Reading-the-Managers
13F · 13D/G · N-PX · LETTERS · TRANSCRIPTS
Reading the Managers
A working method for hedge-fund-manager research in Claude: how to build a manager file from the disclosure record, what that record structurally cannot tell you, and the prompt discipline that keeps the output citable rather than plausible.
Kun Xia Research · September 2026
CONTENTS
You are reconstructing a portfolio you will never see premise
What a 13F structurally cannot tell you constraint
The form stack, and what each one adds sources
The evidence bar, stated in the prompt discipline
Six steps, in this order method
Five prompts that carry the workflow templates
Three runs worth having in practice
How this goes wrong guardrails
What to run, and when cadence
Where this sits in your setup tooling
PREMISE
You are reconstructing a portfolio you will never see
Manager research is an inference problem dressed up as a data problem. The holdings file looks like a fact, so the whole exercise drifts toward tabulating it. The actual question is narrower and harder: what does this manager believe, how much are they risking on it, and has that changed? The filings are evidence toward that question, not an answer to it.
Two failure modes account for almost every bad output.
Fabrication. Asked what a manager said about a name, a model will produce a fluent, well-argued, entirely invented paragraph in that manager’s house voice. It is the single highest-risk task in this whole workflow, because the output is unfalsifiable unless you demand a link.
Over-reading the file. A 13F is a long-only, US-listed, quarter-end snapshot published with a legal lag. Treating it as “the portfolio” produces confident nonsense about managers whose real book is half short, half swaps, and half offshore.
Everything below is built to make those two failures structurally hard rather than merely discouraged.
CONSTRAINT
What a 13F structurally cannot tell you
Put this list into the prompt, not just into your head. A model that has been told what the data omits will hedge correctly; one that hasn’t will fill the gap with narrative.
BLIND SPOTS IN FORM 13F-HR
| GAP | CONSEQUENCE FOR YOUR READ |
|---|---|
| Longs only | Shorts are invisible. A “long” can be the hedged leg of a pair, a merger-arb stub, or an option overlay. Long-short managers are the ones most often misread as directional bulls. |
| Section 13(f) securities only | No cash bonds, no non-US-listed lines, no FX, no commodities, no private positions. A manager whose thesis lives in the capital structure looks like they own nothing. |
| Swaps and total return swaps excluded | Synthetic exposure does not appear. This is the Archegos gap and it has not been closed. |
| Quarter-end snapshot, 45-day lag | By the time the next filing lands, the oldest data is roughly 135 days old. Anything round-tripped inside the quarter never existed as far as the file is concerned. |
| No cost basis, no trade dates | You cannot compute the manager’s entry, their P&L, or whether they added into strength or weakness — only the endpoint delta. |
| Options on a share-equivalent basis | A large “put” line may be portfolio insurance, a financing structure, or a bearish view. The form does not distinguish, and aggregators frequently mislabel it. |
| Confidential treatment requests | A manager building a position can ask the SEC to omit it, then disclose later by amendment. Positions that read as “new this quarter” may be several quarters old. |
| Manager-level aggregation | One filing can combine several funds, sleeves, and separately managed accounts with different mandates. Portfolio-weight math on the aggregate can be meaningless for any individual vehicle. |
| $100m reporting floor | Only managers with at least $100m in 13(f) securities file at all. Emerging managers — often the interesting ones — are entirely absent. |
Practical rule: a 13F supports claims about direction and relative size of disclosed long equity exposure. It does not support claims about conviction, timing, hedging, or the manager’s actual net view. Ask Claude to label every claim by which of those two buckets it falls into.
SOURCES
The form stack, and what each one adds
13F is the worst of these documents and the one everyone uses. The signal density runs roughly in the opposite order to the popularity.
DISCLOSURE SOURCES RANKED BY WHAT THEY REVEAL ABOUT INTENT
| SOURCE | CADENCE | WHAT IT UNIQUELY GIVES YOU |
|---|---|---|
| Investor letter | Quarterly or annual | The thesis in the manager’s own words, with the reasoning and usually the exit conditions. The only source that tells you why. Highest value, least structured, hardest to source legitimately. |
| Schedule 13D | Within 5 business days of crossing 5% with intent; amendments within 2 business days of a material change | An activist thesis, often with an exhibited letter to the board and a stated plan. Effectively a public investment memo with a filing date. |
| Schedule 13G | Passive: 5 business days after crossing 5%. Qualified institutions: 45 days after quarter-end (5 business days after month-end above 10%). Amendments quarterly, faster on 5% and 10% moves | Timely notice that someone crossed a threshold — much fresher than 13F, and the amendment triggers tell you when they kept buying. |
| Form 13F-HR | 45 days after quarter-end | The complete disclosed long US equity book. Best used as a cross-check and a diff engine, not as a thesis source. |
| Form N-PX | Annually, due 31 August, covering 1 July–30 June | How the manager actually voted, including say-on-pay. Cheap, ignored, and the fastest way to test whether a stated governance view is real. |
| Form NPORT-P | Monthly filing; public portions published with a lag | Registered-fund holdings at higher frequency than 13F, including debt. Only covers the ’40 Act vehicles, and the publication cadence is under active SEC rulemaking — verify before building a process on it. |
| Conference decks and transcripts | Event-driven | Sohn, Robin Hood, Capitalize for Kids, Value Investing Congress. A full thesis with slides, dated and attributable. |
| Form ADV Part 2 | Annual | Strategy description, AUM basis, affiliates, custodian, disciplinary history. The entity-resolution document. Read it once per manager, then never again. |
| Form SHO | Not yet live — compliance date extended to February 2028 | Aggregate monthly short positions above thresholds. When it arrives it partially closes the short blind spot. Do not build a workflow on it yet. |
Outside the US, the equivalent trigger disclosures are the real-time source: UK major-shareholding notifications (TR-1, 3% then each 1%), EU transparency filings via national regulators, and Japan’s 5% large-shareholding reports. Non-US managers frequently file none of the US forms at all, so a ticker-first search will silently miss them.
DISCIPLINE
The evidence bar, stated in the prompt
The single most effective intervention: make the citation requirement a hard filter rather than a preference, and make staleness type-specific. A model told “cite your sources” will comply loosely. A model told “rows lacking a URL and a date are dropped before rendering” behaves differently.
TIERS, AND THE STALENESS CLOCK THAT RUNS ON EACH
| TIER | WHAT QUALIFIES | GOES STALE AFTER |
|---|---|---|
| T1 Primary | SEC and national-regulator filings; the fund’s own site (letters, factsheets, decks); official conference recordings; the fund’s own account or podcast | 13F: 2 quarters. 13D/G: 6 months. Letters: 12 months. Conference talks: 18 months. |
| T2 Aggregator | WhaleWisdom, Dataroma, HedgeFollow, 13F.info, Bamsec, Fintel — fine for holdings, penalise one confidence level | Same clock as the underlying document. Swap in the primary URL for anything thesis-related. |
| T3 Press | Named reporter, named masthead, datestamp: FT, WSJ, Bloomberg, Reuters, Barron’s, Institutional Investor | 9 months. Useful for activist news and interviews; too thin for thesis content. |
| T4 Excluded | Anonymous Substacks, X and Reddit, AI-generated summaries, unverifiable newsletters, stock-picker YouTube | Never usable. Not a quality judgement — a verifiability one. |
Two rules follow from the table and are worth restating verbatim in every prompt: quote or tag — direct quotes in quotation marks with an inline link, everything else explicitly marked “paraphrased from [source, date]”; and discussed is not owns — letters routinely name comparables, passed-on ideas, and exited positions.
METHOD
Six steps, in this order
The order matters because each step gates the next. Skipping to step four — which is what “what does Ackman think about X” does — is how you get fabricated letters.
01 Resolve the entity before anything else
Which legal filer, which CIK, which vehicles roll into it, which strategy, what AUM basis, and — critically — which parts of the book never appear in US filings. Managers rebrand, spin out, and split filers; a name is not an identifier.
Output: a header block — filer name, CIK, vehicles covered, strategy, AUM, known non-13F exposure, letter availability.
02 Build the holdings history, not the latest quarter
Pull eight to twelve quarters. One quarter tells you nothing; the shape of a position over two years tells you whether it is a core holding, a build, a fade, or a trade. Store it as a table you can re-derive, with the filing URL against every quarter.
Output: quarter × position matrix — shares, value, portfolio weight, rank.
03 Diff the quarters, and separate trades from price drift
Classify each line as new, add, hold, trim, or exit on share count. Weight and value move with the market; only share count is a decision. A position that doubled in weight while share count fell is a manager trimming a winner, which is the opposite of what the value column implies.
Output: ranked change list with a materiality floor — say, moves above 0.5% of portfolio or 20% of the line.
04 Go find the words
Only now: letters, 13D exhibits, conference decks, interviews, earnings-call appearances, N-PX votes. Search the manager’s own domain first. If nothing substantive exists, say so — plenty of managers file quarterly and never write a word, and a stub that admits it beats an invented thesis.
Output: one thesis card per position that has commentary, every claim quoted or tagged.
05 Reconcile the words against the money
This is where the work actually pays. Where does the letter’s emphasis match the book, and where does it not? A manager writing at length about a 0.8% position and silent on their top holding is telling you something. So is a stated multi-year thesis attached to a line that halved.
Output: a short list of confirmations and contradictions, each with both sources attached.
06 Write the manager file and set the tripwires
A durable markdown file per manager: identity, strategy, current book, thesis cards, words-vs-money notes, open questions, update log. Then define what would change your read — an exit from a top-five line, a 13D, a new letter — and pin the refresh to the filing calendar rather than to your curiosity.
Output: managers/[name].md, plus a dated next-review line.
TEMPLATES
Five prompts that carry the workflow
Each one front-loads the constraint. Fill the brackets and run them in order; the later ones assume the earlier output is in the conversation or on disk.
01 — ENTITY RESOLUTION
Profile the investment manager [MANAGER NAME] as a filer, not as a story. Return only what you can source, with a URL and date on every line:
- Legal filer name(s) and SEC CIK(s); note any predecessor or affiliated filers
- Vehicles that roll into the 13F (funds, sleeves, SMAs) and whether the filing aggregates mandates with different objectives
- Strategy, from Form ADV Part 2 or the firm’s own materials — long-only, long-short, activist, event, concentrated, etc.
- Latest reported AUM and the basis for it (regulatory AUM vs 13F value)
- Non-US or non-equity exposure that would never appear in US filings
- Whether they publish letters, where, and how far back the archive goes If a field cannot be sourced, write "unknown". Do not infer from firm name, strategy label, or what similar managers typically do.
02 — HOLDINGS HISTORY AND QUARTER DIFF
Build the disclosed long equity history for [MANAGER] over the last [8] quarters, from 13F-HR filings. Table: quarter | ticker | issuer | shares | reported value | % of portfolio | rank | filing URL. Then produce a change table for the most recent quarter versus the prior one. Classify each line NEW / ADD / HOLD / TRIM / EXIT strictly on SHARE COUNT, not on value or weight — value moves with price and is not a decision. Where weight rose but shares fell, flag it explicitly as price drift, not accumulation. Apply a materiality floor: only show moves above 0.5% of portfolio value or 20% of the existing line, whichever is smaller. List everything else as a count. State up front: this is disclosed long US-listed equity only. Do not characterise it as the manager’s portfolio or net position.
03 — THESIS EXTRACTION WITH QUOTE DISCIPLINE
For each of these positions held by [MANAGER] — [TICKERS] — find what the manager has publicly said, and extract a thesis card. Search in this order: the firm’s own website and letter archive, SEC filings including 13D exhibits, dated conference presentations and transcripts, then named-byline financial press. Card fields: thesis in one sentence | drivers cited | risks acknowledged | stated mispricing, if any | position sizing context | source list. Rules, applied strictly:
- Every substantive claim is either a direct quote in quotation marks with an inline link, or a paraphrase tagged "paraphrased from [source, date]".
- Anything without a working URL and a date is dropped, not flagged.
- A mention in a letter is not evidence of ownership; label those "discussed".
- If no substantive commentary exists for a position, write a stub saying so. Do not reconstruct a thesis from the firm’s philosophy page.
04 — WORDS VERSUS MONEY
Using the holdings history and the thesis cards already produced, reconcile what [MANAGER] says against what they did. Three lists, each line citing both the filing and the commentary:
- CONFIRMED — stated view and position change point the same way.
- CONTRADICTED — stated view and position change diverge (thesis reiterated while the line was cut; conviction language on a sub-1% position).
- SILENT — top-ten positions with no public commentary in the window. For each contradiction, give the most likely benign explanation alongside the adverse one. Do not resolve it — flag it as an open question with the specific disclosure that would settle it.
05 — CROWDING CHECK ON A NAME
For [TICKER], map which active fundamental managers hold it and what they say. Exclude passive index sponsors entirely (Vanguard, BlackRock at firm level, State Street, Geode, Northern Trust, Schwab, DFA-as-beta) — their weight is index mechanics, not opinion. Include a manager only if the position is at least 2% of their portfolio, with three exceptions: a 13D filer or public letter to the board; a new position discussed in their latest letter; a concentrated fund of 20 or fewer names. Output: holder table with status (current / historical / discussed only), each row carrying a source URL, date, tier and freshness label. Then the driver map — which fundamental driver each holder cites, tagged to a specific quote. A driver appears only if at least one named investor is on record citing it. Then answer three questions plainly: where do these managers agree, where do they split on framing, and is anyone on record saying what the market is getting wrong. If nobody framed it as a mispricing, leave that section empty.
IN PRACTICE
Three runs worth having
| RUN | THE ASK | WHAT IT DOES |
|---|---|---|
| Cold profile | “Build me a file on [manager] before I meet them.” | Full six steps on one manager, eight quarters back. The payoff is step five: the gap between the letter and the book is what you actually ask them about. Chain: 01 → 02 → 03 → 04 → manager file. |
| Quarterly sweep | “13Fs dropped. What changed in my tracked list that touches my coverage?” | Prompt 02 across fifteen managers, intersected with your coverage universe, filtered to material moves. Ninety percent is noise; the run exists to surface the ten percent. Schedule it for the day after each deadline. |
| Pre-initiation crowding check | “Who else is in this name, and what do they see?” | Run before you write, not after. Tells you whether your differentiated view is differentiated, and where the crowded exit is. Deliberately holder-first: surfaces arguments you would otherwise spend a week reconstructing. |
GUARDRAILS
How this goes wrong
FAILURE MODES AND THE COUNTERMEASURE THAT ACTUALLY WORKS
| FAILURE | HOW IT SHOWS UP | COUNTERMEASURE |
|---|---|---|
| Invented letter content | A fluent paragraph in the manager’s voice, with no link — or a link to a real letter that does not contain the quote | URL-and-date-or-drop, enforced as a filter. Spot-check one quote per card against the source yourself. |
| Aggregator restated as primary | WhaleWisdom numbers presented as “per the 13F” | Require the tier label on every row. Aggregators are fine for holdings; cite them as aggregators. |
| Price drift read as a trade | “Increased their position by 40%” when share count was flat | Classify on share count only. Make the prompt say it. |
| Copy-trading logic | A recommendation to follow a manager into a name on 45-day-old data | Use the file as an idea source and a crowding gauge, never as a signal. The lag alone disqualifies it. |
| Share class and listing confusion | An ADR position read as ownership of the ordinary; A vs B vs C classes merged | Verify CUSIP and listing per row. Non-trivially common with dual-listed European names. |
| Survivorship in the manager list | Only the famous and currently-performing managers get tracked | Fix the tracked list in advance and keep the ones that go cold. Their exits are data too. |
| Stale evidence as current | An 18-month-old letter quoted in the present tense | Type-specific staleness labels, and automatic demotion from “owns” to “owned” when the freshest source ages out. |
| Compliance drift | Scraping paywalled letters, or working from material a manager shared privately | Public, verifiable sources only. If your reader cannot open the link, it does not go in the file. |
CADENCE
What to run, and when
Manager research rewards a calendar and punishes curiosity-driven refreshes. Pin the sweep to the deadline, not to the news.
2026 FILING CALENDAR — VERIFY EACH AGAINST EDGAR BEFORE RELYING ON IT
| FILING | DATE | RUN |
|---|---|---|
| 13F — Q4 2025 | 17 Feb 2026 (passed) | Full-year diff; the most informative quarter because annual letters land nearby |
| 13F — Q1 2026 | 15 May 2026 (passed) | Quarterly sweep |
| 13F — Q2 2026 | 14 Aug 2026 (passed) | Quarterly sweep, paired with N-PX below |
| N-PX — year to 30 Jun | 31 Aug 2026 (passed) | Governance check on activist and engagement-led managers |
| 13F — Q3 2026 | 16 Nov 2026 (next) | Quarterly sweep |
| 13D / 13G | Event-driven | Standing alert on your coverage universe — this is the fresh channel, days rather than months |
| Letters | Rolling, ~2–6 weeks post-quarter | Re-run thesis extraction only when a new letter actually exists; check the archive, do not schedule blind |
Two mechanics worth holding in mind while reading the table. Form SHO, which would have given aggregate monthly short positions, has had its compliance date pushed to February 2028 — the short blind spot stays open. And the SEC has proposed cutting NPORT-P publication back from monthly to quarterly; that was a proposal, not a final rule, so confirm the current cadence before building a process around it.
TOOLING
Where this sits in your setup
The investing-ai plugin already covers one half of this cleanly. smart-money-thesis-scanner is prompt 05 productised — ticker-first, with the passive-sponsor exclusion, the 2% conviction floor, the tier system and the hard phase gates already baked in. For “who owns this name and why”, use the skill rather than the prompt.
The gap is the other direction. Nothing in the stack is manager-first: there is no equivalent of coverage-file-builder for a fund, no persistent manager file, and no quarterly diff engine. That is the piece worth building — a skill that takes a manager, resolves the entity, holds eight quarters of history, diffs on share count, and maintains managers/[name].md with an update log. Prompts 01–04 above are effectively its specification.
Two habits that compound: keep the manager files in one folder so cross-manager crowding is a grep rather than a research project, and log every run with its date, because next quarter’s value comes entirely from having last quarter’s file to diff against.
Filing mechanics current as of September 2026 and drawn from SEC materials and law-firm summaries of the 2023–24 beneficial ownership amendments, the Rule 13f-2 compliance extension, and the February 2026 N-PORT proposal. Deadlines shift for weekends and holidays — confirm on EDGAR. Nothing here is investment advice.
