Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Whale Panel

    1

    Buffett, Graham and Lynch each score a stock from its SEC filings under a deterministic rubric. Your agent narrates the numbers and never averages the three verdicts.

    Free

    Security scanned
    Whale Panel

    Whale Panel

    Example session with this skill installed

    Run the whale panel on Coca-Cola (KO). I want Buffett, Graham and Lynch each to weigh in from the latest filings, and tell me where they disagree.

    • Read your context and instructions
    • Compiled the whale panel

    panel narration · prompt v2 · panel_version 1

    Panel

    Buffett — bearish, 50/100 confidence, 18/27: "a fine business priced like a finer one"; intrinsic value sits well under the market cap.
    Graham — bearish, 71/100 confidence, 6/16: "margin of safety -71.2% <= -50% vs Graham Number forces bearish — price alone rules the name out."
    Lynch — bearish, 58/100 confidence, 6/15: "slow grower, PEG 3.34"; "the price is far ahead of the growth."
    Engine count: bullish 0, neutral 0, bearish 3, walked away 0
    Unanimity flag: the engine marks this agreement suspect because the snapshot carries data-quality warnings (ttm_stitched, fundamentals_stale_vs_price) — every rubric that scored read the same flagged input, and agreement on a flagged input is not independent confirmation; see the shared caveats.

    Per-whale gists

    Buffett. The business scores well: ROE 40.7%, operating margin 29.3%, debt/equity 0.01, full marks for management handing back $544M in net buybacks and $10.97B in dividends TTM. The valuation does not: owner earnings of $14.74B discount to an intrinsic value of $193.96B against a market cap of $353.88B, even with stage-1 growth set below the engine's own 5-year earnings CAGR. "Bearish here means 'don't buy at today's price,' not 'bad company.'"

    Graham. Earnings stability is a full 4 of 4, with EPS positive in all 10 periods and growing from 2.48 to 3.19, and dividends paid in 10 of 10. Financial strength loses points on a current ratio of 1.36 below the 1.5 floor. Valuation is 0 of 7: the Graham Number is $23.68 against a price of $82.28, NCAV is negative at -$38.09B, and the deep-overprice override settles the verdict. "A sound enterprise selling at a speculative price."

    Lynch. Coca-Cola comes out of the two-minute drill as a slow grower, with a 5-year EPS CAGR of 7.7% and revenue CAGR of 5.5%. Fundamentals are a perfect 4 of 4, but P/E 25.83 and PEG 3.34 both fail, and no GARP floor lifts the result to neutral. Lynch's open question: "does a dividend story justify this price? The rubric can't answer that."

    Synthesis

    Where the models agree: every scoring whale read the same strong operating record and the same thin liquidity. Buffett and Lynch both report debt/equity 0.01 and a 29.3% operating margin; Buffett and Graham both report a current ratio of 1.36 that fails their respective floors. All three read the same stale, stitched books against the same yfinance market cap. The DCF, the Graham Number, and the PEG all read the price as too high for what the numbers support.

    Where they diverge: three orthogonal models reach bearish by three different routes. The DCF gets there on cash flows, with Buffett noting the raw intrinsic value "is still under the market price" even before the engine's adjustment. The Graham Number gets there on earnings and book value, and by a margin wide enough to trip a forced override. The PEG gets there on growth rate, with Lynch classing the name a slow grower rather than a stalwart. Confidence is not uniform, and it attaches to different things: Graham is most confident at 71 because the override makes the verdict mechanical, while Buffett is least confident at 50 because the engine docked 15 points for fundamentals, management and valuation resting on flagged data. The most confident whale is also rendering the least surprising verdict. Lynch's 58 sits between, with the dividend case explicitly outside the rubric. The two data-quality warnings were treated the same way by all three: each narrated them as caveats affecting valuation, and none walked.

    The thread to pull: Buffett's leverage figure. He flags, as his own inference rather than an engine finding, that a debt/equity of 0.01 "is unusually low for a big branded consumer company" and that an owner "would want to check which debt tag resolved." Lynch scores the same 0.01 for 2 points without comment, so a tag error would move two whales' fundamentals scores at once.

    Shared caveats

    Snapshot snapshots/KO-2026-07-27.json, fetched 2026-07-27; fundamentals from SEC EDGAR via edgartools 5.43.0, market cap from yfinance:fast_info.market_cap.

    ttm_stitched: some trailing-twelve-month net income figures were stitched from year-to-date or annual filings by edgartools rather than reported directly as quarters, across periods ending 2023-12-31 through 2026-04-03. Every whale's valuation rests partly on calculated quarters.

    fundamentals_stale_vs_price: the latest books end 2026-04-03, 115 days before the snapshot date of 2026-07-27. The price is 115 days fresher than the books it is compared against.

    Snapshot age: this is a pinned snapshot dated 2026-07-27, used by request and not refreshed. Both price and books may have moved since.

    All whales: no filings sidecar, so no 10-K text behind any business claim; Form 4 insider data unavailable; no validation findings recorded.

    • Buffett-specific: the debt/equity 0.01 tag question above is his inference, not an engine finding.
    • Lynch-specific: the rubric does not score dividends, so a payout-based case for KO is outside what it measured.

    These are mechanical rubrics plus narration, not investment advice. Full single-whale diagnoses: /buffett KO, /graham KO, /lynch KO.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Ask "what would Buffett, Graham or Lynch say about this stock?" and get three independent, deterministic verdicts scored from SEC filings. The engine computes; your agent only narrates. Nothing is averaged into one score, because three philosophies with different models of value are not a vote.

    What it does

    1. Clones the open-source Whale Panel engine (MIT) and installs it with uv.
    2. Pulls the company's EDGAR filings and a delayed market quote, pins them to a dated snapshot.
    3. Scores that snapshot under a versioned rubric per investor: Buffett (moats, owner earnings, two-stage DCF), Graham (earnings stability, financial strength, net-net, Graham Number), Lynch (classify the company, gate on PEG). Greenblatt (return on tangible capital, earnings yield) is available as a single-whale diagnosis.
    4. Reports one line per whale, the engine's own count line verbatim, where the models diverge and why, and the data-quality caveats.

    Why it is different

    • Same file in, byte-identical JSON out. Every figure your agent prints comes from the engine. The agent is forbidden from computing, estimating or rounding beyond presentation.
    • Disagreement is the product. The DCF, the Graham Number and the PEG reach their verdicts by different routes. The report names the instruments, never "2 of 3".
    • Unanimity is flagged, not celebrated. When all rubrics read the same data-quality-flagged input, the engine marks the agreement suspect.
    • Missing data hard-fails. A whale "walks away" rather than scoring with defaults, and the panel says so.

    What you need

    git, uv (it fetches Python 3.11+ itself), and a name + email for the SEC's EDGAR identity requirement. The skill asks for it; it never invents one.

    Source, rubrics, tests and design records: https://github.com/philipbergman6-glitch/agents

    These are mechanical rubrics plus narration, not investment advice.

    Requires git and uv. US-listed companies with XBRL filings on SEC EDGAR only; foreign filers, banks and very young companies often lack the tags the rubrics need, and a whale then walks away rather than guessing. Greenblatt is single-whale only, not on the panel. Market cap comes from a delayed quote. Each run scores one pinned snapshot; re-fetch to refresh. Rubrics are mechanical; this is not investment advice.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 12 days ago

    • Free with an Agensi account

    Requires

    git
    uv (Python tool runner)
    Python 3.11+ (uv fetches it)

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Shell: git clone + uv sync + the whale CLI. Write: the clone lives in ~/whale-panel and snapshots are written under it. Network: SEC EDGAR filings and a delayed market quote (Cboe, yfinance fallback). Env: EDGAR_IDENTITY is passed inline per fetch and never written to disk.

    Listed12 days ago

    What's inside

    Frequently Asked Questions

    The agent never produces a number. A plain-Python engine fetches the company's SEC filings, pins them to a dated snapshot, and scores that snapshot under a versioned rubric per investor. Same file in, byte-identical JSON out. The agent only narrates that JSON and is forbidden from computing, estimating or averaging. Ask a chatbot the same question and the figures come from its memory.

    git and uv. uv fetches Python 3.11+ itself if the machine lacks it. The skill checks both, tells you what is missing, and installs nothing silently. The engine is cloned once to ~/whale-panel.

    Yes, to comply with SEC EDGAR access policies, the skill will prompt you to provide a name and email address which are used solely for the required identity headers when fetching official filings.

    The skill is free. It is the install-and-run procedure for the open-source Whale Panel engine (MIT): clone, uv sync, fetch filings, run the panel, and the report rules that keep the agent from averaging or inventing figures. The rubrics, personas and tests live in the public GitHub repo.

    The skill is designed to work with any AI agent capable of local code execution and shell access, provided the host machine meets the Python and 'uv' dependency requirements.

    Two things happen. Degenerate or missing inputs make a whale walk away: no score, a stated reason, and the panel proceeds over the whales that did score. Stale or stitched data still scores, but the engine attaches warning codes, docks confidence, and marks unanimous agreement as suspect because every rubric read the same flagged input. The report lists every warning.

    New in Trading & Investing