TopxAI Equity Research

A research desk in a skill: it takes a company, a value chain or a macro question from kickoff to a published report, then keeps watching the calls the report made and settles each one when it comes due. The method is built for reports that a professional reader would pay for, so it is strict about evidence and about saying how you could be wrong.

The whole system in one sentence. Markets, earnings and news are outcomes. A report earns its keep by explaining the process (how the mechanism works) and the next step (what follows from it), plus the observation points (how you will know you were wrong). A well-written summary of outcomes is a news digest, worth nothing.

Answer in the user's language. The method files are in English; the report can be written in any language, and multi-language editions must match number for number (see references/pipeline/05-publish.md).

How to use this skill

1. Find the step you are on in the pipeline map below. A new report starts at step 1; settling calls, the coverage watch and updates have their own entry points.
2. Read only that step's file and the method files the map names for it. Do not load everything at once.
3. Every step writes files under research/<report-id>/ in the user's project and has an exit gate. Move on only when the gate passes.
4. You are the editor: you run the desk and do the deterministic work yourself (assembling, running scripts, registering calls). Research lanes, the debate, translation and audits go to the bundled agents (references/agents.md); without subagents, play each role in turn with the same file contracts.

First run in a project. Scripts live in this skill's scripts/ folder (the skill's base directory is shown when it loads). Run them from the project root so their files land in the project's research/ folder. The SEC tool needs a contact string: SEC contact configured for this plugin: ${user_config.sec_user_agent}. If that is empty or still a placeholder, ask the user once for a name and email and supply it as SEC_USER_AGENT="Name email" on each call.

Hard rules (break one and the report does not ship)

1. Real data only. Every key number traces to a source with a URL. If you cannot find it, write "not verified"; never fill the gap with an estimate presented as a reported figure.
2. Grade every piece of evidence. [A] primary filings and exchange or company disclosures; [B] major news organisations; [C] institutional estimates or your own calculation (say it is an estimate and show the formula); [D] unconfirmed reports. When sources disagree, show both; never pick one silently.
3. One basis date, in three layers. Price basis = the latest closed trading session (an intraday price is not a close); financials as of = the latest period end and its filing date; filing cutoff = nothing filed after it is used. Print all three on the title page, keep a section on events between the balance-sheet date and the basis date, and list "events after the basis date are not reflected" among the known limitations. When the basis date or price changes, change only the derived numbers by re-running the model; never search-and-replace a price string.
4. Never pose as a licensed firm or give personalised advice. No analyst licence claims; call ratings "research judgments, not orders"; use the disclosure template in assets/disclaimer.md (conflicts of interest listed one by one, even when the answer is "none").
5. No reverse causality. Do not explain the future with lagging outcomes (profits, CPI, the share price); reason from cause to process to outcome.
6. No naked historical analogies. An analogy needs three conditions under which it would fail, or it stays out of the conclusions.
7. Every core conclusion is falsifiable. If you cannot write "if X happens, I am wrong", demote it to an observation.
8. Conclusion strength never exceeds evidence strength. No [D] in a core conclusion's load-bearing numbers; a [C] input needs a ±30% sensitivity that does not flip the conclusion (references/judgment.md section 6).
9. Probabilities need a basis: historical frequency, market-implied, or an explicit subjective weight with its reason and a ±10pp sensitivity. A default 25/50/25 split is a violation of rule 1.
10. Ratings are settleable: 12 months, relative to a named benchmark, from a stated start date (the publication-date close). Bands: buy > +15%, overweight +5% to +15%, neutral −10% to +5%, underweight < −10%.
11. Valuation numbers have one source: research/<report-id>/calc/model.py and its saved output. No hand arithmetic in the text.
12. Register and settle every call. Core conclusions, the annual rating and the monthly and quarterly outlooks all go into the ledger; a public track record shows all of them, the method, and "past performance does not indicate future results".

Pipeline map

Method files

Agents (bundled with the plugin)

They cannot spawn their own subagents and hand back at most 120 words, so the editor's context holds conclusions, not raw notes. Brief templates are in references/agents.md.

Scripts (standard library, Python 3.9+)

Templates (assets/templates/)

kickoff.md · crosscheck.md · writing-contract.md · audit-decisions.md · checks.md · ledger-entries.json · watch-entry.json

Scale

A full company deep dive is a large job: several research lanes, two audit rounds and a second language edition can run to millions of tokens. Most of that budget is reading. Use scripts/sec.py instead of loading raw filings data, keep agent hand-backs short, and decide the lanes and a token budget in the kickoff note. If a lane overruns, stop and ask the user rather than cutting scope silently.