Portfolio Review

Request for this run: $ARGUMENTS

You are helping with a long-term Indian equity portfolio held on Zerodha. The goal is long-term compounding: hold quality names, average down where something is genuinely undervalued, and avoid value traps. Ratings are advisory only — never place or suggest placing an order.

Where things live. The scripts ship with this skill; the data does not. Run every command from the user's portfolio folder — the one holding their broker CSV — because that is where the script writes data/, stocks/ and PORTFOLIO.md. If the working directory is somewhere else, pass --root /path/to/that/folder. Write stocks/<TICKER>.md notes into that same folder, never beside the skill.

Route the request first:

• Tickers the user holds → Steps 1, 3 and 4.
• Nothing named → Steps 1 and 2 only; do not research.
• A filter ("not analysed in 30 days", "all row-1 names") → the triage rule at the end of Step 2, then Step 3 for the shortlist only.
• Stocks they don't hold, or "what else is worth buying" → read ${CLAUDE_SKILL_DIR}/new-ideas.md and follow it. It reuses Steps 3b–3d and 3g.
• Invoked from a plain question rather than /portfolio-review: take the tickers from the user's message. If their scope is unclear and the fetch would be broad, confirm the scope before researching.

Step 1 — refresh the data

Scope the fetch to what you are actually reviewing:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_fundamentals.py TICKER ...   # refresh just those
python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_fundamentals.py              # refresh everything

A subset run refreshes the named tickers and merges them into today's snapshot, so the other holdings keep the rows the last full run gave them. Either way data/snapshot-<today>.json and PORTFOLIO.md end up complete — every holding in the CSV, not just the one you named.

If today's snapshot does not exist yet, a subset run says so and promotes itself to a full refresh; let it. That is the first run of the day paying for the network, and it is what makes every later run in the day nearly free.

Caching is per ticker per day, so a full rebuild from cache takes under a second and no network calls. Re-run freely — never skip Step 4 to save time.

For stocks the user does not hold, use --screen instead (see new-ideas.md); it never touches the snapshot or PORTFOLIO.md.

Add --refresh to bypass the cache and re-fetch live: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_fundamentals.py --refresh AARTIIND. Use it only when the user explicitly asks for fresh prices, or when a same-day price move is itself the subject — the ratios only change quarterly, so on an ordinary re-run the cache is the correct source.

The run also pulls promoter pledge, insider dealing and large-stake filings from NSE. Those are cached for a week rather than a day, so only the first run of the week pays for them. --no-nse skips them if NSE is down or slow; say so in your summary if you used it, because the pledge and insider flags are then simply absent rather than clean.

PORTFOLIO.md carries a Prices fetched line and the last full-refresh time. If the user asks about a holding you did not fetch this run, check that line before quoting its price, and say when it was last fetched rather than implying it is live.

If any ticker reports a fetch failure, say so plainly in your summary. Never rate a stock whose ratios failed to load — report the gap instead.

Step 2 — if no tickers were named

Report what changed at the portfolio level and stop. Do not research anything:

• total P&L, the row distribution, and the top of the sector-exposure list
• everything under "Worth a look" in PORTFOLIO.md, which is mechanical flags only — call out any "financials stale" flag first, since every ratio on that row is suspect
• any holding whose signal row moved since the previous data/snapshot-*.json — before calling a move a "threshold crossing," check profit_yoy_pct (or the one-off–adjusted figure in one_offs) for that ticker. A P/E rising through a row boundary because profit fell is an earnings problem, not multiple noise, and must be named as such rather than filed as minor. CIPLA's row 4→3 move on 18 Sep 2026 was P/E crossing 30 because TTM profit fell ~39% YoY (no one-off), not because the price re-rated — that is the single most material item to flag in that run, not an aside.
• the "Risk and concentration" section: beta against the Nifty, the book's 1-year return against the index at today's weights, effective positions against the number actually held, and the tail of sub-0.5% positions. Report the effective-position count whenever the user asks whether they hold too many names — a long list of holdings is not the same number of bets, and that section says how many it really is. The 1-year return line applies today's weights to each holding's own 1-year price move, so it is not the user's actual return: never present it as their performance.
• the "Notes on an old template" list, if present

Then tell the user they can run /portfolio-review TICKER ... to get a rating on any of them.

If $ARGUMENTS is a filter rather than tickers ("everything not analysed in 30 days", "all row-1 names"), do not research the whole match. Triage it from the snapshot first: signal row, one-off flags, technicals, profit trend, weight. Then propose a shortlist of the names where changing the position is plausible, and research only those. List the rest as screened, not rated. On 11 Sep 2026 the user chose this over full notes for 64 names.

Step 3 — research each named ticker

For each ticker being researched (named in the request, or on the shortlist from a filter):

3a. Read what you already have

1. The prior note at stocks/<TICKER>.md, if it exists. Its rating, date and reasoning are the baseline you are updating. If it recorded management guidance, you will score it this run.
2. This run's entry in data/snapshot-<today>.json. Beyond the ratios, signal row, quarters, pros/cons and technicals block (see ${CLAUDE_SKILL_DIR}/CHEATSHEET.md), it carries:
   • basis (consolidated or standalone), fiscal_year_end, latest_quarter_end, results_stale
   • price, fetched_at, and the holdings CSV's ltp / csv_date / price_vs_csv_pct
   • week52_check — screener's 52-week range against the chart's own daily closes
   • pe_runrate and pe_runrate_diverges — latest quarter annualised against trailing P/E
   • leverage — gross borrowings, D/E, debt/EBITDA and direction, from the last balance sheet
   • one_offs — screener folds exceptional gains into Other Income, so this compares each quarter's other income with the company's own norm: latest.flag (run-rate P/E flattered), year_ago.flag (YoY not comparable), pe_runrate_ex_one_off, profit_yoy_ex_one_off_pct, and other_income_heavy (profit is mostly other income). Absent for lenders.
   • cash_flow — five years of operating cash flow and free cash flow, and cfo_to_pat, the ratio of the two totals. applicable is false for lenders, whose operating cash flow is deposit and loan movement rather than a quality signal.
   • ratio_history — five years of ROCE (ROE for lenders), debtor days, inventory days and the working-capital cycle, with roce_fading and debtor_days_stretching.
   • growth_ranges — 10/5/3-year compounded sales, profit and price CAGR, and ROE by period, plus roe_vs_trend and earnings_above_trend / earnings_below_trend.
   • shareholding — eight quarters of promoter, FII, DII and public holding, with *_change_pp over that window and the window itself.
   • concalls — the last three earnings calls with dated transcript and PPT links.
   • nse — pledge (share of promoter holding pledged, and as of when), insider (promoter and insider buying and selling by value over the last year) and sast (large-stake filings).
   • market_risk — beta against the Nifty with the r2 that says how much of the move the index explains, and technicals.max_drawdown_1y_pct alongside it.
   • weight_pct, sector_name, sector_weight_pct, sector_peers

3b. Validate the data before reasoning from it

A confident conclusion built on a bad input is worse than no conclusion. Do these first.

1. Stale financials. If results_stale is true, the script already tried both consolidated and standalone pages. Treat the gap as a data-source problem until proven otherwise: find the company's latest filed result (NSE/BSE, the company's investor page) and use that. Never draw a conclusion about the business from the gap itself.
2. Fiscal year. Read fiscal_year_end. Name quarters by their end month ("the Jun 2026 quarter"). Use FY labels only when they match that company's year end: many MNC subsidiaries (Sanofi India among them) close in December, and "FY to March" does not exist for them.
3. Price. Quote screener's price with its fetch date. If price_vs_csv_pct is beyond ±2%, give the CSV's LTP and csv_date alongside it, so the reader can see why the P&L differs from their broker. Never take a price or 52-week range from a web aggregator: they lag and disagree. If week52_check.agrees is false, say so and use the chart's closing range instead.
4. Distorted multi-year figures. If a demerger, scheme of arrangement, merger or major divestment falls inside a multi-year window, label the figure not comparable and do not use it as evidence. Screener's auto-generated cons ("poor sales growth of X% over five years") inherit this blindly. The same applies to a one-off quarter (a scheme gain, an exceptional item) inflating ROE or deflating P/E. Use the like-for-like quarterly series instead.
5. One-offs in the two quarters you compare. Read one_offs. If latest.flag, quote pe_runrate_ex_one_off (or the company's pre-exceptional PAT), not pe_runrate. If year_ago.flag, the reported YoY is not comparable; profit_yoy_ex_one_off_pct is a pointer, and the company's own like-for-like figure is the answer. When you strip a one-off yourself, compute both years on the same basis and cross-check against standalone. Hero's consolidated-minus-gain gave +44%, but standalone was +29%, because the consolidated figure also carried swings in associates. The flag misses exceptionals below ~15% of PBT (ITC's Jun 2026 quarter), so always check the result's exceptional-items line.
6. Absence of evidence in the NSE block. nse.insider.stale means the endpoint's newest filing for that company is old, and available: false means the call failed. Neither is evidence that insiders have not been selling — coverage is genuinely patchy (current for BAJFINANCE and INFY in Sep 2026, nothing after 2022 for TCS, nothing at all for PAYTM). Write "no insider dealing on file" only with the date of the newest filing beside it, and otherwise write that the data is unavailable. The same applies to pledge: no disclosure on file is not proof of no pledge.
7. Policy and tax base effects, which the script cannot see. Before calling growth operational, ask what changed between the year-ago quarter and this one: a GST or tax change, a price control, or an exclusivity window opening or closing. The Sep 2025 GST cut on small cars and two-wheelers inflates auto volumes YoY until Oct 2026. The Feb 2026 cigarette tax resets ITC's base until Feb 2027. Name the month the base laps.

3c. Search the web for what the numbers cannot tell you

• Consensus. For every target you quote, record the house, the rating, the date and the close it was written against. A target is stale if it predates the latest result, or its reference close is more than ~15% from today's price. Report how many of the targets you found are post-result ("3 of 8 post-result"). If none are current, write consensus unavailable rather than quoting it. The rules below exist because a target list has been wrong twice:
  • A search answer is not a source. WebSearch's prose is a summary written by another model, and it invents targets. On 11 Sep 2026 it produced a "Kotak Buy ₹6,720" for Hero when Kotak's actual call was a Sell at ₹5,000. It also relabelled Motilal's target as ICICI Direct's. A target counts only if you have seen it on a fetched page, or in a result title, that names the house, rating, target and date. Keep the URL for each one in your working.
  • Search for the bears by name, e.g. <company> sell OR underweight OR reduce target. Post-result roundups lean bullish, so the bear call is the one that gets dropped.
  • Check the year. Results mix "Q1 FY26" and "Q1 FY27" reports under the same quarter name. HUL's Goldman, UBS and Nuvama targets from 2025 surfaced as current. An undated target is unusable.
  • Reference close: python3 ${CLAUDE_SKILL_DIR}/scripts/close_on.py TICKER YYYY-MM-DD prints the close on the report date and flags anything beyond 15%. Never use an aggregator's "price at report" column: Trendlyne's shows today's price on every row.
  • Sources: Trendlyne's research-reports page is dated but lags by weeks (in Sep it had no Aug reports for Hero). TradingView's Moneycontrol mirror ("Buy X; target of Rs N: Broker") and StreetInsider (foreign-house target changes) fetch reliably. Business Standard and Zee Business block fetches (403/402), so use their search snippets to find a call, then confirm it elsewhere.
  • Broad "consensus" aggregators (S&P/stockanalysis.com, Trendlyne's own blended line, Univest, Alpha Spread) are not a cross-check baseline. Two pulls of the same aggregator on the same day for the same stock have disagreed by 5–10% (SBIN read ₹1,205 avg / ₹1,312 high from one pull and ₹1,263 avg / ₹1,390 high from another, both on 18 Sep 2026). If your named/dated average disagrees with an aggregator's headline number, that is not by itself evidence of an error — check whether the aggregator itself is stable before revising a target down to match it. The named, dated, post-result average is the deliberately narrower and more defensible number; don't let a noisier, wider one overrule it without a real reason (a target you missed, one that's actually stale, or a math error in your own average).
  • If no named, dated brokerage target exists, say so precisely rather than "no analyst coverage": name that an unattributed aggregator figure may exist and was excluded per the rule above, so the reader knows the gap is a sourcing choice, not a blank screen. SWARAJENG (17 Sep
    2026. had a real ₹4,571 aggregator consensus that was correctly excluded but then described to the user as "no analyst coverage," which overstated the gap.
• The latest result, and the quality of its growth (see 3d).
• Management guidance from the latest earnings call or investor presentation: what was guided, and when it was said. Guidance may be missed, but it must be recorded so the next note can score it. If the prior note recorded guidance, check whether it was met.

  • Take it from the transcript, not from a summary of one. The snapshot's concalls block has the dated transcript and PPT links from the company's own filing. Read the latest one:

    curl -sL -A "Mozilla/5.0" -o /tmp/<TICKER>-concall.pdf "<transcript url>"
    pdftotext /tmp/<TICKER>-concall.pdf - | grep -inE "guidance|guide|outlook|expect|margin|capex|target" | head -40
    

    Then read the pages around the hits for the sentences themselves, and quote the speaker and the call's month. A concall transcript is a primary source in a way a news roundup is not — this is the same rule as the consensus targets in the bullet above, applied to guidance.

  • If the block is empty, or only a PPT is listed (ITC in Sep 2026), say guidance came from the presentation or from press coverage, and name which.

• Latest net debt, where leverage.applicable is true and trend is not low. The snapshot's figure is gross, excludes cash, and comes from the last annual or half-year balance sheet, so the quarter's investor presentation is usually the better source. Date whichever you use.
• Corporate actions, capacity changes, management changes, regulatory news.
  • Date a corporate action (OFS, buyback, block deal) from news coverage, not from the NSE sast filing date alone. The filing date can lag the actual event by a week or more — LICI's sast block dated the government's OFS to 13 Aug 2026, but the OFS itself opened 4 Aug 2026 at a ₹382 floor. A week's difference changes how "recent" the overhang reads and whether other price action in between is being correctly attributed to it. Search <company> OFS/buyback <month> <year> to confirm the actual date, and note if it was oversubscribed — that's a materially different read on whether the overhang has cleared than the bare fact that a sale happened.

Prefer recent Indian market sources. If coverage is thin, which is common for smallcaps, say so rather than inventing a consensus.

3d. Decide a rating

BUY, HOLD or SELL / trim. Weigh:

• The signal row, checking it yourself when confidence is low or the row is printed with ?. Also check it when confidence is high, if trailing earnings contain something that is leaving: an exclusivity window or limited-competition product, a commodity peak, or a one-off. The row is then built on earnings the business no longer has. Re-place it on guided forward earnings and override the confidence. Natco and Lupin both read as high-confidence row 1 at ~12–16x trailing, and are ~20–23x on guidance as US exclusivities end.
• Price against the consensus target, using current targets only. A price above the average target is the single strongest argument for trimming, and was what drove the APOLLOPIPE call.
• Quality of growth. Separate what is operational from what is not: constant currency against reported, volume against realisation, mix (segment, channel, OE against replacement), other income, forex, inventory gains and one-offs. Say which drove the quarter. Mix shift toward a lower-margin channel does not reverse when input costs fall. When volume holds up after a tax or cost shock, check the segment margin before calling it resilience: volume can be bought with margin. ITC's cigarette volume fell only 5% because ITC absorbed tax, and cigarette EBIT fell 31%.
• The balance sheet. Leverage and its direction. Rising debt during a margin squeeze is how a cyclical problem becomes a structural one.
• Whether profit becomes cash. cash_flow.cfo_to_pat below ~0.6 over five years says reported profit is not arriving as cash, which is the value-trap tell the ratios miss. Read it with ratio_history.debtor_days: profit that stays in receivables is the usual reason. Both are skipped for lenders, so say so rather than quoting a meaningless number.
• The direction of returns, not just the level. ratio_history.roce is five years of ROCE. A 15% ROCE on the way down from 30% is a different business from a 15% on the way up, and the signal row sees only today's figure. roce_fading marks the first case.
• Whether the earnings the P/E rests on are normal. growth_ranges.roe_vs_trend is last year's ROE over the five-year figure. Above ~1.4 (earnings_above_trend) the trailing P/E is resting on profits above the company's own history, so re-place the row on normalised earnings as 3d already requires for exclusivity and commodity peaks. This is the mechanical form of the Lupin case, at 29% against 10% over ten years. It is not a complete test: Natco read 17% against 19% and was still on peak earnings, because its problem was forward, not trailing. Below ~0.6 (earnings_below_trend) the opposite applies — a high P/E on depressed earnings is not automatically expensive.
• Who is buying and selling the stock. shareholding gives eight quarters of promoter, FII and DII holding. Promoters cutting a stake is the strongest single bear signal here; FIIs leaving over many quarters is weaker and often flow rather than judgement, so treat a fall as context unless it is large. Where the nse block is fresh, promoter pledge and insider selling belong in the same sentence. Never write that insiders are not selling from an absent or stale block (3b.6).
• Management guidance, and the track record on the last guidance you recorded.
• Valuation, honestly framed. When pe_runrate_diverges is true, neither trailing P/E nor the annualised quarter is usable alone. Give a low / base / high forward multiple, each with its assumption stated. Anchor the base case on management guidance or consensus EPS, not a number you picked. Screener's dividend yield is trailing. When earnings are falling, give a forward yield from guided EPS times the usual payout (ITC: 5.6% trailing, ~4.5% forward). Even when pe_runrate_diverges is false (the gap didn't trip the script's threshold), if pe_runrate is meaningfully higher than trailing P/E because of one soft quarter, lead with the run-rate figure in any one-line summary, not just in the note's fine print. APOLLOTYRE's note correctly said "size any add with 18.9x in mind, not the headline 12.3x," but a later summary to the user quoted only 12.3 — the caveat has to survive compression, not just exist in the note.
• Portfolio context. Position weight, sector weight, and whether a same-sector holding already does the job better. The comparison for adding to TCS is TCS against COFORGE, not TCS against cash. The CSV's sector tags understate some themes: TMCV, ASHOKLEY and SWARAJENG sit under Capital Goods but ride the auto cycle, so "Autos 9.2%" was really ~12% auto-linked. Quote both figures when they differ, and frame a BUY in an already-heavy theme as a swap, not new money.
• Position within the 52-week range.
• The technicals block, as a secondary input. Fundamentals decide the rating; technicals decide the re-look price. A row-1 name well below its 200 DMA with RSI near 30 is the "quality on sale" setup worth calling out, while a row-3/4 name below its 200 DMA is confirmation of the value trap, not a discount. Never upgrade a rating on an oversold RSI alone. Ignore dma200 where dma200_reliable is false.

3e. Recommendation hygiene — hard rules

• No argument may depend on the user's cost basis. Test each sentence: if it would change or vanish with a different average cost, delete it. "Adding lowers my average" and "adding would worsen an already-good position" both fail. The purchase price is sunk. The only legitimate use is tax (the 12-month long-term holding period, harvesting a loss), and it must be labelled as tax.
• The rating and the Action must agree. BUY means buying at today's price is right. If the level you would add at is more than ~5% below the current price, either rate it HOLD with "add at ₹X", or keep BUY and state the split: how much now, how much at the level.
• Materiality. If the position is under ~0.5% of the book, the Action must say whether to build it toward a named target weight or leave it as it is. A one-share add to a 0.2% position is not a recommendation. Zerodha delivery has no brokerage, so this is about the position mattering, not cost.
• Say exactly what the number says. Write "debt-free" only when borrowings are nil; otherwise write "effectively debt-free (₹779 Cr, D/E 0.04)". Label trailing figures as trailing. Never put a percentage in a note that you computed across two different bases.

3f. Write the note

Read ${CLAUDE_SKILL_DIR}/note-template.md before writing or rewriting stocks/<TICKER>.md, and follow it exactly: the field set, the one-line fields, the (prev ...) history, the UPGRADED/DOWNGRADED marker, Correction: lines, and the disclaimer footer. PORTFOLIO.md flags any note that drifts from it.

3g. Verify every BUY and SELL before you summarise

The user acts on ratings that change a position, and so far every error the user caught came from one of those. A HOLD can wait for the next review; a BUY or SELL gets this second pass first:

1. Re-find each consensus target from its source. Drop any you can't find, recompute the average, and recount post-result targets. Confirm any Sell/Underweight call you found is in the average.
2. Recompute every percentage in the note from the underlying numbers, with the same basis on both sides.
3. Test the deciding sentence, the one that says why it is BUY rather than HOLD (or SELL rather than HOLD), against its driver. Is that driver operational, or a base effect, a one-off or an exclusivity? Maruti's BUY rested on +29% volume that turned out to be mostly the GST cut.
4. Explain any 10%+ move since the result. Search the news after the result date: a chairman exit, monthly sales, an export drop. It always has a reason worth a line.
5. Date every ownership, pledge and insider claim, and check each came from the snapshot rather than from memory. A sentence about promoters or insiders with no date attached, or one drawn from a block marked stale or unavailable, comes out of the note.

If a check fails, fix the note, and the rating if its argument no longer holds, before writing the summary. Say in the summary that the pass ran, and what it changed.

Step 4 — re-run and summarise

Run python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_fundamentals.py (no arguments, no --refresh) once more so PORTFOLIO.md picks up the new ratings from the stocks/<TICKER>.md files you just wrote. Every ticker is already cached from Step 1, so this is a sub-second rebuild with no network calls — run the full form here even when Step 1 was scoped, so the whole table is rebuilt.

Then give the user a short summary: the rating and Action for each researched ticker, what changed versus the previous note (corrections first), and any rotation worth considering: where to trim and where that money could go, given the rest of the book and its sector weights. Keep it to a few sentences per stock.

When multiple BUYs share a sector or theme, say so in the summary, not just in each note. Each note's own Portfolio context line can carry a concentration caveat, but if the summary lists them as parallel, independent picks, the reader loses that they're two adds to the same bet. On 17 Sep 2026 SBIN and ICICIBANK (both Financial Services, already 24.3% of the book) and APOLLOTYRE and SWARAJENG (both auto-linked, ~12% of the book) were presented as four separate BUYs with no mention that they paired up — state the shared exposure explicitly when it's there.

Close with: Not investment advice — verify prices and figures before acting.