Instructing Opulent
Be opinionated and specific. Opulent can investigate ambiguity, but it should not have to guess your preferred outcome, source, policy, or design.
Provide clear directives
State the job, the intended result, and any decisions already made. "Improve our reporting" leaves the source, format, audience, and success test open. This does not:
Reconcile the April revenue export against the general ledger, explain every mismatch, and return an editable workbook plus a one-page controller summary.
Pick work that fits the system
Choose tasks with a concrete result and a way to check it. Provide the files, systems, examples, templates, and links Opulent must follow. Visual work needs a reference design or exact layout requirements. Domain work needs the policy, the calculation rules, and an owner for exceptions.
Make context authoritative
Tell Opulent which source wins when inputs disagree. Name the canonical artifact or system of record, then identify the working copies, examples, and prior deliverables that may inform the job without overriding it. If a stale workspace copy conflicts with the named authority, require Opulent to stop and surface the conflict.
Connect the repositories, documents, business systems, and specialized tools the task needs. State whether a missing connection blocks the job or permits a lower-quality fallback, and require approval before a consequential workflow changes its execution path.
Establish frequent checks
Spell out validation. Use checks that fit the work:
| Kind of work | Checks to require |
|---|---|
| Data | Row counts, formulas, schemas, reconciliations, invariants |
| Research | Citations, source conflicts, coverage notes |
| Documents | Style checks, redlines, export review |
| Code | Tests, lint, type checks, builds, CI |
| Visual workflows | Screenshots and recordings |
| Operations | System receipts, logs, before-and-after state |
Set checkpoints and independent work packages
Break a complex job into steps with their own completion state. Run independent packages in parallel and keep dependent work in sequence.
Give every managed session the same return format for artifacts, evidence, status, and exceptions. The lead run can then compile the result without interpreting a different structure from each child.
Tell Opulent to test its own work
Ask Opulent to open the finished artifact, run the application, operate the browser flow, or query the system after writing. The agent should exercise the result through the same surface a person or system will use, before asking for review.
A read-after-write check catches the failure mode that hurts most: a write that reported success and did not land.
Use Skills and Knowledge for repeated guidance
A Skill contains the steps for a type of job. Knowledge holds facts, policies, conventions, and context used across runs. Put changes to the method in the Skill and changes to stable rules in Knowledge. Leave one-time exceptions in the run, and never use Memory to bypass an approved rule.