# Business Decoder > Build a first-principles, plain-English understanding of any publicly listed company — the front-of-funnel lens applied BEFORE any valuation or thesis work. Covers what they do, how they make money, the need they serve, their customers, value-chain seat, ownership and control, the tailwind or headwind they ride, and whether it is a good, mediocre, or fragile business — translating ALL domain jarg… - **Type:** Skill - **Install:** `agentstack add skill-viktor-milev-claude-skills-business-decoder` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [viktor-milev](https://agentstack.voostack.com/s/viktor-milev) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [viktor-milev](https://github.com/viktor-milev) - **Source:** https://github.com/viktor-milev/claude-skills/tree/main/business-decoder ## Install ```sh agentstack add skill-viktor-milev-claude-skills-business-decoder ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Business Decoder The first lens, applied before any numbers. Given any listed company — one you may be hearing about for the first time — this skill produces a plain-English memo that lets a finance-fluent generalist (think a CIO who is not a biologist, physicist, or engineer) *fundamentally understand the business from first principles*: what it does, how it makes money, the real need it serves, why it is or isn't a good business, and the structural force it rides. It sits at the front of the research funnel — **upstream of `investment-council`** (which pressure-tests a thesis or trade) and **upstream of all valuation work** (a dedicated valuation skill). Its job is comprehension and a *worth-deeper-work?* steer, never a buy/sell call and never a valuation. The defining quality bar is **de-jargoning**: every term from the underlying science or engineering is translated into a faithful plain-English analogy, with the precise term kept in a parenthetical so the reader can verify. Finance and strategy vocabulary is fair game; domain-technical vocabulary is not. This bar is hardest — and matters most — in deep-tech and biotech, so it is treated as a cross-cutting discipline, not a section. See `references/de-jargoning.md`. --- ## CHANGELOG **v1.3** — Surgical patch from an independent adversarial QC audit (no spine, template, or Numbers-Line-core changes; description still frozen sub-1,024). Tightened the **mode/framing interaction**: the mode recommendation is a non-blocking statement the user may override, and the one permitted framing question is *identity only*. Fixed the **"dated-source flags" dangling reference** (dated per-fact sourcing lives in the dossier/working notes; the memo summarizes provenance in section 10). Added a **banner-from-section-10-first** ordering note. Added a **non-listed-entity branch** to Step 0. Disambiguated the **Phase-2 ingest phrase** ("ingest this dossier" / "here's the research", retiring the overloaded "go deep with this"). Extended the share rule with a **market-indeterminate carve-out** (revenue-bearing but undefined denominator → name the indeterminacy, don't manufacture a number). Clarified the honest-bear cap as **one *tight* paragraph (density may scale, length stays capped)** — a deliberate scope boundary, not expanded. Deferred by deliberate call: Standard-vs-Deep rationale wording (low-confidence, weak counter-evidence), §9-verdict-as-reasoning-point (verdict is downstream of the §8 bear, already a scoped point), and the "section 0" label (cosmetic). **v1.2** — Maintenance review (single-run architecture reaffirmed; no redesign). Added a **silent terminal de-jargon verification pass** at the end of Step 3 (run the de-jargon test over every translated term before finalizing). Replaced the soft "reasoning effort" aside with **scoped extended-reasoning** bound to the two judgment points where value concentrates (faithful translation under uncertainty; the honest-bear adjudication). **Consolidated** the "character-not-magnitude" and "market-structure / estimated-share" rules to single canonical statements (in The Numbers Line) with cross-references — requirements unchanged; the v1.1 pre-revenue carve-out and the paste-ready Phase-1 brief left intact. **Reframed length** as complexity-scaled guidance (tight for a well-covered large-cap; longer expected for contested/obscure deep-tech) rather than a fixed page cap; "30-Second Version must stand alone" retained. **Trimmed** the in-body Industry-Lenses summary to a pointer + lens list (full detail stays in the reference file). Added **output-hygiene rules**: strip the HTML skeleton's authoring comment on inject (memo body appears once); no raw source labels in memo prose. Added an explicit **WARNING-banner threshold** (raise it whenever section 10 will flag any *load-bearing* fact as unverified). Frontmatter `description` unchanged (sub-1,024-char invariant preserved). **v1.1** — Added competitive **market structure** (monopoly / duopoly / oligopoly / fragmented) and the company's **estimated market share** (rounded to ~10% increments) as an explicit, stressed element of section 6 and the Phase-1 brief. Classified both as allowed structural facts under the numbers line, with a pre-revenue carve-out (characterize the competitive *field* when no end-market exists yet). Frontmatter `description` unchanged (sub-1,024-char invariant preserved). **v1.0** — Initial release. Single skill, two internal modes (Deep / Standard). Phase-1 Research Brief lives inside the skill. Twelve-section memo spine, "character not magnitude" numbers line, "Field Notes" light HTML theme on demand. --- ## Trigger Conditions **Mandatory triggers:** `decode this business`, `decode this company`, `business decoder`, `what does X actually do`, `how does X make money`, `understand this business`, `explain this company in plain English`, `first-principles read on X`, `break down what X does`. **Strong triggers** (when the intent is comprehension of a *named listed company*, with no thesis or valuation attached): `is X a good business`, `I just heard about X — what are they`, `is this a real business`, `what problem does X solve`, `who are X's customers`, `where does X sit in the value chain`, `what is X, in plain terms`. **Do NOT trigger** — route elsewhere: - A buy/sell/short call, a mispricing claim, position sizing, "is X a good *trade*," "should I buy X," "what's wrong with this thesis" → `investment-council`. - Valuation, multiples, margins, growth rates, returns on capital, price targets, comps, discounted-cash-flow, tearsheets, consensus estimates, earnings models → a dedicated valuation skill. - Fund or asset-manager due diligence, performance/risk metrics → a fund-evaluation skill. - A Bloomberg beta/regression screenshot → a dedicated beta/regression reader. The clean line: this skill answers **"is this a real and good *business*, and what is it?"** It hands off the moment the question becomes **"is this a good *trade*, and what's it worth?"** --- ## The Numbers Line (read before writing) This is comprehension, not valuation — but "is it a good business?" is inseparable from the *character* of its economics. So the line is **character, not magnitude**. **This section is the canonical definition of what is allowed and what is banned; the Hard Rules and memo template reference it rather than restating it.** - **Allowed** — qualitative economic character: recurring vs one-off vs project-based vs binary-milestone revenue; price-maker vs price-taker; mission-critical vs discretionary spend; razor/blade; toll-booth vs commodity; and **competitive market structure** (monopoly / duopoly / oligopoly / fragmented many-player). And **structural facts**: revenue mix %, customer/supplier/product/geographic concentration %, segment split, rough scale (e.g. "~$2bn revenue, ~70% from one product line"), and the company's **estimated market share of its relevant market**, rounded to ~10% increments (an estimate is fine — "roughly a third of the market," "~40% share" — precision is not the point). - **Pre-revenue carve-out:** if the company has no end-market yet (e.g. a clinical-stage biotech with no product), end-market share is *not yet applicable* — say so, and instead characterize the competitive **field** of the technology/platform (how many serious players chase the same approach, and whether this name is a front-runner or one of a crowd). - **Market-indeterminate carve-out:** if the company *has* revenue but its relevant market is itself undefined or contested — the denominator can't be cleanly drawn, common on exactly the obscure deep-tech names Deep mode exists for — do not manufacture a precise share. Give the most defensible rough bucket only if one genuinely exists, name *which* market you are measuring, and flag the indeterminacy as an open question in section 10 (and the banner). The mandate is to *address* the market, never to invent a number to satisfy it. - **Banned** — any number that feeds a valuation: margins, multiples, growth rates, returns on capital, free-cash-flow figures, price targets, "cheap/expensive." If you catch yourself reaching for one, stop — that belongs to a downstream skill. If the analysis genuinely cannot proceed without a magnitude, name the gap in the memo's open-questions section rather than supplying the number. --- ## Mode Selection One skill, two modes. Decide at framing; state which mode you're in. - **Deep mode** — the default for a *fresh, obscure, or technically hard* name (any deep-tech, biotech, photonics, or specialist industrial; any company you'd otherwise only be able to describe from its own deck). The skill emits a **Phase-1 Research Brief** — a paste-ready directive — then stops. The user runs it with the **Research** feature on (which a skill cannot toggle itself — say so plainly), and re-invokes business-decoder with the returned dossier for Phase 2. - **Standard mode** — the fast path for a *well-known, well-covered* name, or whenever the user says "quick," "fast," or "just do a pass." One invocation: do intensive inline retrieval (searches scaled to complexity) and write the memo directly. - **Phase-2 ingest** — if the user has *already* supplied a research dossier (or pastes Research output, or says "ingest this dossier" / "here's the research"), skip straight to writing the memo from it. Default lean: for non-trivial or hard names, recommend Deep mode (emit the Phase-1 brief) and name Standard as the quick alternative the user can choose instead — offered as a statement, not a question that blocks. Never force two turns on an easy large-cap. --- ## Session Protocol ### Step 0 — Frame Identify the company and **resolve its identity**: legal name, ticker(s), the specific listing (an American Depositary Receipt versus the local line, the parent versus a listed subsidiary). If identity is genuinely ambiguous — dual listings, a name collision, a very recent merger/spin, or the user gave only a vague descriptor — ask **one** clarifying question, then stop. Otherwise proceed. If X turns out **not to be a publicly listed entity** (a private company, a division, a product, or a protocol — phrasings like "what does X do" and "I just heard about X" will sometimes land on these), say so plainly and either proceed on the available non-filing sources while flagging the absent annual-filing tier in the WARNING banner, or, if it falls outside the skill's listed-company scope, recommend the more appropriate route. Never invent a listing. Classify the **industry lens** and read the matching section of `references/industry-lenses.md` for what to emphasize and what jargon will need translating. If no lens fits cleanly, apply the universal spine with first-principles judgment — the spine always holds; the lens only shifts emphasis. Select the **mode** per the Mode Selection logic above. State the chosen mode as a recommendation, not a blocking question — on a hard name that means emitting the Phase-1 brief and stopping; the user overrides simply by asking for the other mode. The only question that may halt framing is the identity question above. ### Step 1 — Deep mode only: emit the Phase-1 Research Brief Produce the brief below, paste-ready, then stop and hand control back. Do not start analyzing. ``` PHASE-1 RESEARCH BRIEF — business-decoder Company: [name / ticker / listing] Run this with the Research feature ON, then paste the dossier back into a new business-decoder turn (say "ingest this dossier"). Gather, with a dated source for each fact, enough to answer: - In one plain sentence a non-expert grasps: what does this company sell, and to whom? - How does revenue actually arrive (subscription / one-off / project / aftermarket / regulatory milestone)? Who writes the cheque? - What need or problem does the customer hire them for? Is it mission-critical or discretionary? What would the customer do if this company vanished? - Where do they sit in the value chain, and who holds the power in that chain? - What is the competitive market structure (monopoly / duopoly / oligopoly / fragmented), who are the main rivals, and roughly what share of its relevant market does the company hold (an estimate rounded to ~10% is fine)? If pre-revenue, characterize the competitive field of the technology instead. - Who owns and controls the company (founder/key person, family, state/SWF, PE, diffuse ETF/index)? Control mechanisms (dual-class, golden share)? - What structural tailwind or headwind are they tied to — and does it actually flow through to THEIR revenue, or are they merely adjacent to the story? - Where is the business concentrated or fragile (one customer, one product, one geography, one regulator, key-person dependence)? Source priority (capture provenance for each): 1. Latest annual filing — Business section (e.g. Form 10-K Item 1) AND Risk Factors (Item 1A), or the local-regulator equivalent. Risk Factors especially: the one place a company is legally forced to say what could go wrong. 2. Latest investor presentation; homepage; product and pricing pages. 3. Latest earnings or investor-day TRANSCRIPT (management under analyst questioning is often clearer than the deck). 4. At least one INDEPENDENT / ADVERSARIAL read — competitor framing, sell-side initiation, trade press, a short report if one exists. Flag it as such. 5. For deep-tech/biotech: a layperson explainer of the underlying science, used to UNDERSTAND only — never reproduced. 6. Catch-all, situation-specific: clinical-trial registries, regulator actions, patents, ownership/13F holdings, litigation, named suppliers/customers, job postings (an underrated tell for what's actually being built and sold). Return a sourced dossier. Note anything you could NOT find. ``` ### Step 2 — Ingest sources and live-verify Whether from a Phase-1 dossier (Deep) or inline retrieval (Standard): assemble the sources in priority order, and **live-verify the load-bearing facts** — what they sell, how revenue arrives, ownership/control, the concentration figures, the tailwind linkage. Treat company sources as the company's *claims*, not facts; weight the Risk Factors section and any independent source against the self-narrative. This skill is **retrieval-dependent**. If retrieval is unavailable or sources can't be obtained in full, the run is *degraded, not blocked*: proceed, but flag every unverified load-bearing fact as memory-not-verified and raise the WARNING banner (Step 4). The adversarial read is **best-effort** — if none could be found, say so in the banner rather than inventing one. ### Step 3 — Write the memo Use the exact template in **The Memo** below. Apply the de-jargoning discipline throughout (`references/de-jargoning.md`). **Length scales with the name.** A well-covered large-cap should land tight (≈1.5 pages); a contested, obscure, or technically hard deep-tech name will — and should — run longer (2–3+ pages). Write to the complexity of the company, not to a fixed page count: never pad, never truncate. The 30-Second Version is mandatory and must be genuinely sufficient on its own. **Spend reasoning where the value concentrates.** Engage real reasoning effort narrowly, at exactly two points: (a) the faithful translation of any term whose plain-English analogy involves genuine uncertainty, and (b) the honest-bear adjudication in section 8. Do not spread effort evenly across the memo — these two are where the craft and the judgment live. **Before finalizing, run a silent verification pass over every translated term.** Apply the de-jargon test from `references/de-jargoning.md`: could a smart generalist who read *only* that sentence (a) state what the company sells and (b) not be embarrassed saying it in front of a domain expert? Rewrite any term that fails. This is a check you perform internally, not visible chain-of-thought in the memo — and it must preserve "faithful first, … ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [viktor-milev](https://github.com/viktor-milev) - **Source:** [viktor-milev/claude-skills](https://github.com/viktor-milev/claude-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-viktor-milev-claude-skills-business-decoder - Seller: https://agentstack.voostack.com/s/viktor-milev - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.