelog, not a synthesis. Same rule applies in stronger form when --since is not set: the report reflects the whole field around the topic, not the recently-touched slice.

Phase 2 — Synthesize

The narrative arc (per feedback_synthesis_voice.md):

central insight → constraint → counter-example → one-line

This is the only acceptable structure. If you find yourself writing "Background:" / "Current state:" / "Outlook:" — stop. That's the source notes' structure, not synthesis.

Discipline: the report must produce framing the source notes don't already express. If the synthesis is "the primary note rearranged," it failed. Ask: what do I see when I look at investing + history + geopolitics + tech + Brazil together that I don't see in any single note?

Lens behavior (--lens flag, default neutral):

Voice rules — baseline: docs/vault-note-guide.md → "Voice and Writing Standards" (analytical not editorial, no bull/bear or "the market isn't pricing X" claims, exact figures with sources, tensions over conclusions, no bold, no emojis). Report-specific: the reader is a macro-focused investor who has the vault context — sharp, not sterile.

Calibrate for high general knowledge, low subject-specific knowledge. The reader is an informed macro investor — assume they know sigma, beta, IRR, EPS, NII, ROE, CDS, ETFs, OPEC, shipping incoterms (CIF, FOB), off-take, Mitsui / Sumitomo / other major trading houses, G20 heads of state (Milei, Lula, Boric, Xi, Trump), basic commodity terminology. Do NOT gloss any of those.

DO gloss subject-specific jargon on first use in the report body, with a 3-10 word parenthetical. This does NOT mean dumbing down or cutting detail — keep every figure, every attribution, every nuance. It means a reader with strong macro context but no lithium-industry (or rare-earth, or Brazilian-fiscal, etc.) vocabulary can still parse the synthesis.

Examples of what to gloss:

• Industry-specific abbreviations most finance readers would not know: tonnes per year of lithium carbonate equivalent (LCE, the standard industry unit), 1.6nm A16 node (backside power delivery generation), WTI basis (Cushing vs Brent spread)
• Domain policy / political shorthand: "anti-involution" — Beijing's campaign against overcapacity-driven price collapses in strategic industries, Arcabouço Fiscal (Brazil's 2023 fiscal framework replacing the spending cap)
• Specialist geographic or geological terms: Puna Plateau salt flats, Smackover Formation (underground brine reservoir across Arkansas-Louisiana), CoWoS (TSMC advanced packaging)
• Niche pricing agencies, research houses, or indices: the pricing-agency Fastmarkets, Benchmark Mineral Intelligence, the Shanghai Metals Market (SMM) price
• Newly-introduced proper nouns for recent events: Nova Andino Litio JV (Dec 2025 Codelco-SQM partnership) on first mention; subsequent mentions just the name

Examples of what NOT to gloss:

• Statistical and market-structure terms: σ, beta, basis points, standard deviation, z-score
• Standard financial metrics: IRR, EPS, NII, ROTCE, CDS, NAV, AUM
• Shipping / trade incoterms: CIF, FOB, EXW, CFR
• G20-level political figures and central bankers: Milei, Lula, Boric, Trump, Xi, Powell, Lagarde
• Major well-known trading houses and banks: Mitsui, Mitsubishi, Glencore, Trafigura, Goldman Sachs
• Terms already glossed earlier in the same report
• Concepts that have their own [[wikilinked]] vault note (the wikilink itself is the pointer)

The test: if a finance professional who follows markets daily but doesn't know the specific subject would pause and wonder what a word means, gloss it. If they'd roll their eyes at the gloss, don't.

Cross-vault weave — when content from multiple vaults bears on a single point, integrate inline (don't section-by-vault):

The real today reflects a 1,100bp carry advantage [[Selic]] vs Fed funds, anchored by the [[Arcabouço Fiscal]] credibility regime — a fiscal architecture Brazil has only had for two cycles, against a five-decade history of regime collapses (History: Plano Real). The carry is the easy story; the durability of the regime is the harder one.

That paragraph pulls investing (rate differential), Brazil (fiscal framework), and history (regime durability) into one analytical thread. That's the test.

Phase 3 — Gap surface

Before closing the report, scan what you read for:

• Dead wikilinks in the primary note (entities mentioned but no note exists) — list under "Gaps."
• Stub notes referenced as if substantive — flag them.
• Cross-vault counterparts that should exist but don't — e.g., the investing note covers a topic with deep historical roots but there's no history vault entry.

The Gaps section is the feature that makes /report a vault tool, not just a prose generator. It tells the user what to research next.

Phase 4 — Write the report

Path: investing/Reports/YYYY-MM-DD-<topic-slug>.md

Confirm date: date +%Y-%m-%d. Use that exact date in filename and frontmatter. If file already exists for today and topic, append -2, -3, etc.

Format:

---
name: <Topic>
type: report
topic: "[[<Primary Note>]]"
lens: <neutral|allocator|contrarian|what-changed>
deepdive: <true|false>
generated: YYYY-MM-DD HH:MM
sources_read: N
tags: [report]
---

# <Topic> — <one-line theme>

[2-4 paragraphs synthesizing across vaults. Narrative arc, not section dump.
Inline `[[wikilinks]]` to entities. Inline `[History: X](obsidian://...)` to cross-vault.]

## What's new (only if --lens what-changed or --since)

[Chronological delta. Specific dates, exact figures.]

## Other vault references

- [[Note A]] — what it adds
- [[Note B]] — what it adds
- [Geopolitics: Note C](obsidian://...) — what the other lens adds

## Gaps

- [[Missing Entity]] — referenced in [[Primary]] but no note exists
- [[Stub Note]] — only N lines of substance; consider `/deepdive`

Length: target 400-800 words in body (excluding frontmatter and footer sections). Shorter is fine if the topic is narrow.

Final check before save:

• All [[wikilinks]] preserved (printable rule)
• No bold in body
• No bull/bear framing
• Exact figures, attributed
• Synthesis arc, not source-note rearrangement

Phase 5 — Daily note + chat output

1. Log to today's daily note under a ## Reports section (create if missing):

   ## Reports
   
   - 14:23 — `/report <Topic>` → `[[YYYY-MM-DD-<topic-slug>]]` (lens: <flag>, N sources, deepdive: <yes|no>)
   

   If --deepdive edited or created notes, also ensure the /deepdive-style ## Notes created/expanded and ## Edit log entries exist for those source-note edits.

2. Echo the report body to chat so the user reads it without opening the file. Mention the saved path in one line at the end:

   Saved to investing/Reports/YYYY-MM-DD-<topic-slug>.md. Disposable — re-run anytime.

Reports/ folder hygiene

• investing/Reports/ is the canonical location.
• Reports carry tags: [report] and type: report so they can be filtered out of compliance scans, vault-of-record gates, and cross-vault gates (none of those apply to disposable synthesis). Source notes edited by --deepdive still require their normal compliance gates.
• Reports are NOT entity notes — check_note_compliance.py should not run against them. If a hook trips on a report, exclude investing/Reports/ in the hook config.
• Old reports stay on disk indefinitely. The user manually clears investing/Reports/ when it gets noisy. No auto-expiry.

Failure modes to avoid

• Re-stating the primary note. If the report could be generated by reading just the primary note alone, the cross-vault gather added nothing. Re-do the synthesis or surface the actual cross-vault tension.
• Vault-by-vault sections. "Investing perspective:", "History perspective:" is concatenation, not synthesis. Weave them inline.
• Editorial framing. "This is bullish for X" is wrong. "The setup that would make X bullish requires Y; today Y is at Z" is right.
• Stealth deepdive. Never edit source notes unless --deepdive is explicit. Normal /report writes only the report itself plus the daily-note report line.
• Runaway deepdive. --deepdive prepares the anchor(s) for a better report; it does not chase every gap recursively. Essential first-order support only, then surface the rest.
• Dropping wikilinks. [[NVIDIA]] stays as [[NVIDIA]] in the output, never collapsed to "NVIDIA" or "Nvidia". The repo printing rule applies.