Kanchi Dividend Sop Skill for Claude

Kanchi Dividend Sop

Overview

Implement Kanchi's 5-step method as a deterministic workflow for US dividend investing.

Prioritize safety and repeatability over aggressive yield chasing.

When to Use

Use this skill when the user needs:

• Kanchi-style dividend stock selection adapted for US equities.
• A repeatable screening and pullback-entry process instead of ad-hoc picks.
• One-page underwriting memos with explicit invalidation conditions.
• A handoff package for monitoring and tax/account-location workflows.

Prerequisites

API Key Setup

The entry signal script requires FMP API access:

export FMP_API_KEY=your_api_key_here

Input Sources

Prepare one of the following inputs before running the workflow:

• Output from skills/value-dividend-screener/scripts/screen_dividend_stocks.py.
• Output from skills/dividend-growth-pullback-screener/scripts/screen_dividend_growth_rsi.py.
• User-provided ticker list (broker export or manual list).

Expected JSON Input Format

When using --input, provide JSON in one of these formats:

{
  "profile": "balanced",
  "candidates": [
    {"ticker": "JNJ", "bucket": "core"},
    {"ticker": "O", "bucket": "satellite"}
  ]
}

Or simplified:

{
  "tickers": ["JNJ", "PG", "KO"]
}

The optional value-dividend-screener and

dividend-growth-pullback-screener handoffs use stocks[].symbol.

Both build_sop_plan.py --input and build_entry_signals.py --input

accept that shape directly, as well as the native candidates[].ticker and

tickers[] shapes above.

For deterministic artifact generation, provide tickers to:

python3 skills/kanchi-dividend-sop/scripts/build_sop_plan.py \
  --tickers "JNJ,PG,KO" \
  --output-dir reports/

For Step 5 entry timing artifacts. --yield-floor is mandatory — it is

the Step-1 yield gate; without it every row fail-safes to STEP1-RECHECK

(a row can never reach a PASS tier without Step 1). Pass --profile /

--safety-bias for run_context, and --events-json for the Step 4b scan

(absent ⇒ every row is treated as SKIPPED and a TRIGGERED name is capped

to HOLD-REVIEW — never silently clean):

python3 skills/kanchi-dividend-sop/scripts/build_entry_signals.py \
  --tickers "JNJ,PG,KO" \
  --alpha-pp 0.5 \
  --yield-floor 3.0 \
  --profile balanced --safety-bias medium \
  --events-json reports/kanchi_events_2026-05-17.json \
  --output-dir reports/

Workflow

1) Define mandate before screening

Collect and lock the parameters first:

• Objective: current cash income vs dividend growth.
• Max positions and position-size cap.
• Allowed instruments: stock only, or include REIT/BDC/ETF.
• Preferred account type context: taxable vs IRA-like accounts.

Load references/default-thresholds.md and apply baseline

settings unless the user overrides.

2) Build the investable universe

Start with a quality-biased universe:

• Core bucket: long dividend growth names (for example, Dividend Aristocrats style quality set).
• Satellite bucket: higher-yield sectors (utilities, telecom, REITs) in a separate risk bucket.

Use explicit source priority for ticker collection:

• skills/value-dividend-screener/scripts/screen_dividend_stocks.py output (FMP/FINVIZ).
• skills/dividend-growth-pullback-screener/scripts/screen_dividend_growth_rsi.py output.
• User-provided broker export or manual ticker list when APIs are unavailable.

Return a ticker list grouped by bucket before moving forward.

3) Apply Kanchi Step 1 (yield filter with trap flag)

Primary rule:

• Step-1 yield = the regular forward yield = `latest_declared_regular

dividend × cadence-implied frequency / price (WS-1 dividend_basis.py`).

Never use profile.lastDividend / TTM — it lags the latest declared raise

(defect D5) and silently bundles specials (D4).

• Apply the profile floor (income-now 4.0% / balanced 3.0% / growth-first

1.5%) to the regular yield only.

Trap & freshness controls (machine-emitted by dividend_basis.py):

• special_dividend_flag → exclude specials; report regular vs ttm yield.
• variable_policy_flag → FAIL (CALM-style; not an income base).
• cut_flag → FAIL; suspension_flag → FAIL.
• freeze_flag → HOLD-REVIEW (income cash-cow exception decided in

Step 8 synthesis only if safety is clean & unblocked).

• Data Freshness Gate: if the regular yield is within ±0.20pp of the

floor (floor_borderline) and the latest declared dividend is not

confirmed from an authoritative source, emit STEP1-RECHECK — **never a

hard FAIL** (this is the CFR D5 fix).

4) Apply Kanchi Step 2 (growth and safety) — sector-dispatched

Safety is sector-specific — a uniform GAAP/FCF triad mis-judges banks

(FCF meaningless) and regulated utilities (FCF structurally negative).

Use references/sector-step2-modules.md; the deterministic dispatch is

scripts/payout_safety.py.

• Always compute the payout triad: GAAP-EPS payout, Adjusted-EPS

payout, FCF payout. The safety verdict uses Adjusted-EPS + FCF

(consumer), or the sector module (bank / utility / insurer).

• adjusted_eps_source = UNAVAILABLE ⇒ cap HOLD-REVIEW (fail-safe;

never a silent PASS).

• GAAP↔Adjusted EPS divergence > 25% ⇒ Step-4 one-off flag.
• A merger completed within 4 quarters presumes GAAP EPS is distorted

⇒ force the adjusted path or HOLD-REVIEW (FITB/Comerica golden case).

• Regulated utilities: negative FCF is not an auto-FAIL — judge on

FFO/debt + allowed ROE + rate-case + equity-issuance risk.

When trend is mixed but not broken, classify as HOLD-REVIEW instead of

hard reject.

5) Apply Kanchi Step 3 (valuation) with US sector mapping

Use references/valuation-and-one-off-checks.md and apply

sector-specific valuation logic:

• Financials: PER x PBR can remain primary.
• REITs: use P/FFO or P/AFFO instead of plain P/E.
• Asset-light sectors: combine forward P/E, P/FCF, and historical range.

Always report which valuation method was used for each ticker.

6) Apply Kanchi Step 4 (one-off event filter)

Reject or downgrade names where recent profits rely on one-time effects:

• Asset sale gains, litigation settlement, tax effect spikes.
• Margin spike unsupported by sales trend.
• Repeated "one-time/non-recurring" adjustments.

Record one-line evidence for each FAIL to keep auditability.

6b) Apply Kanchi Step 4b (forward structural-event scan)

Step 4 is backward-looking; Step 4b catches *pending/recent* structural

events (the MKC-Unilever miss, D3). For each surviving candidate, run a

WebSearch + issuer-IR/SEC check using the source hierarchy: issuer IR

→ SEC filing (8-K/10-Q/10-K/proxy/S-4) → exchange/company deck →

reputable wire → finance portals (secondary only). Record findings into a

curated events JSON and pass it via build_entry_signals.py --events-json.

• Only a major structural event caps the verdict to HOLD-REVIEW

(tx > 10% mcap, share issuance > 10–20%, leverage +0.5x EBITDA,

control/listing/HQ change, merger-of-equals / RMT / spin-off / large

asset sale, dividend/rating/leverage-policy change, sector-specific

materiality, or rolling-24m cumulative M&A > 15% mcap). Minor bolt-ons

are a CAUTION note only.

• Pessimistic cap: FAILED-DEGRADED / SKIPPED / NO_EVENT_FOUND

on a Step-5 TRIGGERED name ⇒ HOLD-REVIEW + T1 BLOCKED. WebSearch

unavailable (web app / offline) is treated the same — never a silent

skip. CLEAN_CONFIRMED (primary source checked) is stronger than

NO_EVENT_FOUND (search only).

7) Apply Kanchi Step 5 (buy on weakness with rules)

Set entry triggers mechanically:

• Yield trigger: current yield above 5y average yield + alpha (default +0.5pp).
• Valuation trigger: target multiple reached (P/E, P/FFO, or P/FCF).

Execution pattern:

• Split orders: 40% -> 30% -> 30%.
• Pre-order blockers: if a candidate has any unresolved

pre_order_blockers[] (from WS-1/2/3 — variable/cut/suspension,

adjusted-EPS-unavailable, GAAP/Adj divergence, bank credit, utility

FFO/debt, event-scan failed/skipped, stale dividend, …) OR

t1_blocked is true, the first tranche is **blocked or downsized to a

≤20% tracking tranche** — not 40%.

• Sector cluster risk: when ≥ SECTOR_CLUSTER_WARN_COUNT same-sector

names pass (e.g. many small banks share one macro beta), emit a

portfolio-level CLUSTER-RISK warning.

• Require one-sentence sanity check before each *unblocked* add: "thesis

intact vs structural break".

8) Produce standardized outputs

Always produce:

• Screening table with the actionable verdict tier: CLEAN-PASS,

PASS-CAUTION, CONDITIONAL-PASS, HOLD-REVIEW, STEP1-RECHECK,

FAIL (synthesized by verdict.py from Step 1 + Step 2 + Step 4b +

blockers). Include evidence per row.

• One-page stock memo (use references/stock-note-template.md) with the

per-ticker provenance block (price/dividend/payout/event sources,

unresolved_blockers, evidence_refs[]).

• Limit-order plan with split sizing, blocker gate, and invalidation.
• Top-level run_context (profile, yield_floor_pct, safety_bias,

universe_source, excluded_asset_types) so a 3%-run result is never

silently reused inside a 4%-run.

Output

Return and/or generate:

• SOP screening summary in markdown.
• Underwriting memo set based on

references/stock-note-template.md.

• Optional plan artifact file generated by

skills/kanchi-dividend-sop/scripts/build_sop_plan.py in reports/.

• Optional Step 5 entry-signal artifacts generated by

skills/kanchi-dividend-sop/scripts/build_entry_signals.py in reports/.

Cadence

Use this minimum rhythm:

• Weekly (15 min): check dividend and business-news changes only.
• Monthly (30 min): rerun screening and refresh order levels.
• Quarterly (60 min): deep safety review using latest filings/earnings.

Multi-Skill Handoff

Run this skill first, then hand off outputs:

• To kanchi-dividend-review-monitor for daily/weekly/quarterly anomaly detection.
• To kanchi-dividend-us-tax-accounting for account-location and tax classification planning.

Guardrails

• Do not issue blind buy calls without Step 4, Step 4b and safety checks.
• Do not treat high yield as value before validating coverage quality.
• Use the regular forward yield for Step 1, never a special/TTM-inclusive

figure; near-floor + unconfirmed ⇒ STEP1-RECHECK, not FAIL.

• A failed/skipped event scan on a TRIGGERED name ⇒ HOLD-REVIEW + T1

blocked. Never silently skip Step 4b.

• Keep assumptions explicit; adjusted_eps/data missing ⇒ fail-safe

HOLD-REVIEW, never silent PASS.

Resources

• scripts/thresholds.py: single source of truth for all SOP

thresholds + SCHEMA_VERSION (downstream schema-evolution guard).

• scripts/dividend_basis.py: WS-1 regular/special/variable/freeze/cut +

Data Freshness Gate engine (pure, offline).

• scripts/payout_safety.py: WS-2 sector-aware GAAP/Adjusted/FCF payout

triad + completed-merger linkage.

• scripts/event_scanner.py: WS-3 isolated forward/recent corporate-action

scanner + materiality gate + pessimistic cap.

• scripts/verdict.py: WS-5 actionable-tier synthesis + run_context +

evidence_ref helpers.

• scripts/build_entry_signals.py: orchestrator (Step 5 targets + WS-1/2/3/5

integration). Flags: --yield-floor, --events-json, --profile,

--safety-bias, --universe-source.

• scripts/build_sop_plan.py: deterministic SOP plan scaffold generator.
• scripts/tests/test_golden_p0.py: P0 merge gate — end-to-end frozen

verdicts for CALM/ORI/CMCSA/MKC/CFR/cut (run via scripts/run_all_tests.sh).

• references/default-thresholds.md: human-readable threshold mirror.
• references/sector-step2-modules.md: Step 2 safety indicators by sector.
• references/valuation-and-one-off-checks.md: Step 3 valuation + Step 4 one-off.
• references/stock-note-template.md: one-page memo + provenance block.