agentleFS
tusharagg1tusharagg1/aifolimizer/.claude/skills/position-review/SKILL.md

position-review

Skill2 starsChanged 5 months agoMITView on GitHub

Review a holding (or sweep top holdings) and return a HOLD / TRIM / SELL verdict with entry/stop/target. Routes each name to the right analysis - earnings-analyzer if earnings are imminent, earnings-postmortem if it just reported, the adversarial bull/bear pipeline for high-conviction names, lighter stock-analysis otherwise. Use when the user asks "should I hold or sell X", "review my NVDA position", "review my top holdings", "position review", or for an automated nightly top-holdings sweep (the unqualified whole-book "review my holdings" belongs to portfolio-review). Logs each verdict for forward tracking.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before routing, load prior decisions so verdicts stay consistent across sessions:

    line 18
  2. Very specific:

    Call get_profile FIRST. Then gather routing signals (parallel):

    line 25
  3. Distinctive:

    After the routed analysis yields a lean, cross-check exits:

    line 41
  4. Distinctive:

    Run the heavy adversarial pipeline inline (main context) for at most the 1-2 highest-stakes flagged names.

    line 52
  5. Distinctive:

    Never invent data - if a routed tool returns empty, say "data unavailable" and lower conviction.

    line 75
  6. Distinctive:

    Don't run the adversarial pipeline for more than ~2 names in one sweep - token cost balloons and the nesting limit blocks parallelism.

    line 83

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Position Review (routing orchestrator → hold/sell verdict)

What this is

A router, not a new analysis. It picks the cheapest sufficient analysis per holding and emits one decision. Claude is the orchestrator - it calls shared MCP tools and, for deep names, runs the adversarial pipeline inline. Skills are not self-invoked.

Modes

  • Single ticker ("review my NVDA position"): route + verdict for that one name.
  • Sweep ("review my holdings" / automated nightly): take top-N holdings by weight (default 6) and route each. Keep total output tight.

Stage 0 - Decision Memory (load FIRST)

Before routing, load prior decisions so verdicts stay consistent across sessions:

  • mcp__aifolimizer__get_cross_ticker_lessons with max_lessons=3 - portfolio-level win/loss patterns
  • For each name reviewed, load mcp__aifolimizer__get_ticker_decision_history (ticker=…, max_decisions=5) and mcp__aifolimizer__get_ticker_reflection (symbol=…, n=3).

Reconciliation rule: if a prior decision exists and your new read flips it, state explicitly WHY it changed (new data / catalyst / price move). Never silently contradict a logged decision - that drift is exactly what this prevents.

How to run

Call get_profile FIRST. Then gather routing signals (parallel):

  1. mcp__aifolimizer__get_portfolio - holdings, weights, cost basis, return
  2. mcp__aifolimizer__get_personal_context - province / marginal_tax_rate_pct / account_waterfall to ground per-name tax + account framing. If present=false, note the framing is generic and suggest the profile-setup skill.
  3. mcp__aifolimizer__get_earnings_calendar - earnings proximity per name (pass watchlist symbols= only if reviewing a non-held name)
  4. mcp__aifolimizer__get_triggered_alerts (since_hours=48) - recent price/RSI/concentration flags
  5. mcp__aifolimizer__get_earnings_results for names that may have just reported

Routing table (apply per ticker, first match wins)

ConditionRoute toWhy
Earnings within ~7 daysearnings-analyzer flow (get_fundamentals + get_technicals + expected move)Pre-earnings risk dominates the decision
Reported in last ~5 days OR surprise flaggedearnings-postmortem flow (get_earnings_results + get_news_headlines)Beat/miss reaction sets the near-term path
Weight ≥ 8% OR a triggered alert OR user asked for depthadversarial-research pipeline (adversarial-research) run INLINEHigh stakes → full bull/bear/risk debate
Everything elsestock-analysis flow (get_fundamentals + get_technicals + get_positioning_signals)Cheap single-pass read is enough

After the routed analysis yields a lean, cross-check exits:

  • Verdict SELL/TRIM on a Non-Reg loss → call get_tax_loss_candidates; note harvest opportunity + 30-day superficial-loss caution.
  • Any BUY/ADD lean → get_positioning_signals crowding guard (consensus ≥70 → downgrade, don't add into the crowd).

Once the verdict is set, get concrete levels from mcp__aifolimizer__get_trade_ticket (ticker, action=<verdict: HOLD|TRIM|SELL>, conviction). It pulls cost basis + held quantity from the live session, so:

  • HOLD → exit_ladder (T1/T2/T3 profit-taking from current price) + stop_loss_price + position.stop_below_cost. This is the "where do I take profit / cut loss" plan.
  • TRIM/SELL → market-exit ticket (no ladder by design - show stop_loss_price only). Render these instead of hand-computing stop/target.

Subagent-nesting constraint (important for the sweep)

Subagents cannot spawn subagents. The adversarial pipeline already spawns 6 agents, so in a multi-holding sweep you CANNOT nest it per ticker. Pattern:

  • Run the heavy adversarial pipeline inline (main context) for at most the 1-2 highest-stakes flagged names.
  • For the rest, run the flat single-pass stock-analysis/earnings flow inline.
  • Bound cost: never run more than ~2 deep pipelines per sweep; downgrade the rest to flat.

Output (per ticker)

TICKER · weight X% · VERDICT: HOLD / TRIM / SELL · conviction
  route: <earnings / postmortem / adversarial / stock-analysis>
  held:  avg $.. · ret +X% · stop below cost? Y/N
  HOLD → exits: T1 $.. (sell 40%) · T2 $.. (35%) · T3 $.. (25%) · stop $Y
  TRIM/SELL → stop $Y  (no ladder by design)
  reason: <1-2 lines - the decisive factor>
  tax: <Non-Reg harvest note if SELL/TRIM at a loss, else omit>

Sweep mode: lead with a one-line roster (HOLD x4 · TRIM x1 · SELL x1), then the blocks, worst-conviction-holds and all SELL/TRIM first.

After output - log decisions

For each name log the verdict (this feeds the forward win-rate / track-record loop). Pick ONE tool - they take different params:

  • Default path → mcp__aifolimizer__log_recommendation: skill="position-review", action, conviction, rationale, and levels as target_pct + stop_pct (PERCENT from entry; entry is captured live at call time - do NOT pass absolute prices).
  • Adversarial route ran → mcp__aifolimizer__log_trade_decision: skill_used="position-review", action, conviction, entry_price / target_price / stop_price (ABSOLUTE prices), thesis_summary.

Rules

  • Verdicts are HOLD / TRIM / SELL only - this skill reviews EXISTING positions; it does not open new ones (that's top-trades-today / cash-deployment).
  • Never invent data - if a routed tool returns empty, say "data unavailable" and lower conviction.
  • Sweep output under 500 words total; single-ticker under 300.
  • Respect get_profile account types for tax framing; do not hardcode capital.

Gotchas

  • Earnings calendar is yfinance - sparse for TSX names. For .TO holdings with large weight, note "verify earnings date on IR page."
  • get_triggered_alerts is read-only (does not re-evaluate). If the user wants fresh evaluation, mention run_alerts_now - only if asked (adds latency).
  • A name can match two routing conditions (e.g. big weight AND earnings soon) - earnings proximity wins; fold the size concern into the reason.
  • Don't run the adversarial pipeline for more than ~2 names in one sweep - token cost balloons and the nesting limit blocks parallelism.
  • TRIM verdicts: the engine omits target/RR by design - show stop only, don't flag the null as an error.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.