Guides
Each guide follows one real task from question to checked answer. It shows what to ask, which tools the LLM calls and what they return, the kind of answer to expect, and how to check that answer in PeopleSoft before you rely on it.
The guides use a small custom application as their example: a training request system whose
definitions start with ZZ_, including the ZZ_TRAINING_REQ component, the ZZ_TRAIN_REQ record,
and the ZZ_TRN_LOAD Application Engine program. The SQL guide uses the delivered JOB and
DEPT_TBL records.
Get better answers
Section titled “Get better answers”These habits help in every guide:
- Name what you know. An exact name, such as “the
ZZ_TRAIN_REQrecord”, lets the LLM go straight to the right tool. If you only know what users see, describe it, such as “the Job Data page”, andsearch_definitionsfinds the definition behind it. - Say the market when a component isn’t in
GBL. - Ask for sources. Ask the LLM to list the definitions it looked at, with their paths. That makes the answer easy to check.
- Check before you change anything. AI Assist reads definitions; it doesn’t change them. Before you act on an answer, open the definitions it names in Application Designer.
- Remember which database you’re asking about. The server reads one database, usually development, and production can differ from it.
Tell the LLM about your system
Section titled “Tell the LLM about your system”Most AI IDEs read a project instructions file on every request: CLAUDE.md for Claude Code,
.github/copilot-instructions.md for GitHub Copilot in VS Code, and project rules in Cursor. A few
lines about your PeopleSoft system save you from repeating them in every question:
## PeopleSoft
- Use the aipeoplesoft tools to look up PeopleSoft definitions. Don't guess record, field, or component names.- The tools read HRDEV: HCM 9.2 on PeopleTools 8.62, Oracle Database.- Our custom definitions start with ZZ_. Delivered definitions are last changed by PPLSOFT.- Components use market GBL unless a task says otherwise.- In SQL, correlate on every key that describe_record returns, and handle EFFDT and EFFSEQ on effective-dated records.