Fields
The curated field list: which fields an agent sees, at which level, in what order.
Open the picker
In the entity builder, click the Core Record card.
Each row is a field with five checkboxes — List, Standard, Detailed, Filter, Write — and a per-field instruction. The picker counts your choices against each level’s cap (10 / 20 / 30) as you go. → Levels
The level checkboxes decide what an agent receives; Filter and Write decide what it may ask and change:
- Filter — Find Records accepts this field in filters and
sortBy. Untouched, it follows the Detailed checkbox; check or clear it and ContextWorks records an explicit choice, so filtering and returning move independently. A filterable field in no level appears below the table as a filter-only pill: agents filter on it, no payload carries it. → Query filters - Write — agents may change this field through Update Records. An admin sets this, or a manager of the entity’s home domain. No agent can change the field until the entity and org write switches allow writes too. → Write access
Suggest Filterable mines your org’s recently-run reports for the fields people actually filter on — matched fields get declared filterable, the rest arrive as filter-only pills. Edit the result; it’s a starting point.
Add a field
Click Add Field and search the object’s fields by label or API name. The list shows only fields you can access; if one is missing, check your own field-level security.
To add a field from a parent record, walk the lookup with the picker breadcrumb — Account.Owner.Name on an Opportunity entity. Use this for one or two high-value parent fields; for several, use a reference instead.
Order the fields
Fields render in the picker’s order, in both markdown and JSON, and order carries weight with an LLM. Put the fields that identify and qualify the record first, supporting detail after.
Drag a row by its handle to reorder, or focus the handle and use ↑ / ↓. All three levels share one order, so reordering changes all three at once.
The rendered field block has no section headers — each labeled value speaks for itself. Summary Metrics and related lists keep their own labels.
Meaning: label, instruction, synonyms
The Meaning column opens one editor per field, holding three slots: Label (what the agent calls the field), Instruction (what the value means here), and Synonyms (what people call it). A filled glyph means the field already carries something.
Show Meaning swaps the level, filter, write and create columns for an editable box per row, so you can write meanings for many fields in one pass. → Field meaning
Field types
| Type | How it renders |
|---|---|
| Text, picklist | As-is |
| Number, currency, percent | Formatted with separators and symbols; currency respects the org’s settings |
| Date, datetime | Formatted, not raw |
| Checkbox | true / false |
| Lookup | The referenced record’s name where available, otherwise the ID |
| Formula | Its calculated value |
| Address | Postal form on one line |
| Geolocation | latitude, longitude |
| Long text | In full — check the token cost |
Compound fields (BillingAddress, a geolocation) are cheaper than their components: one curated field, one readable string. Use the components when you want only part — country alone, say, for territory assignment.
Empty values
ContextWorks leaves out a field that has no value on the record — no dash, no null. It works record by record, with nothing to configure. Quality measures completeness per field once you have traffic.
Fields added automatically
Two cases need no curation:
- ContextWorks queries any field a formula references so the expression can evaluate. It renders only if you also curate it.
- ContextWorks queries the columns a related list or reference needs on those records, and their entity’s curation governs what renders.
Best practices
- Include what changes an answer. If this field were different, would the agent say something different?
Renewal_Date__c— yes.LastModifiedDate— almost never. - Curate
listfirst and hardest. They are the columns this record shows everywhere else; choose them so a human could identify the record. - Leave out mostly-empty fields. A field populated on 4% of records costs tokens on every load and informs almost none.
- Watch for redundancy.
Amount,Expected_Revenue__c, andWeighted_Amount__ctogether invite the wrong pick. Include the authoritative one and explain it with an instruction. - Include IDs when the agent may act. A workflow that chains into an update needs the record ID in the payload.
Limitations
Removing a field hides it from this view only; a user who can query the object can still reach it elsewhere. Use field-level security for access control. The other direction is safe: curate a sensitive field and callers without access to it still won’t receive it. → Security
What’s next
- Instructions — attach meaning to what you curated.
- Write access — what agents may change, and the chain that gates it.