Portfolio Review

Portfolio Review

Render the three numbered sections inline in the conversation, one pass — mechanism per references/design-tokens.md § Mechanism and output anatomy per references/rendering.md. Do not render into a document, canvas, webpage, side panel, file, or download.

Sequence — call, write, then call. Each section is written in full, visual and prose, before the next tool call is issued. Never batch the calls and write afterwards.

analyze_portfolio → write header + section 1analyze_scenarios → write section 2get_opportunities_and_issues ‖ get_fund_health_and_ideas → write section 3→ write closing

The two step-3 calls run together; nothing else runs in parallel. One visual per section, never one for the whole review.

• Scope check first. If the advisor named one specific analysis, call that tool and stop. This skill is for the open-ended ask — "review this", "run the numbers", "pull up their account".
• If no saved portfolio, household, model, pasted holdings or resolvable subject exists, route to guided-portfolio-builder or the appropriate lookup path before reviewing.
• Never make the advisor name a tool or a field.

Rendering minimums if references are unavailable

Do not tell the advisor a reference file was unavailable. If the host cannot load the reference files, still follow these minimums:

• Try the richest inline visual form first; markdown tables are fallback only.
• Snapshot metrics render as same-row stat tiles, not a two-column metric/value table.
• Scenario impacts render as a chart or bar visual, not a plain percentage table.
• Fund health renders as a multi-column card grid when cards are available. Every fund card includes the Funds to Explore tail; if none are returned, say No Funds to Explore returned.
• Opportunity signals render as status-tinted cards with an Available path, not as a plain markdown table, when cards are available.
• Clear counts require returned clear checks. If the response returns only fired rules, state only the number flagged.
• Fund-health commentary is visual-first: the fund cards carry each flagged fund's drivers, status, Funds to Explore, and links. Prose after the cards is at most three bullets: highest-priority fund, why it outranks the others, and coverage gaps.
• For red or flagged fund-health opportunities, encourage reviewing the flagged funds in chat. A methodology URL is labeled Fund health methodology, never as a fund-details tab or route.
• Closest model renders as a designed comparison prompt: model card, split-match note, and primary call-to-action Run side-by-side comparison in this chat. The match is on equity/fixed-income split only, not holdings. Do not render closest model as plain prose plus a markdown table when cards are available.
• Surface returned Advisor Center links. If any section returns comparison_url, methodology_url, fund URL, product URL, or route, include the relevant link near the block it supports.
• Never show tool names, field names, ids, or missing-reference commentary in the advisor-facing answer.

Links this domain carries

Cite them per references/rendering.md § Citing. The comparison view renders live under its own analysis date and fee settings, so its figures can differ from the review.

Step 0 — Resolve the subject

Every analysis tool takes subject: {type, id}. Both members required. No inline positions.

Route on what the advisor gave:

• Id already in context — use it. Do not re-list.
• A name, or nothing — call list_portfolios and list_household_portfolios, present the candidates, let the advisor pick. Never guess between two subjects sharing a name; total_market_value and holding_count tell them apart.
• A model — list_models, then pair model_id with its group's comparison_id_type.
• Pasted holdings or a spreadsheet — save first, see below.

Take comparison_id_type off whatever read or created the subject. Never build a TYPE:id string or comma-join ids.

Hold the subject's record before step 1 — the list_* entry, or the create_* / get_model response. An id that arrived with no record: call the matching list_* once and take its entry. The header and stat row read from this record.

Benchmark — only when the advisor named a model or benchmark in the ask. Resolve it through list_models to its own {type, id} and thread that one through steps 1 and 2. Never choose a benchmark for them.

Saving unsaved holdings

create_portfolio writes real data. State that, wait for an explicit yes. Then:

• Resolve identifiers first. CUSIP → cusip, ISIN → isin. A CUSIP in isin returns "no such security".
• Tickers and names go through search_securities. On resolution: ambiguous, present candidates and ask — never pick a share class for them.
• Spreadsheet: state the inferred column mapping, get confirmation. A blank or unparseable identifier — name the row and stop, never drop it silently.
• portfolio_type is required by rule though schema-optional. Ask; do not default.

Review the returned id as PROSPECT_ID.

Step 1 — Portfolio snapshot · analyze_portfolio

• nav absent = a model. Percentages only, no dollars.
• Household: account_contributions renders as Risk by account between allocation and equity geography, and leads the prose — which account carries the risk is the story.
• Benchmark named: call analyze_portfolio on the benchmark too for the benchmark tile row. Otherwise no benchmark row.
• Closest model — when the response carries closest_model, call list_models and take the row whose model_id matches its id. Hold that row and the subject's equity and fixed-income slice figures from asset_allocation. Nothing renders in section 1; both are carried to the closing. No closest_model: no closing block, and never substitute the named benchmark for it.

Write section 1 now, before calling analyze_scenarios.

Step 2 — Scenarios · analyze_scenarios

• Pass initial_investment only when the subject has no nav and the advisor named an amount.
• Pass benchmark only when one was resolved in step 0. Never pass include_contributors or include_all_scenarios — both are follow-ups.
• Display order and dollar basis per references/rendering.md § 2. Impacts are modelled, not forecast: never call one a prediction, a probability or a likelihood.

Observations — take each of these in order whose data is present, stop at five; write fewer than three only when fewer are present, never pad:

1. The worst case and its cost.
2. Whether the response's own worst scenarios are materially worse than the shortlist this deployment always reports — only when the response distinguishes the two.
3. The spread between worst and mildest.
4. Rate sensitivity on its own — only when a Treasury scenario is present.
5. Benchmark-relative direction — only when relative_impact exists.

Each bullet cites a figure.

Write section 2 now, before calling the step-3 tools.

Step 3 — Opportunities, fund health, Funds to Explore · get_opportunities_and_issues + get_fund_health_and_ideas

• Empty signals = nothing flagged. Say that; never omit the section.
• Flagged fund = one or more drivers fired. status is rendered as given, as a dot and words, and orders the cards; flag count orders within a status.
• Name non_fund_holdings and unknown_holdings by count. No position vanishes silently.

Commentary — point form, after the cards. Signals and fund-level drivers already sit in visual cards per references/rendering.md § 3; do not restate each card in prose. Write at most three bullets: which opportunity or fund to inspect first · why it outranks the others, citing the separating metric · coverage gaps such as non-fund or unresolved holdings. No paragraph per flagged fund.

Write section 3 now, then the closing.

Closing

• Caveats — inferred column mapping, resolved identifier, placeholder investment amount, equity-sleeve scope, benchmark assumed. Only those that applied.
• Closest model — only when step 1 carried one. Render the comparison prompt in references/rendering.md § Closest model, filled from the held list_models row and allocation figures. The primary next step is to run the side-by-side comparison in chat. The Advisor Center link is secondary when returned, not the main action. The match is on equity/fixed-income split only, not holdings — say so.
• Follow-ups — three to five, each one a tool this domain actually carries and each earned by something this review found: a model or benchmark comparison when a closest model was named, deeper risk, ESG or carbon, simulated performance, wealth projection, style factors, the full scenario set, or scenario contributors. Never offer one the findings do not motivate.
• Report/PDF handoff is reactive only. Do not proactively mention report generation in the closing. If the advisor asks for a report or PDF, route through ac360-capability-router.

Throughout

• Cite, attribute and report gaps per references/rendering.md — every section. The advisor decides; this skill organises what came back.
• Links are part of the output. When a returned Advisor Center, comparison, methodology, fund or product link supports a rendered block, surface it near that block. Do not finish a review with no links if any returned link exists.
• Never drop a section. Within a section, lead with its most material finding. The closing names the most material finding of the whole review.
• Read-only. create_portfolio, update_portfolio, clone_portfolio, the household writes and update_advisory_fee write real data. Ask first, every time, and say what will be written.
• One subject, all three sections. Never switch mid-review.
• Credential and gate failures are plumbing, not findings about the portfolio.
• Advisor-facing, not client-ready. Concise and analytical working material. No talking points, no reassuring framing, no marketing tone.
• No tool names, field names or ids in anything the advisor reads.