Equity Research Customised Process

Equity Research — Customised Process (Thesis-First)

Produce a defensible 10–18 page investment memo on a single stock by:

1. Understanding the business and industry deeply
2. Forming a direction (long / short / pass) early
3. Developing 2–4 thesis pillars
4. Steel-manning the counter-arguments as risks
5. Auditing pillars for falsifiability and materiality
6. Building one model with sensitivity tables
7. Iterating if model surprises arise
8. Writing the memo with a "What would change my mind" section

Output: a 10–18 page markdown memo (~5,000–8,000 words; cover + 2 appendices; the cover/Section 1 is itself a 1-page verbal-pitchable summary) + supporting Excel model + research archive directory the user can browse and revisit.

Core operating principles

Principle 1: Skill does data, user does judgment

Skill never picks a direction. Skill never grades pillar quality. Skill surfaces evidence and structure; user makes calls.

Principle 2: Q&A interludes — every phase boundary is a checkpoint

The skill must NEVER auto-advance between phases. After producing each phase output, the skill explicitly pauses with a message like:

"Phase X complete. Output: [name of file or summary]. Ask any question about the business / output / source data, request edits, or say 'continue' / 'next' to advance."

User questions during interlude are unrestricted in scope and number. User may:

• Ask for clarification on any concept or data point
• Request the skill explore a specific topic deeper
• Request alternative drafts
• Edit the output directly
• Drill into source documents (10-K excerpts, transcript pages, etc.)

Skill responds, then re-prompts: "Ready to continue, or more questions?"

User-driven cadence is essential. Some phases get heavy interrogation (Phases 2, 3, 7, 8, 9, 10, 11, 12); others get a quick "continue" (Phases 4, 5, 13). Skill must respect both. Phase 11 (model build) is a HEAVY phase, not a quick-continue one — it is where the thesis becomes auditable numbers, so it draws the most line-by-line interrogation ("show me before you edit", "where does the pillar feed in?", "why this growth rate?"). Budget for it: render the assumptions + pillar→driver map and pause for review before wiring scenarios (phase11 "Review gate before scenarios").

Principle 2b: Procedural questions are answered by the skill, not asked of the user

The workflow already prescribes sequencing — which phase follows which, what gets read when, when a manual fetch is needed. These are the skill's to answer, not the user's. Do NOT use AskUserQuestion (or an open free-text "what should I do next?") for a decision the workflow already owns. If the user asks a procedural question ("should we read the notes now or in Phase 5?", "is Phase 1 done?"), answer it from the workflow spec — don't bounce it back as a choice.

Reserve AskUserQuestion for genuine judgment the workflow deliberately leaves open: direction (long / short / pass), which pillar survives the audit, whether to accept or reject a materiality flag, which of several drafts to keep. When unsure whether a choice is procedural or judgment: if the answer is written somewhere in the skill, it's procedural — answer it; if it requires forming a view, it's judgment — the user decides. The sell-side done/skip/defer prompt is judgment (the user owns their data-gathering); "what phase is next" is procedural (the skill owns it).

Principle 3: Manual fetch when programmatic isn't possible

For paywalled / TOS-restricted sources (CapIQ, Bloomberg Terminal, FactSet research), skill prompts the user to fetch manually and drop files into the appropriate directory, then waits for "done." See references/phase1-context-load.md for the workflow.

Principle 4: Vocabulary discipline

• Direction = the rating (long / short / pass; BUY / HOLD / SELL)
• Pillar = one specific testable argument supporting the direction. Must have 5 elements: claim / driver / mechanism / magnitude / timeframe
• Killing condition = a pre-specified event/number that would prove a pillar wrong, written at thesis formation, used for live monitoring after publication

Never call a directional rating a "thesis." A thesis is the direction + the pillars supporting it. Pillars are the testable units.

Workflow at a glance

How to execute the workflow

When the skill triggers, the orchestrator does the following in order:

1. Confirm ticker and intent. Ask the user "Ticker?" if not provided. Confirm the goal is full thesis-first research (not a quick lookup).

2. Set up directory structure at ~/Claude Projects/Equity Research/[TICKER]/ with subdirectories:

   filings/         # 10-K / 10-Q / DEF 14A (US issuers) OR 20-F / 6-K (foreign private issuers). HTM as filed; PDF optional.
   transcripts/     # Earnings call transcripts (Motley Fool / Seeking Alpha / IR site)
   ir-materials/    # Investor day decks, latest earnings deck, 10-K/20-F language diff
   sell-side/       # USER POPULATES manually — CapIQ consensus DATA exports (both modes); note PDFs only if research_notes_available
   extractions/     # Structured primary-source extractions (20F_extraction.md, headline_anchors.md) — produced by Phase 2 Step 0
   working/         # Skill-synthesized files (incl. street_view.md when notes unavailable)
   deliverables/    # Final model + memo
   

   At Phase 1 the skill also sets research_notes_available (starting assumption false, confirmed with the user at the Phase 1 Step 8 gate) in working/context.md — false means CapIQ consensus data is available but written research-note PDFs are not, so the note-derived layer comes from the earnings-call analyst Q&A + free sources. Phases 5, 6, 10, and 11 read this flag downstream (the other phases are source-mode-independent).

3. Execute phases sequentially. For each phase:

   • Load the corresponding references/phaseN-*.md file
   • Follow its instructions
   • Produce the output file
   • Pause with the standard Q&A interlude prompt
   • Wait for user "continue" / "next" before advancing

4. Reference files (loaded on demand per phase):

   • Phase 1 → references/phase1-context-load.md — source fetching (EDGAR/transcripts/IR/market data), source hierarchy, two-mode CapIQ gate (Step 8), Street-view fallback (Step 9)
   • Phase 2 → references/phase2-company-brief.md — Step 0 structured extractions (hard gate), then the ~1,500-2,000w business synthesis
   • Phase 3 → references/phase3-industry-brief.md — market size/structure, competitive set, moats, tailwinds/headwinds
   • Phase 4 → references/phase4-driver-trees.md — revenue/margin decomposition; tree customization for hybrid businesses
   • Phase 5 → references/phase5-consensus-map.md — two-tier consensus (CapIQ headline + driver-level), dispersion flags, anchor existence matrix
   • Phase 6 → references/phase6-asymmetry-hunt.md — 4 asymmetry types, evidence standards, setup-bias + conviction-ceiling checks
   • Phase 7 → references/phase7-direction-brief.md — decision-ready synthesis brief; user commits long/short/pass
   • Phase 8 → references/phase8-pillar-template.md — 5-element pillar drafting, materiality flag, iterate-with-user loop
   • Phase 9 → references/phase9-steel-man.md — strongest counter-pillars + evidence-cited rejections → risk list
   • Phase 10 → references/phase10-pillar-audit.md — falsifiable/material/defensible gates + killing-condition drafting
   • Phase 11 → references/phase11-model-build.md — method gate, fresh derivation, model build + formatting, sensitivity, valuation_outputs.yaml
   • Phase 12 → references/phase12-iteration.md — pillar↔model reconciliation (always writes its file), 4 surprise modes
   • Phase 13 → references/phase13-pitch-template.md — memo structure/templates, discipline checklist, docx/pdf build

5. Reuse existing skills where appropriate. The default forward-multiple build in Phase 11 leverages financial-analysis:comps-analysis patterns; the DCF mechanics (when DCF is the selected method) leverage financial-analysis:dcf-model. Pillar tracking format in Phase 10 borrows from equity-research:thesis.

6. Always end the workflow with the pitch document and a recap message summarizing rating, target, pillars, and what to monitor.

Important behaviors

• NEVER fabricate figures. This is the highest-priority rule. Every quantitative claim in any phase output (Phases 1–10 and the memo) must be verified from a primary source on disk (filings, transcripts, shareholder letters, sell-side notes, market-data exports). If a number is not available from a source, write [not disclosed] or [needs sell-side pull] — never invent a plausible-sounding placeholder. The ONLY phase where unverified numbers are permitted is Phase 11 (Build model), where forecast inputs are assumptions by definition — and even then, every assumption must have a documented basis (Street consensus, mgmt guidance, computed from historical trend, or explicit pillar claim from Phase 8). Plausibility ≠ fact. If you catch yourself rounding "I think it's around X" into "X" — stop and verify or flag.

• No fake precision. Never attach specific page numbers, take rates, percentages, or counts to a claim unless that specific number has been verified in the source during this session. The failure mode is plausible precision — e.g., "10-20% transaction fee" or "p.7 of the JPM note" — both can be made up while feeling researched, because the substance you're attaching them to may be defensible. If you are confident about the substance but cannot verify a specific number / page, write the substance in generic form ("per Wolfe initiation", "industry standard for DSPs", "sell-side estimates", "trade press reporting") — never invent precision. The test before citing a page number: can you point to the line on the page right now? If not, generalize the citation. This rule is a tighter sibling of "NEVER fabricate" — the former covers literal made-up numbers; this one covers fake precision wrapped around real substance.

• Pre-write claim test for every number. Before writing any specific number (percentage, $/€ amount, count, page reference) in any phase output, run the claim test:

  • Is it disclosed in a primary source I read THIS SESSION? → cite with source + page
  • Is it computed from primary disclosures? → cite as [computed from X, Y]
  • Is it from a sell-side note? → OPEN AND VERIFY the page before writing the citation
  • None of the above? → omit the number. Write [not disclosed] or describe qualitatively. Never write "approximately X" as a workaround for "I don't know X."

• No orphan numbers in outputs. Every numeric claim must have an inline citation [source], [computed], or [not disclosed]. Machine-assisted: run scripts/validate_citations.py <file> on every phase output before saving — it produces a review list of uncited numeric prose lines; each flagged line must be either cited or confirmed as the analyst's own model output / forward threshold (which is its own source). Zero unexplained flags before saving.

• My modeling vs. sourced fact must be visually distinct. When a number is my own modeling extrapolation (not from any source), state explicitly: "my modeling: ~X% based on [methodology]" — never embed it as if it were disclosed or sell-side. The reader needs to know which side of the line each number sits on.

• Phase 11 fresh derivation — back-of-envelope numbers from earlier phases never become model inputs without independent primary-source derivation. Pillar magnitudes asserted in Phases 8/10 (e.g., "+50-150bps GM vs Street," "+€110-330M FY28 GP delta," "approximately 2-7% PT impact") are qualitative structuring tools. The model in Phase 11 re-derives every input from primary sources (20-F line items, shareholder-letter actuals, sell-side consensus xlsx) — shares, tax rate, capex, WC, the applied multiple and its peer/own-history/sell-side-implied anchors (and WACC components only if DCF is the selected method). If the rebuilt model disagrees with a pillar magnitude, that's a Phase 12 surprise trigger, not an error to silently align away. Track record across runs: every back-of-envelope chain caught at least one error when surfaced for review. See references/phase11-model-build.md Step 0 for the full rule.

• Every numeric claim has an inline citation. Three accepted forms:

  • [source, p.N] — verified, number and page checked: e.g., €17,186M [Q4 25 letter, p.3]
  • [source] — substance verified, exact page not opened: e.g., ~70% royalty share [Wolfe init]
  • [est, not disclosed] — acknowledged as not a fact: e.g., Family tier ~25% of subs [est, not disclosed]

  No bare numbers in any phase output. A claim like Family ~25% must be either cited or flagged as an estimate. Enforceable at proofread time via scripts/validate_citations.py.

• Source-fallback rule — never skip a needed data point. When the first-choice source for a data type doesn't contain a needed number, fall back to the next layer rather than omitting the data point. Specifically:

  • Headline number missing from shareholder letter → check the 10-Q (quarterly) or 10-K / 20-F (annual)
  • Detailed cost line missing from MD&A → check the audited P&L line items in the F-pages / Item 8
  • Segment metric missing from segment note → check the operational review in MD&A
  • KPI missing from press release → check the investor day deck or earnings call transcript

  If after the fallback ladder the data point is genuinely not disclosed anywhere, mark inline as [not disclosed] — never just omit. Omission silently corrupts comparisons (you can't tell whether the value was zero, unknown, or deliberately skipped).

• Data source hierarchy — by purpose, not by file. Each data type has a designated source. Pulling the same number from the wrong source corrupts comparisons.

  • Company-reported actuals, headline numbers & KPIs (revenue, OI, NI, GM, EPS, subs, MAU, ARPU) → Shareholder letter / earnings press release (clean multi-year tables; identical numbers to the audited filing, faster to find).
  • Detailed audited line-items (R&D / S&M / G&A separately, segment P&L, share count detail, tax reconciliation, lease/debt schedules, risk factors, governance, exec comp) → 10-K / 20-F / proxy. The shareholder letter doesn't split these.
  • Past-quarter actuals vs. pre-print consensus (beat/miss table) → CapIQ Consensus xlsx (the xlsx has both the Actual and the Median pre-print consensus in one file — avoids cross-source mismatch). Do NOT pull the actual from the shareholder letter when comparing to CapIQ consensus — use CapIQ's own actuals row for internal consistency.
  • Next-quarter guide vs. Street consensus → Web search (Reuters/FT/Bloomberg/Investing.com). Press captures the Mean used in headlines; CapIQ Median can differ from the press consensus.
  • Forward FY+1/2/3 consensus → CapIQ xlsx Median row.
  • Qualitative strategic context + Street framing (forward direction, mgmt tone, market reaction, what analysts are worried about) → Earnings call transcript (incl. analyst Q&A) + financial press; sell-side analyst notes too if research_notes_available. Not for numerical claims.

  Two source modes (research_notes_available, set in Phase 1): CapIQ consensus data (estimates, comps multiples, own-history, aggregate PT) is available in both modes and stays the primary for everything numeric. Only the written research-note PDFs are conditional — when unavailable (the default), their qualitative layer (driver-level estimates, Street framing, per-bank PT bridges, industry primers) is supplied by the earnings-call analyst Q&A + mgmt guidance + free public sources (working/street_view.md). See references/phase1-context-load.md Step 0 / Step 9.

  Always re-pull the headline anchor numbers from the most recent shareholder letter at the start of any phase that uses them. A year-stale anchor (e.g., FY24 actuals used after FY25 results are out) will systematically corrupt every downstream figure. See references/phase1-context-load.md for the full source-hierarchy table.

• Always compare the latest quarterly results to consensus expectations. Whenever a phase output references the most recent quarter, include the pre-print consensus alongside the actual figure for revenue, EPS, MAU/subs (or other KPIs central to the business), and forward guidance. Source the consensus from the CapIQ Consensus xlsx (Refinitiv I/B/E/S aggregation — available in both modes), company-provided reconciliations, or sell-side notes if research_notes_available. If consensus isn't available, say "consensus not retrieved" — never guess at the magnitude of beat/miss.

• Always identify and research the material live events/overhangs driving the stock. The filing's risk factors are generic and stale — they won't surface an active investigation, a fresh lawsuit, a short report, a pending deal, or a guidance cut. Run the Phase 1 Step 5b "Material events & overhangs scan" (web research + recent 8-K/6-K, against the explicit trigger list) and research each material item to depth in Phase 2 §4 (status: resolved/ongoing/pending; outcome range; thesis read) — writing it up as a dedicated section of the company brief (Phase 2 is a single file — do not scatter into separate working/<topic>.md notes). Never rely on the filing alone for what the market is actually pricing. TCOM/working/company_brief.md §5 (the SAMR regulatory deep-dive) is a worked template.

• Find the crux mechanically, then go deep on it. The most important question on a name is usually the one the company does NOT cleanly disclose (Temu inside PDD's "transaction services," AWS once buried in Amazon "Other," international-margin inside TCOM's single segment) — so never let the reported segment structure define what matters. Phase 1 Step 5c "Key debate / crux scan" derives it from market signals you tabulate: the analyst-question tally (the most-asked topic on the call IS the crux), the "what's not broken out?" check, the price/multiple, and the bull/bear one-liner. Then carry the named crux: Phase 2 surfaces & describes it (breadth) → Phase 4 quantifies its economics (reconstruct undisclosed segment margin / take-rate / sub-business P&L via triangulation + peers + call Q&A, flagged [est, derived]) → Phase 6 adjudicates. "Not disclosed" is the start of the analysis, not the end; leaving a load-bearing number unreconstructed is itself a failure. The test: have I answered the question a sharp PM asks first about this name?

• Length is never a cap on working documents. Word/page counts in any phase spec are indicative only. Completeness comes first — include everything material; never drop a required element to hit a length. (The final pitch memo's page range is guidance, not a hard cap.)

• Never skip the Q&A interlude. Even if the user has been silent, prompt them at every phase boundary.

• Never write a thesis pillar without all 5 elements. Drafts that lack any of (claim / driver / mechanism / magnitude / timeframe) must be flagged and rewritten before being shown to user.

• Never auto-pick a direction. Phase 7 produces a tentative read with conviction level, but the user commits.

• Save killing conditions verbatim. What's drafted in Phase 10 is what appears in the Phase 13 pitch under "What would change my mind." Don't paraphrase between phases.

• Pass is a valid output. If Phase 12 iteration doesn't converge after 1–2 rounds, recommend "pass" and document why. Better than a forced thesis.

• Two source modes — set research_notes_available at Phase 1, by asking the user, not by silent default. The user has CapIQ consensus data in both modes but the written research-note PDFs only when research_notes_available: true (starting assumption false). Note access varies run-to-run, so the skill must surface the mode choice at the Step 8 gate — request the CapIQ data exports (required in both modes) and, in the same message, offer the user the option to also supply written notes (Mode B) with a short list of what to grab. The user's reply sets the flag; only fall through to false without asking if the user already declined notes this session. Under false, the note-derived layer is replaced by the earnings-call analyst Q&A + mgmt guidance + free sources; do NOT block waiting for notes that won't come. (See references/phase1-context-load.md Step 0 / Step 8.)

• Manual fetch waits forever. When prompting the user to drop the CapIQ data exports into sell-side/ (Step 8), wait indefinitely. Don't time out and proceed without them unless the user says "skip." (Under research_notes_available: false you wait for the data exports, not for written notes.)

• Never declare Phase 1 complete before the Step 8 fetch decision. The CapIQ-data pull is a gating step of Phase 1, not an optional add-on. Issue the Step 8 request as its own message and wait for an explicit done / skip / defer. Do NOT bundle it underneath a "Phase 1 complete — ready for Phase 2?" summary, and do NOT offer to advance to Phase 2 until the user has answered. The default assumption is "the user is gathering the exports" — silence means keep waiting, not advance. (When research_notes_available: true, the same gate also covers the note PDFs.) (See references/phase1-context-load.md gating rule.)

• "Phase 1 complete" requires a complete archive AND an honest audit trail. Step 8b is a two-part gate. (A) Completeness: every required artifact — ≥2yr filings, last 4 earnings-call transcripts, last 4–5 quarterly 6-Ks, IR decks, market data, material-events scan, CapIQ data — must be PRESENT or JUSTIFIED-ABSENT (attempted via the step's full source ladder + confirmed genuinely unavailable, sources named). "PENDING / fetch-later / didn't-try" is not a passing state, and under Mode A (research_notes_available: false) the call transcripts are mandatory and non-deferrable — they substitute for the missing written notes. (B) Consistency: source_index.md + context.md must match the actual folder (every export/note logged, research_notes_available recorded, no stale "PENDING"). A folder missing an unjustified artifact — or an index that disagrees with the folder — is not complete. (See references/phase1-context-load.md Step 8b.)

Pre-Phase-13 audit checklist (mandatory before drafting memo)

Before opening Phase 13 to draft the memo, the skill runs a four-check internal-consistency audit across the working/ files. Do NOT advance to Phase 13 if any check fails — fix the working files first. The memo is a synthesis layer; if the inputs disagree, no amount of memo iteration will fix it.

Run the audit; report failures explicitly to the user; pause for fixes before Phase 13. If all four pass, advance.

User preferences layer

Across multiple memos a given user develops settled stylistic preferences that, captured once, should drive the first draft of every future memo rather than re-iterating from defaults each time. Examples of preferences worth capturing: bullets-vs-prose balance, "Thesis" vs "Pillar" terminology, methodology placement (body vs appendix), inclusion or omission of optional sections (comps cross-check, sell-side reconciliation, Management subsection, named long-duration-competitive-threat sub-section), preferred header conventions, page-count target.

Capture preferences in either:

• The user's auto-memory: ~/.claude/projects/-Users-[username]/memory/ (persistent across all conversations)
• A skill-local preferences.md if user wants the preferences scoped just to equity research

Reference these at Phase 13 start so the first draft of every future memo matches established conventions. Settled preferences that conflict with the template default should follow the user's preference, not the template.

Style notes for outputs

• Plain English over jargon. The user is a student building intuition, not an MD pretending sophistication.
• Quantify whenever possible. Pillars without numbers are vibes.
• Cite sources with file path and page/section. E.g., "FY24 10-K p. 47" or "Q3 FY25 transcript, CFO opening remarks."
• Prefer markdown tables over prose for comparative data (driver trees, consensus maps, sensitivity outputs).
• The memo is 10-18 pages (~5,000-8,000 words) with cover + 2 appendices. Cover is the verbal-pitchable summary; body sections carry the depth; Appendix A has the valuation mechanics (forward-multiple triangulation + equity bridge by default, or DCF + WACC build if DCF is the selected method), Appendix B has key model assumptions. If the overall memo spills past 18 pages, tighten — but the previous 5-8 page target was too tight for an interview-grade memo with a valuation appendix.