aiib-research-skill

AIIB Investment Research Skills

Three portable Agent Skills for AIIB-style developing-markets infrastructure investment research, usable in claude.ai (web + desktop) and Claude Code (terminal + web). A top-down deal-sourcing funnel in three stages:

Mode A  sector scan      → sub-sector economics + anchors          (context)
Mode S  company sourcing  → graph-expand anchors → PRIVATE candidates (the longlist)
Mode B  company dossier   → deep-dive each candidate, in sector context (the deep-dive)

• Mode A — aiib-sector-scan — input Country · Sector (e.g. "Indonesia · Renewables") → exhaustive sub-sector enumeration + cheap triage, a full 9-field A–I deep-dive on the most-investable few (deep-on-all is opt-in), plus the anchors (listed leaders, DFI investees, deal winners) that sourcing expands from.
• Mode S — aiib-company-sourcing — input a sub-sector (+ its economics) → a ranked list of mostly- PRIVATE, mandate-fit companies, found by graph-expanding anchors (competitors, suppliers, customers, co-investors, PE/DFI fund-siblings), fund-following, value-chain decomposition, and sweeping deal trackers / PPP pipelines / credit-rating lists / local-language sources.
• Mode B — aiib-company-dossier — input a company (or a Mode-S candidate) → a structured dossier, screened within its sub-sector economics: background + moat · forward guidance & catalysts · financials (margins, earnings quality, peer multiples) · valuation & bull/base/bear scenarios · AIIB-mandate alignment incl. ESG · key people & management quality · risk assessment.

▶ See it work first

Three real, live outputs (Claude Code, 2026-06-23) show the whole funnel in action — start here:

1. India renewables — sector scan (Mode A)
2. India solar — private-company pipeline (Mode S)
3. Sunsure Energy — investment dossier (Mode B)

They're one run end-to-end: A's anchors fed S; S's top private name (Sunsure) fed B. Comps are live yfinance, claims are provenance-tagged, and the work runs on cheap Sonnet subagents.

Design principles

• Provenance-tagged. Every claim is marked 🟢 web-sourced / 🔵 connector / ⚠️ training-unverified, so unverified figures are visible. Built for a web-search floor: with no network it still runs, fully ⚠️-tagged under a "no live sources" banner.
• Mandate-anchored. Each skill bundles references/aiib-mandate.md — AIIB's 6 sectors + 4 thematic priorities + 2030 targets, sourced from aiib.org — so mandate fit is cited, not hallucinated.
• Subagents where available, never required. On Claude Code the skills delegate each sub-sector / company to a parallel subagent (.claude/agents/); on the claude.ai chat app (no subagents) they run sequentially. Same output either way.
• Tiered free data, web-search floor. When a code tool + network are available, the skills pull free structured financials directly — yfinance (broad international coverage), SEC EDGAR, stooq, or OpenBB's free providers — via the bundled scripts/fetch_financials.py, tagged 🟢; otherwise they fall back to web search, and finally to ⚠️-tagged training knowledge. No API keys required.

Data availability by surface (important)

The structured-data tier needs a code-execution sandbox with network access. That differs by surface, so the same skill produces richer or thinner financials depending on where it runs — by design, it degrades gracefully and the provenance tags always show which tier was used:

Takeaway: for the fullest, numbers-from-live-feeds dossiers, run on Claude Code. On the claude.ai chat app it still works end-to-end — it just leans on web search rather than direct data pulls when the sandbox has no network. Either way, every figure is provenance-tagged so you can see what it relied on.

Platform support

The research methodology is cross-platform; the packaging and the live-data tier are Claude-native. Verified against vendor docs (Anthropic + OpenAI, early 2026):

For ChatGPT, the methodology is delivered as a Custom GPT running on web search — see ports/chatgpt-custom-gpt/ (instructions + knowledge files + setup). The in-sandbox data pulls (scripts/fetch_financials.py) don't run there, because "the Python environment used for data analysis cannot make external web requests or API calls" (OpenAI docs).

Credits & licenses

• This repo: MIT (see LICENSE).
• The fundamental + qualitative analysis framework adapts claude-equity-research (MIT), reframed for AIIB-style direct/PE infrastructure investing — the public-equity trading apparatus (technical analysis, options flow, 12-month price targets) is intentionally omitted.
• OpenBB is AGPL-3.0 and is not bundled here; the skills only invoke it at runtime if the user already has it installed. yfinance (Apache-2.0) and SEC EDGAR cover the same free ground, so they are the default.

Install

claude.ai (web / desktop) — Pro, Max, Team, Enterprise (code execution enabled)

Each skill folder is self-contained. In Settings → Capabilities → Skills, upload each of:

• .claude/skills/aiib-sector-scan/
• .claude/skills/aiib-company-sourcing/
• .claude/skills/aiib-company-dossier/

(Subagents in .claude/agents/ are not used on claude.ai; the skills detect this and run sequentially.)

Claude Code (terminal or web)

Clone this repo and open it (or point a Claude Code web session at it). The skills in .claude/skills/ and subagents in .claude/agents/ are picked up automatically — subagent fan-out works here.

Usage

Mode A:  "Research Indonesia · Renewables for AIIB."      →  sub-sector economics (triage + deep top few) + anchors
Mode S:  "Find private geothermal developers in Indonesia." →  ranked list of mostly-private candidates
Mode B:  "Build a dossier on <company>."                  →  7-section investment dossier

Typical flow: A (context + anchors) → S (source the private companies) → B (dossier each).

See real runs (live, 2026-06-23, clean Claude Code session):

• examples/india-renewables-sector-scan.md — Mode A: 12 sub-sectors, 2 waves of 6 Sonnet workers, primary-sourced, comps verified live via yfinance (46 🟢 / 2 ⚠️).
• examples/india-solar-company-sourcing.md — Mode S: 121 companies (96 private) over 3 iterative rounds, 13 source-expander workers (zero recursion), deterministic dedup, ranked deal pipeline with PE/DFI backers.
• examples/sunsure-energy-dossier.md — Mode B: a 7-section dossier on a private Tier-1 name from the sourcing list, screened within the sub-sector economics; company financials from rating rationales + DFI disclosure, listed-peer comps via yfinance.

These three are one funnel run end-to-end: A (India renewables) → S (solar private pipeline) → B (Sunsure dossier).

Controlling cost / the Sonnet subagents (Claude Code)

A skill can't force Claude to delegate or pick a model — that's the model's call + your session config. Two levers that do work, verified live:

1. Run the session on Sonnet — claude --model sonnet (or /model sonnet interactively). Then the whole run — orchestrator and any subagents — is Sonnet. Cheapest, always works, and Sonnet is plenty for web research + filling the template.
2. Tell it to delegate, in your prompt — add e.g. "delegate each deep sub-sector to a subsector-researcher subagent" (Mode S: source-expander; Mode B: company-dossier-researcher). Verified: this makes it spawn the purpose-built, non-recursive agents on Sonnet while the orchestrator stays on your session model — so the heavy searching runs cheap. Without this nudge, on a small/triaged scan the model often just does it inline (fine, but on your session model).

Either way the run is bounded — triage + round caps + per-worker search budgets + non-recursive agents mean it can't run away, whichever model it lands on.

Disclaimer

• Not affiliated with AIIB. This is an independent, unofficial tool. It is not affiliated with, endorsed by, sponsored by, or sanctioned by the Asian Infrastructure Investment Bank (AIIB). "AIIB" and related names and marks are the property of their respective owner and are used here only for descriptive reference (nominative fair use).
• Source of mandate data. The AIIB sector and thematic-priority information in references/aiib-mandate.md was sourced from AIIB's public website (aiib.org) as of 2026-06-22 and may be incomplete or out of date. Consult aiib.org for authoritative, current information.
• Not investment advice. Output is AI-generated research for informational purposes only. It may contain errors, omissions, or fabricated figures — that is exactly why every claim is provenance-tagged (🟢 web / 🔵 connector / ⚠️ training-unverified). It does not constitute investment, financial, legal, or tax advice. Verify every figure independently and do your own due diligence before relying on anything here.
• No warranty. Provided "as is", without warranty of any kind. Use at your own risk; the authors accept no liability for any decision made on the basis of this tool's output.