Skip to main content
Structured Finance includes built-in skills: reusable capabilities that combine a set of prompts and domain logic into a single task the agent can carry out on request. For example, the adhoc-strats skill turns a plain-language question like “show me top 10 vehicle models by balance” into a full stratification query, without you having to build the aggregation yourself. You can ask the Agent what’s available:
and it will return a list of built-in skills. You can also add your own skills or have the Agent create one for you.

​
Built-in skills

These skills work across all supported Auto ABS deal types (GMCAR, SDART, BMWOT, TAOT, CALT, EART, and so on) and integrate with the harmonization and strats pipelines you’re already using. Skills are customizable to your organization’s needs. When you confirm use of a skill, the Agent identifies the necessary steps and gems, executes them, and returns traceable results — for example, surfacing state or vintage-quarter concentrations to support a pool-cutting decision.

​
Get the full cheat sheet for a skill

You can ask the Agent for all cheat sheet prompts:
The Agent then returns a list of prompts: result of adhoc Strats cheat sheet then run specific ones by reference:
You can then run a prompt directly by number — for example, run 2.2. Running 2.2 Show me top 10 vehicle models by balance with avg LTV, for example yields a result similar to the following: result of running 2.2 Show me top 10 vehicle models by balance with avg LTV

​
Ad-hoc stratification (adhoc-strats)

The adhoc-strats skill accepts natural-language prompts for exploring a loan pool. Prompts are organized around common analysis patterns:
  • Pool Health — summary stats (contracts, balance, WA FICO/LTV/APR), credit tier splits, new/used breakdowns
  • Concentration Risk — top states, vehicle models, and vintages by balance; concentration threshold flags
  • Tail Risk — filtering to subprime FICO, high LTV, or combined tail segments
  • Stacked Risk — multi-factor filters (e.g., FICO under a threshold AND LTV over a threshold)
  • Pricing Adequacy — APR by credit tier, mispriced-risk detection, new vs. used APR spread
  • Delinquency & Performance — DPD bucket breakdowns, delinquency by state or credit tier
Prompts support threshold tuning (e.g., “under 660” vs. “under 620”), percentage-of-pool framing, and dimension pivots (e.g., “by state” or “by vintage”). Thresholds for what counts as subprime, high LTV, or high APR vary by deal type — prime loan vs. subprime loan, loan vs. lease.

​
Cross-deal comparison (asset-comp)

The asset-comp skill runs on any of these trigger phrases:
  • asset comp / add to comp / compare to other deals
  • compare this tape / extract comp / insights on this deal
  • write research note / collateral analysis / cross-deal comparison
  • GMCAR comp / download prospectus / refresh comp

​
Operations

​
Input modes

  1. Harmonized CDM tape — uses your current pipeline’s output_preview data
  2. SEC prospectus — downloads and parses prospectus PDFs (e.g., “download prospectus for GMCAR 2025-2”)

​
Example: extracting a comp

“extract comp from this tape” returns a preview — no changes are made to the comp table until you confirm:

​
SDART 2025-4 — Comp Extraction (Preview)

Next steps from a preview:
  • “add to comp” → insert into cross_deal_comp.csv
  • “compare to others” → side-by-side with peer deals
  • “insights on this deal” → full cross-deal analysis
Cross-deal comparison works because every tape is harmonized to the same CDM, which means that metrics line up without additional mapping. The Agent also self-retries on errors encountered during execution.

​
Add a new skill

Built-in skills cover common workflows — harmonization, ad-hoc strats, cross-deal comparison — but you can add a skill scoped to your own team’s practices. For example, a desk might codify its own tail-risk definitions as a skill, since what counts as “subprime” or “high LTV” varies by deal type. A skill such as desk-risk-thresholds could encode your team’s specific cutoffs and stress-test segments as a reusable prompt set, so anyone on the desk gets consistent results without re-specifying thresholds each time. To add a new skill:
  1. Click the +, then select New Tab, then select Skill.
  2. In the dialog, enter a Skill Name and click Create.
  3. On the new page, enter a description and content for the skill.
You can also ask the Agent to create a skill for you.
For more information on skills, see Add and use skills