banker-memo（银行家备忘录）

Banker Memo

Banker-grade research memo from raw-data/ snapshots — prompt-driven, not template-driven.

What this skill does

Takes a deliverable dir that already has raw-data/*.json (produced by the CN Phase 3.5 pipeline) and drives an agent through a senior sell-side analyst framework to produce:

1. analysis.md — 8-section banker research memo (2500-4000 words, peer-benchmarked, with explicit data flags)
2. slides-outline.md — PPT blueprint (10-15 slides, content-driven — NOT a fixed 8-slide template)
3. data-provenance.md — every hard number traceable to a raw-data filename stem

Why prompt-driven (not Python template)

The 0.9.4-era build_deck.py used an 8-slide Python template that stamped the same structure onto every company. Problem: the result is a data dashboard, not banker analysis — no industry context, no peer benchmarking, no SOTP reasoning, no "4C's" credit framework, no banker opinion.

Prompt-driven flips this: the agent (which has read thousands of real sell-side reports during training) writes the narrative; the Python layer only does orchestration (dispatch + gate audit).

Best-fit cases

• User explicitly asks for 投行级 / banker-grade research
• Deliverable already has raw-data/ from cn-client-investigation Phase 3.5
• Target is CN (A-share / H-share / non-listed with unified social credit code)
• The previous Python-templated output felt shallow / repetitive / stat-card-heavy

When NOT to use

• raw-data/ is missing — run cn-client-investigation Phase 3.5 first
• Target is US-listed — this skill is CN-focused (industry framework / 4C's assume CN credit context)
• Just want a quick fact-sheet — use strip-profile skill instead

Workflow

Step 1 — Pre-flight

Verify raw-data/ has:

• At least one aigroup-market-mcp-*.json file (for listed) OR a primematrix-basic_info.json (for non-listed).
• Listed companies SHOULD have 5 Tushare files (basic_info / company_performance / stock_data / daily_basic / income) for the full framework.

Step 2 — Build parameterised prompt

Use references/banker_prompt_template.md and substitute placeholders:

• {ts_code} / {name_cn} / {industry} / {raw_dir} / {out_dir}
• {file_list} — discovered from raw_dir
• {uscc} — parsed from the primematrix filename prefix

A helper scripts/build_banker_prompt.py does this. See references/banker_prompt_template.md for the canonical prompt body.

Step 3 — Dispatch agent

Run openclaw agent --agent main --thinking high --json --timeout 600 --message "$(cat prompt.md)". Typical runtime 3-6 minutes per company with high-thinking.

Step 4 — Validate output

Run the standard 7-gate validate-delivery.py --strict-mcp on the output dir. The banker memo obeys the same authenticity / provenance rules as Python-templated outputs — gates unchanged.

Step 5 — PPT compile

Run build_deck.py (v0.9.6+) which now reads slides-outline.md for layout cues. Slide count = whatever the agent planned (10-15 typical, not a fixed 8).

Hard constraints enforced by the prompt

1. Every hard number cites a raw-data filename stem, format X 亿元（src: income） or Y%（src: company_performance）
2. Wind / 万得 / 同花顺 / Bloomberg / 彭博 forbidden (redundant with source_authenticity_check gate)
3. No vague numbers — "XX 亿元左右" / "大约" / "估计" must be tagged [EST] with reasoning
4. QoQ deltas in pp units (avoids HARD_NUMBER regex false-positives)
5. Peer comparison numbers tagged [EST, per sector consensus] — never tied to a specific forbidden source

Worked example (0.9.5 smoke test)

BOE 000725.SZ run on 2026-04-20:

• Input: ~/deliverables/bj-smoke-v2/000725_sz/raw-data/ (6 JSON files already fetched)
• Agent high-thinking runtime: ~5 minutes
• Output analysis.md: 14.8 KB (vs 1.6 KB from v0.9.5 Python template)
• Agent self-flagged Data Flag 1: income-derived 净利率 2.68% vs company_performance 2.09% — 0.59pp口径差异 → must verify pre-credit decision
• Agent self-flagged Data Flag 2: company_performance YTD 累积 ≠ Q4 单季, computed +0.74pp Q4 NPM change from diff
• Peer comp: 5 companies (TCL 科技 / 维信诺 / LG Display / 群创光电) all tagged [EST, per sector consensus]
• Valuation: SOTP (LCD PB 0.8-1.0x + OLED PS 1.0-1.5x) → 合理 PB 1.0-1.3x; target price 3 scenarios (3.5-4.0 / 4.2-4.5 / 5.0-5.5)
• Credit view: 4C's with specific numbers + 授信额度 50-80 亿 / 期限 1 年 / 利率 LPR+60-130bp / 增信要求 (设备抵押 60% + 应收质押 70%) / 财务承诺 (负债率<55%, 利息覆盖>2x)
• Slides outline: 12 slides (not fixed 8) — includes section dividers + bar chart (revenue/profit) + line chart (quarterly ROE) + valuation scenario table + 4-color risk card
• validate-delivery --strict-mcp — all applicable gates PASS

Output standard

Quality checklist

• [ ] 8 sections all present (ES / Profile / Industry / Financial / Peer / Valuation / Risk / 4C's)
• [ ] Every \d+(亿元|%|元|倍) in analysis.md has a provenance row
• [ ] Peer comparison has ≥3 companies, all tagged [EST, per sector consensus]
• [ ] Valuation section has ≥2 methods (relative + SOTP or DCF) + 3 scenarios
• [ ] Risk section uses a table with severity levels (not a bullet list)
• [ ] 4C's section gives a specific credit conclusion (额度 + 期限 + 利率 + 增信 + 财务承诺)
• [ ] slides-outline.md specifies layout type per slide (card / table / chart / divider)
• [ ] validate-delivery.py --strict-mcp OVERALL PASS