huasanai/ndlq-weekly-summary

Niu Delinqi Weekly Course Summary

Goal

Reduce a long live stream to a formal summary that lets readers quickly grasp the main narrative while accessing every company mentioned in full.

Do not compress courseware into an abstract. The live stream transcript is the primary evidence for judging Niu's views, attitude, and importance; courseware is only used to correct terminology, numbers, headings, and material background. Reference summaries are only for structure and readability — do not inherit their facts or opinions.

"Comprehensive" does not mean equal weight: develop important themes in the body, and cover every named company in a full tally at the end. This way nothing is missed, and a hundred courseware slides do not drown out what Niu truly emphasized.

Before You Begin

1. Confirm inputs: live stream transcript, courseware, course date, and optional reference draft. If the user does not specify an output, create a Feishu Doc by default; the user may choose Markdown, DOCX, PDF, or another location.
2. If a source file is missing, proceed with the available materials but clearly communicate capability boundaries:
   • Transcript only: can summarize Niu's views, but confidence in terminology and number correction decreases;
   • Courseware only: can only produce a courseware summary, not "Niu's live stream opinion summary";
   • No reference draft: does not affect content extraction.
3. Read references/growth-loop.md, locate or initialize the cross-period state directory. Before each session, read current-context.md, jargon-glossary.json, and the most recent opinions; do not rely on conversational memory.
4. Default output is Feishu Docx: read config.json in the state directory and use the parent node specified; if no parent node exists, create the document in the personal document library. If the current Agent provides a Feishu Doc Skill/tool, read its operational rules before publishing; if no Feishu capability exists, save locally as Markdown and mark as pending publication. Always create a new document — do not overwrite old versions. Override default configuration when the user explicitly specifies another location or format.
5. If you need to verify changing stock codes, listing status, announcements, or prices, prioritize primary sources such as exchanges, regulatory bodies, and company announcements. External verification may only correct errors or supplement context, not fabricate Niu's views.
6. Fully read references/speaker-profile.md and use Niu's identity, investment style, and language habits to disambiguate, denoise, and identify speech functions. The speaker profile is a weak prior: it must not override evidence from the current live stream or be written as a fixed introduction in the formal output.

Mandatory Intermediate Artifacts

Do not write directly from a long transcript into an article. First create the following internal materials in the working directory; they do not enter the group's formal output:

1. Structure Index: opening, courseware walkthrough, Q&A, closing, and original paragraphs/timestamps.
2. Transcript Correction Table: original word, corrected word, basis, confidence level.
3. Topic Evidence Cards: opinion attribution, attitude, importance level, logic chain, numbers, and counter-evidence.
4. Full Company Ledger: every named listed company, unlisted company, and clearly identified security tool.
5. Course Identity Asset Inventory: methodology, book list/tools, changes to course pools such as Niu Ji 50, student jargon, and Niu's personal interest affiliations.
6. Opinion Revision Record: judgments explicitly reversed, downgraded, or upgraded during the current session or across sessions; the latest clear statement takes precedence, but the reason for the change must be preserved.
7. Current Session Growth Packet: new jargon added this session, opinion events, data snapshots, user corrections, and output address; record in the fixed run-packet structure.

When fields and judgment methods are needed, read references/evidence-model.md.

Workflow

1. Extract and Preserve Evidence Anchors

• Extract the full transcript text, preserving original paragraph numbers or timestamps.
• Extract courseware page numbers, titles, body text, and key chart information.
• Preserve the original text; do not overwrite the sole evidence directly.
• Match courseware to transcript by company name, proper nouns, numbers, dates, and teaching order; leave blank when uncertain, do not force a correspondence.

2. Clean Voice Transcript

• Remove non-informative fillers, mechanical repetitions, and recognition noise, but retain original text that expresses attitude strength or personal style.
• Correct directly when courseware and context uniquely confirm.
• When high probability but not unique, write _(suspected: ...)_.
• When confirmation is impossible and it affects conclusions, write _(original unclear)_; noise that does not affect the main narrative may be removed.
• Focus review on companies, people, stock codes, industry terms, units, years, prices, and orders of magnitude.
• Use the speaker profile to rank candidate words, e.g., buy-side/sell-side, Niu Ji 50, brokerage research, industry jargon, and long-term focus tracks; but direct correction is only allowed when courseware or context uniquely confirms.
• First check jargon-glossary.json in the state directory. Mappings marked user_confirmed can be used directly for correction; model-inferred or single-transcript guesses may only enter a pending queue, never pollute the official jargon library.

Do not smooth ambiguous or contradictory sentences into logically perfect conclusions for the sake of reading fluency.

3. Restore Teaching Order

Identify and preserve the semantic sequence:

Opening proactive judgment → Courseware walkthrough → Q&A landing on companies and prices → Closing review/supplement

Opening and Q&A usually better reflect Niu's true priority than courseware titles. When Niu skips a slide, revisits one, or explicitly says "no need to cover this," take the live behavior as definitive.

4. Determine Opinion Attribution, Attitude, and Importance

For each piece of important information, label the attribution: Niu's opinion, courseware opinion, questioner's statement, externally verified fact, or researcher commentary.

Unified attitude labels:

• 🔥 Niu Emphasizes: proactively raised, repeatedly circled back, or uses strong deterministic words;
• ✅ Niu Agrees: explicitly bullish, endorses, adds to a pool, or considers watchable;
• 🟨 Niu Reserved: directionally agrees but has doubts about valuation, timing, or delivery;
• ❌ Niu Rejects: explicitly considers expensive, logic unsound, buzzword riding, or fundamentals unsupported;
• 😄 Joke/Sarcasm: exaggerated returns, rotation jargon, or clearly informal judgments;
• 📄 Courseware View/Not Evaluated: only reads the title or repeats material, Niu does not express an opinion;
• ❓ Pending Verification: unclear transcript, ambiguous terminology, or important facts not yet confirmed.

A company may simultaneously hold "long-term agreement + short-term reservation." Do not compress into a single bullish or bearish stance.

Niu is skilled at proactive review and correction. When signals such as "I used to think / now I change" or "last week I said / this week I revise" appear, list the opinion change separately: old judgment, latest judgment, reason for change, and subsequent validation. The latest clear statement takes priority, but do not delete the context created by the old judgment.

For cross-period comparison, use the most recent valid opinion from current-context.md and mark the current session as: first appearance, continuation, strengthening, weakening, reversal, or condition change. Write "not updated this session" when there is no current evidence; do not pass off historical attitudes as current statements.

Importance levels:

• A-level: opening main narrative, repeatedly emphasized, clear price/valuation/ranking, Q&A revisits, opposes courseware, or forms a reusable framework;
• B-level: fully developed, supports the main narrative, provides industry screening or company comparison;
• C-level: only reads the title, explains what a company does, or contains no new judgment.

Write A-level content in full, B-level moderately, and merge C-level into quick coverage or just the company ledger.

5. Execute Course Identity Asset Hard Override

Normal importance must not override the following:

• Niu's methodology such as "Dao and Art," cyclical view, market view: independent heading, A-level treatment;
• Books, tools, indices, or ETFs explicitly recommended by Niu: full list, at least B-level;
• Additions or removals to/from Niu Ji 50, Niu Ji Index, and other course pools: independent heading, with removal, addition, reasons, and replacement logic placed together;
• Student jargon such as "Pioneer Map," "Big Cannon Hand," "Butterfly Fan": retain the original words, explain in a highlighted box or narrow table;
• Niu's personal investments, holdings, financing participation, or project affiliations: prominently mark interest affiliation.

"Niu Ji 50" is the name of the course stock pool, and does not imply exactly 50 stocks. Historical/non-latest stock pools can be used to correct abbreviations and discover course relationships, but they cannot prove current pool membership, nor can absence from a table negate Niu's current remarks.

6. Build the Full Company Ledger

Scan at least twice:

1. Scan the entire transcript by company name, alias, product name, and stock code;
2. Use the topic evidence cards to cross-check whether every company has been entered into the ledger.

Merge abbreviations, old names, and slips of the tongue, but retain candidates that cannot be uniquely confirmed. For each company, record:

• Formal abbreviation, code, market / whether unlisted;
• Appearances and reasons for mention;
• Niu's attitude and importance level;
• Core content and judgment logic;
• All prices, valuations, market caps, time windows, catalysts, and risks;
• Type of future stock price judgment, conditions, time horizon, and confidence level.

Classify stock price judgments into the following types to avoid mixing different terminology:

1. Clear target price;
2. Target market cap or valuation multiple, convertible but must show assumptions;
3. Directional judgment, e.g., "still has room," "high-risk," "fall first then see";
4. Conditional path, e.g., "if asset injection completes, then ...";
5. Current price / historical price reference, not a prediction;
6. Joke, sarcasm, or repeating someone else's target, not Niu's formal forecast;
7. No clear stock price judgment given.

When Niu has made no prediction, write "No clear stock price judgment given." Do not derive a target price for him. Detailed fields are in references/evidence-model.md.

7. Review Core Logic

Only comment on A-level and a few critical B-level topics. Check fixed four layers:

Did the event actually happen?
    ↓
How does it change industry supply-demand, cost, price, or discount rate?
    ↓
How large is the target company's exposure to revenue, profit, and cash flow?
    ↓
Do current valuation, capex, dilution, and time value support shareholder returns?

Commentary must answer: why A leads to B, what data is missing, and what would falsify it. Clearly label commentary as "Researcher Commentary" — do not mix it into Niu's original meaning.

8. Generate the Formal Output

Write the conclusion first, then expand in teaching order. Use the structure of references/output-template.md, and delete empty sections based on the week's content.

Core requirements for the formal output:

• Title uses "Niu Delinqi (Niu Zong)"; body uniformly uses "Niu Zong";
• Open with a three-minute overview giving the main narrative, directional ranking, and most critical targets;
• Core companies must clearly state code, price/valuation terminology, stock price judgment, conditions, and risks;
• Course pool adjustments, methodology, book list, and jargon must have prominent positions;
• End the body with the "four-column" full company ledger: Company (Code) | Attitude | Core Content | Future Stock Price Judgment;
• Do not reproduce courseware point by point, and do not write C-level material into long paragraphs;
• Do not include local paths, evidence anchors, internal method explanations, or large excerpts of paid courseware in the public output.

Target Typography Rules

First assign semantic labels to core targets, then use scripts/stance_markup.py for deterministic rendering:

• [[VERY_BULLISH:Company (Code)]]: Niu is very bullish, repeatedly emphasizes, core holding, or explicitly recommends; displayed as orange bold in Feishu;
• [[GENERAL:Company (Code)]]: generally bullish, neutral, reserved, merely discussed, or attitude not strong enough; displayed as bold in Feishu;
• [[VERY_BEARISH:Company (Code)]]: Niu is very bearish, explicitly advises avoidance, does not include, or says has no value; displayed as green bold in Feishu.

In the course-specific color scheme, green means "explicitly avoid," opposite to the common "green = recommend." Therefore the opening legend must explain. Only use orange or green for explicit strong attitudes; "good company but expensive," short-term caution, or no expressed opinion should all be only bold. When long-term strongly bullish but short-term cautious, orange may be used, along with text explaining both time scales; if attitudes conflict and cannot be merged, use only bold.

9. Adapt to Publication Carrier

• By default, generate Feishu-parsable Markdown, allowing mixing of Feishu inline XML color tags. Run stance_markup.py --format feishu before publication to ensure no semantic placeholders remain.
• Use Feishu's supported blockquote/callout; do not output raw <quote>...</quote>.
• For ease of PDF export, prefer tables of 2–4 columns; convert overly wide information into grouped cards.
• Do not use Mermaid diagrams that require horizontal scrolling; use vertical ASCII art, narrow tables, or plain bullet points for complex logic.
• If the user requests "new version," always create a new document; do not overwrite old summaries.

10. Pass the Quality Gate Before Delivery

Execute references/quality-gate.md item by item. If any high-risk item fails, first return to the evidence card or company ledger to correct, do not simply polish to conceal.

11. Complete One Growth Cycle at the End of Each Session

After the formal output passes quality inspection, generate this session's run-packet.json according to references/growth-loop.md, then execute:

python3 scripts/memory_tool.py ingest --state-dir <state directory> --packet <run-packet.json>
python3 scripts/memory_tool.py validate --state-dir <state directory>

This step appends opinion events, saves data timestamps, absorbs user-confirmed jargon, and rebuilds current-context.md. It does not automatically rewrite SKILL.md: single-session observations enter memory; procedural rules are only promoted when the user explicitly requests permanence, or when the same kind of issue appears at least twice, preventing drift with over-summarization.

Each "growth cycle" must leave at least one reusable increment; if there are no new jargon or opinion changes this session, still record "Opinion continued / not updated this session" and a data as_of to prove cross-period comparison was completed. If publication fails but the body has passed quality inspection, it may be recorded as publication_status=pending; do not write a successful run record if the body has not passed quality inspection.

Boundaries

• This is a course opinion summary and research aid, not real-time stock recommendations or a personal trading plan.
• Specific stock judgments must retain the live stream date and conditions; they cannot be repackaged as still-valid advice.
• Do not publish paid courseware original text, full transcripts, copyrighted courseware, or large portions of third-party research reports.
• Do not pass off external verification results as Niu's opinions, nor write researcher judgments as Niu's original words.