Audit an agent’s sources
Checks whether an answer is genuinely grounded in the sources it claims.
[ANSWER][SOURCES]
Topic · 64 prompts
Agents & automation
Explore agents & automation prompts with clear inputs, complete instructions and ways to adapt the result. Explore “Audit an agent’s sources” and “Design human approval gates”.
Design human approval gates
Prevents a helpful assistant from silently sending, buying or deleting.
[WORKFLOW]
Find why an agent failed
Diagnoses an agent run without turning guesses into fixes.
Generate an agent test suite
Creates normal, edge, adversarial and failure tests with observable pass criteria.
[CONTRACT]
Map an agent’s tool permissions
Sorts tools into read, draft, approval and prohibited capabilities.
[AGENT PURPOSE][TOOLS]
Universal agent instruction template
A fill-in-the-blanks operating contract for a narrow agent.
[AGENT NAME][OUTCOME][AUDIENCE][INPUTS]
Turn an idea into an agent build specification
Produces a platform-neutral design before anyone pays for API calls or grants access.
[USE CASE][AUDIENCE][DATA][OUTCOME]
Ask Codex to set up a project agent
Requests a scoped project workflow and tests before any automation is added.
[TASK][FOLDER][ACCEPTANCE TESTS]
Ask ChatGPT to set up your first agent
Turns one recurring task into a bounded, testable assistant without pretending tools are connected.
[TASK][RESULT][FILES OR EXAMPLES]
A neighbourhood improvement field agent
Plans a small, lawful community improvement with real stakeholder checks.
[PUBLIC SPACE OR SERVICE][AREA][TIME AND BUDGET]
Organise messy files with a dry-run manifest
Plans reversible file organisation without deleting or silently overwriting anything.
[EXACT ROOT]
Monitor a topic and report only meaningful changes
Defines a repeatable research monitor with source quality, deduplication and quiet unchanged runs.
[TOPIC][ALLOWED PRIMARY SOURCES][SCHEDULE][ACTIONABLE THRESHOLD]
Test a site as several real user types
Walks key journeys with screenshots, failures and reproducible evidence rather than a vague quality score.
[SITE][PERSONAS][STAGING OR PRODUCTION][TEST ACCOUNTS AND DATA]
Create a bounded inbox-triage agent
Sorts incoming messages, drafts replies and escalates important cases without sending on its own.
[TEAM][ALLOWED INBOX AND TIME RANGE]
Safe website operations agent
Creates a persistent operator that diagnoses, changes, verifies and reports without overclaiming.
[SYSTEM][OUTCOME][TESTS][ENVIRONMENT]
Bounded research agent
Defines a research agent that cites evidence, stops at clear limits and exposes uncertainty.
[OBJECTIVE][ALLOWED SOURCES][BUDGET]