How Seentio Maps Strategy Names to Templates

When you ask Claude Code "build me Greenblatt's strategy," there's a quiet but important step happening behind the scenes: Claude doesn't compose the strategy spec from training memory. It looks the name up in a curated dictionary first. This guide explains why we built that dictionary, how it works, and what happens when a name doesn't match.

What you'll learn

How the lookup_strategy MCP tool turns fuzzy user requests into precise template references, the alias dictionary that powers it, and the fallback path when nothing matches.

Why a lookup tool exists

Picture two users:

• User A says: "Build me Magic Formula." They mean Joel Greenblatt's textbook value-quality strategy from The Little Book That Beats the Market. Earnings yield + return on capital, top 30, hold a year.

• User B says: "Build me Magic Formula." They read a quant blog last week that called a single-factor low-P/E screen "the magic formula." They want low P/E, top 50, hold quarterly.

If Claude infers the spec from its training knowledge, it'll pick Greenblatt's version every time (because that's the dominant reference in academic literature). User B will get something they didn't ask for. They might not notice until the backtest comes back looking unfamiliar — or worse, until real money is rebalanced into the wrong portfolio.

This is what we call the name-collision problem. Strategy names are ambiguous; AI inference is non-deterministic; the combination produces silent mistakes. The fix is a deterministic name-to-template lookup with explicit aliases and disambiguation.

How lookup_strategy works

When the user mentions a strategy by name, Claude calls lookup_strategy(query=<their words>) as the first move. The tool hits our backend's /api/strategies/templates-lookup endpoint, which scores each template against the query.

Scoring tiers

The "confident match" rule

A match is declared confident when: - The top score is ≥ 0.7, AND - No other alternate scores ≥ 0.55 (no near-tie ambiguity)

If both conditions hold, the response shape is:

{
  "query": "Greenblatt",
  "confident_match": { "id": "greenblatt_magic_formula", "name": "...", "score": 0.95 },
  "alternates": [],
  "guidance": "Use confident_match.id with get_template..."
}

Claude reads the confident match and proceeds to fetch the full spec via get_template, then run dry_run_ranking, then create_strategy.

When matches are ambiguous

If the top score is below 0.7, OR there's a near-tie (multiple alternates above 0.55), the response surfaces alternates instead of picking one:

{
  "query": "value quality",
  "confident_match": null,
  "alternates": [
    { "id": "greenblatt_magic_formula", "score": 0.65 },
    { "id": "top_n_large_cap",          "score": 0.55 }
  ],
  "guidance": "Multiple matches — ASK the user which one they meant before creating."
}

The MCP system prompt tells Claude: when there's no confident match, don't guess — ask the user to pick from alternates by name and description. This forces an explicit confirmation rather than a silent inference.

The current alias dictionary

Today (May 2026) we ship four templates with the following aliases:

Greenblatt Magic Formula

• greenblatt, magic formula, joel greenblatt, little book that beats the market, earnings yield roic, value quality, ebit ev

Top N Dividend

• dividend, high yield, income, dividend champions, dividend aristocrats, income strategy

High Momentum

• momentum, trend following, jegadeesh titman, 12-1 momentum, price momentum, trend

Top N Large Cap

• large cap, mega cap, top n, fama french size, size factor, market cap weighted

You can see them all by asking Claude:

What strategy templates are available?

Claude calls list_strategy_templates and returns the catalog with each template's id, name, description, default cadence, factor count, and top_n. Useful for browsing.

What happens when no template matches

If you ask for something exotic — "Build me Piotroski F-score" or "Buffett-style quality" — lookup_strategy returns:

{
  "query": "piotroski",
  "confident_match": null,
  "alternates": [],
  "guidance": "No built-in template matches 'piotroski'. Compose a custom ranking block per the cookbook in your MCP system prompt. Restate the strategy intent back to the user before calling create_strategy."
}

Now Claude falls back to custom strategy authoring. It uses the cookbook of factor derivations baked into the MCP system prompt:

Claude composes a ranking block using these expressions, then restates its interpretation back to you before calling create_strategy. Something like:

I don't have a built-in Piotroski template. The closest custom version uses three quality components from our snapshot: ROA + operating margin + gross margin, equally weighted. We don't have year-over-year change data needed for the full 9-component F-score. Want me to proceed with the simplified version?

That restate-before-creating pattern is the safety check for the no-match case. The dry-run preview (showing actual tickers) is the second safety check.

What about fields we don't have at all?

Some strategies need data Seentio's snapshot doesn't include. The classic example: debt_to_equity. EODHD's API has the field name in the schema, but it's universally null in our daily snapshot — we'd need a different data feed to populate it reliably. We document these in our field inventory and lookup_strategy will surface a warning if you accidentally try to use one.

When this happens, Claude is instructed to propose a substitute and ask rather than silently failing. For debt_to_equity specifically, the substitute is operating_margin (operational efficiency) or 1/pb (leverage indirectly via book value).

Tips & common questions

• The lookup is your safety net. Even when you're confident a strategy has a clear name, asking Claude to look it up first costs nothing and protects against the name-collision problem. Make it a habit.

• Aliases are case-insensitive. "GREENBLATT", "greenblatt", "Greenblatt" — all match. Spaces, hyphens, and underscores are normalized: "magic_formula" and "magic formula" both work.

• The catalog grows over time. We ship templates conservatively — each one is a vetted academic strategy with documented behavior. If a strategy you'd want is missing, email us at contact@app-seentio.com. Common requests: Piotroski F-score, Quality (QMJ), Low Volatility, Momentum + Quality combos.

• Custom strategies still work. When the lookup fails, Claude composes a ranking block from the cookbook. You're not stuck — you just need to confirm the interpretation before saving.

• The summary echo-back is the second safety net. After create_strategy saves, the response includes a plain-English summary of EXACTLY what got saved. Claude surfaces this verbatim. Read it. If anything's off, ask Claude to update or delete-and-recreate.

• Lookup failures are logged. Every lookup_strategy call shows up in your S3 conversation log at wyn-associates/open_ledger/copilot_sessions/{user_id}/mcp_{date}.jsonl. We use these to grow the alias list — if "garp ratio" returns no match for 5 different users, we know to add it.

What's next

• Create your first strategy using Claude Code — try saying "Greenblatt" first to see lookup_strategy in action.
• Run a backtest on whatever strategy you create.
• Read how the alert system fires so you know what to expect post-activation.

Need more help? Visit our Chat page to ask Seentio's AI assistant, or reach out at contact@app-seentio.com.