Portfolio Plan Builder

Portfolio Plan Builder

Interview-driven setup and maintenance of PORTFOLIO-PLAN.md — the file every other skill reads for the user's intent that no broker's MCP can know. Works over whichever broker(s) BROKERS.md names active — resolve any capability against reference/BROKER-CAPABILITIES.md. reference/READ-ONLY-POLICY.md (hard rule) and reference/RESEARCH-STANDARDS.md (frameworks, disclosure) apply.

Write scope. The one skill that writes a file the user owns: PORTFOLIO-PLAN.md at their project root, from reference/PORTFOLIO-PLAN.example.md. Nothing else — never a broker action, never another file, never an edit the user hasn't seen. (Note: BROKERS.md is a separate file the user edits themselves — this skill doesn't write it, and reads it only to know which broker(s) to pull grounding data from.)

A conversation, not a form. The plan's value is the thinking it forces, so probe and push back — a plan recording whatever the user said first is worth less than one that survived a few good questions. Interview in rounds of 2–4 questions, grounded in what they actually hold. Never dump the template as a questionnaire.

When this runs

• File doesn't exist → offer this before doing the analysis the user asked for. Don't proceed on guessed targets, and don't build it unasked either: say what's missing, what it unlocks, and offer (a) build it now, 5–10 minutes, (b) answer just the minimum for this one question, (c) proceed without it, stating the assumptions used.
• A needed section is blank or stale → fill that section, then hand back. Don't turn a rebalancing question into a full plan interview.
• The user asks to create, review, refresh, or change the plan → full or section-scoped pass, as they prefer.

Steps

1. Locate or create. Check the project root. If absent, copy the template and confirm PORTFOLIO-PLAN.md is git-ignored — it will hold real financial data. If present, inventory what's filled, blank, marked n/a (a decision — don't re-ask), and stale per its _Last reviewed:_ stamps.

2. Ground the interview in real data before asking anything. One batched pass over every active broker — noting that this covers equity and ETFs only unless INDmoney is active and its net-worth capability is verified working (see BROKER-CAPABILITIES.md): equity holdings, open F&O/intraday positions, available-margin, one batched LTP call, and fundamentals/screener capability for sector and market-cap classification (Groww-only today). Compute the current bucket split, top-10 weights, and sector weights.

   This turns abstract questions concrete: not "what's your mid/small-cap target?" but "you're at 41% mid/small-cap today, mostly in three names — is that where you meant to be?" Users answer the second honestly and guess at the first.

3. Interview in this order — the template documents each field; what matters here is the sequence and what to press on:

   1. Goals & horizon — what the money is for, when it's needed, what an unacceptable outcome looks like. Everything else follows, so start here even though the template starts elsewhere.
   2. Risk limits — appetite, then anchored against their book: "a 30% drawdown on today's holdings is about ₹X — hold, add, or sell?" Single-stock, sector, and per-fund/AMC caps, each marked hard or soft.
   3. Target allocation — propose a split derived from their answers and their current book, then let them correct it. Proposing beats asking cold, as long as the reasoning shows and it's clearly a draft. Set drift bands per bucket.
   4. Rebalancing rules — cadence, threshold, correction order, selling constraints, untouchables.
   5. Fixed income + inventory — targets, then the inventory table. Say plainly that direct bonds/NCDs/FDs/SGBs are invisible to every broker's MCP configured here, so this table is the only source of truth for bond-ladder-planner and rate-watch.
   6. SIP register — this is the whole mutual fund inventory, not just SIP amounts: no broker's MCP has verified working fund data today (Groww has none; INDmoney's is unverified — see BROKER-CAPABILITIES.md), so ask for every fund held with units, average cost, and current value, then the SIP rows on top. Say plainly that without this table their fund sleeve is invisible to every skill and portfolio-review will be reporting on the equity book alone. Flag lump-sum-only funds and paused SIPs explicitly.
   7. Tax context — note where lot-level data isn't available rather than assuming.
   8. Position theses — for the top holdings, why held plus a checkable invalidator (a number, event, or date). Highest-leverage section: without one, earnings-watch and portfolio-review can only report that the price moved. If they can't name one for a large position, that's the finding — record it as an open question rather than inventing a thesis.
   9. Exclusions, deployable capital, income goal, watchlist, output preferences — quick round, mostly preference.

4. Grill — the part that earns the file. Reflect contradictions back with the numbers, once each, as a question with both branches rather than a verdict on their judgment:

   • Targets that don't sum to 100%, or a fixed-income target that contradicts the bucket table.
   • Stated appetite vs. the actual book (self-described conservative, 60% small-caps), or vs. behaviour they describe.
   • Horizon vs. known outflows — money needed in 18 months sitting in a seven-year bucket.
   • Limits already breached today: is the limit the intent, or is the position? One has to move.
   • Bands so tight they'd trigger constant trading, or so wide the plan never binds.
   • Overlap they may not see — funds converging on the same top holdings, an ETF duplicating direct holdings.
   • An income goal the book can't plausibly produce; show the projected figure against it.

   Accept "yes, deliberately" and record it as such, ideally in the decision log, so no skill re-litigates it later.

5. Write it. Fill the template, keeping its exact section headings and anchors — other skills key off them.

   • Only what the user actually said. No inferred numbers, no placeholder targets. Undecided → blank; decided there's no constraint → n/a.
   • Follow the template's conventions block (dates YYYY-MM-DD, rupees, percentages of current market value, exchange tickers).
   • Stamp every section touched with today's _Last reviewed:_.
   • Show the diff and get confirmation before writing, and honour the file's own "may skills propose edits" preference on later runs.
   • Never commit it, and never echo real figures into any example or tracked file.

6. Close the loop. State which sections are now filled, which are still blank and what that blocks (see the template's "How the skills use this file" table), and the natural next step — usually rebalancing-planner against the fresh targets, or back to the original question. stock-research is the way to test any shaky thesis that came out of the interview.

Maintenance mode

For an existing plan, audit rather than re-interview:

• Sections stale past ~6 months (a quarter for tax and SIP), by their own stamps.
• Plan vs. reality now: limits breached, buckets outside their bands, theses that have hit an invalidator, watchlist triggers fired, decision-log rows due to revisit.
• Fixed-income rows that have matured, and SIPs that changed.

Present that as a short findings list — formal version: reference/templates/portfolio-plan-builder.md — then offer to update only the rows that moved. A yearly full pass is worth it; a monthly one isn't.

Presentation

Lead with what the interview covers and roughly how long. Keep rounds short — questions as a numbered list, one line each, with the grounding number attached where there is one. Close with the filled/blank summary table and the next step. No disclosure block for the interview itself; if step 4 ends up giving a positioning opinion, it applies to that part.