OracleAGI: ADVANCED Prompt Engineering System (ChatGPT-4)

TL;DR

OracleAGI uses 25+ chained prompts and multiple loop cycles to iteratively generate and refine unusual business ideas.

Briefing Cornell Notes

Briefing

OracleAGI is a multi-agent, multi-step prompt engineering system designed to generate and iteratively refine “out-of-the-box” business ideas by running dozens of chained prompts in loops. The core mechanism is a pipeline where eight distinct expert agent personas—covering business strategy, marketing, and software-related perspectives—evaluate a single starting “ID,” produce multiple new ideas, and then feed improved outputs back into the same structure to push creativity further.

The system begins with 25+ chain prompts implemented in a custom setup (shown via Visual Studio Code). Each agent persona has a tailored role description; for example, one agent is positioned as an expert in business finance and strategy with an MBA background, while others are framed as marketing specialists or software-focused thinkers. The starting point is a single idea ID (the transcript uses a cloud service concept for booking a sunny day for an outdoor event as an example). That ID is injected into every agent’s prompt so all personas work from the same seed.

From there, OracleAGI runs in loops. In the first loop, two business-oriented agents generate three “crazy” business ideas, which are saved to a file named new id.txt. The transcript gives examples such as a holographic memory store, a telepathic social network, and time travel tourism. A subsequent loop takes those improved ideas as new inputs and prompts the agents to find correlations across the set, then brainstorm five even crazier ideas together. Those five outputs are saved to another file (described as new idea 2), and the process continues.

A later loop reduces the candidate set by selecting the best performers: from a larger pool (the transcript describes 20 IDs at one stage), the agents pick three top ideas and save them to a new file. The final step then asks for a ranked, well-structured list of the five most “crazy” and best ideas, each presented with a title and a TL;DR. Example outputs include an Earth-to-space elevator, an instant food replicator, future map brain implants, a dream exploit exploration service, and a biofeedback theme park.

The practical pitch is that the system can be run on ChatGPT-4 (with an estimated ~40-minute runtime and ~$6–$7 cost), though the creator recommends “chapter turbo” for testing. The workflow is presented as straightforward to set up despite the large number of prompts, and the creator signals that OracleAGI has not been fully explored yet—suggesting further experimentation with its reasoning and iterative properties. The overall takeaway is that creativity is treated like an engineering problem: seed an idea, run persona-based critique and generation in loops, persist intermediate results to files, and finish with ranking and structured summarization.

Cornell Notes

OracleAGI is a chained, multi-agent prompt system that generates and refines highly unusual business ideas through repeated loop cycles. It uses eight expert agent personas (e.g., business finance/strategy, marketing, and software) that all start from the same seed “ID,” then produce new idea IDs saved to text files. Later loops compare and correlate multiple candidate ideas, expand the idea set (e.g., generating five new ideas), then narrow it down by selecting top candidates (e.g., choosing three). The process ends by ranking and formatting the five best ideas with titles and TL;DR summaries. The value is a structured way to push ideation beyond a single-pass brainstorm by forcing iterative evaluation and synthesis across personas.

How does OracleAGI turn one starting idea into a larger set of new concepts?

It injects a single seed “ID” (a starting business idea) into the system prompt of every agent persona. In the first loop, the business persona agents generate three new idea IDs, which are saved to a file (named new id.txt). Those improved IDs then become inputs for the next loop, where agents are prompted to find correlations across the set and brainstorm five additional ideas together. Each loop writes outputs to new text files so later steps can reuse the expanded candidate pool.

What role do the “expert agent personas” play in the ideation process?

Personas constrain how each agent contributes. The transcript describes eight different expert personas spanning business finance/strategy, marketing niches, and software-related roles. Because each persona has a role-specific prompt, the system produces ideas that reflect different angles—strategy, go-to-market, and technical feasibility—rather than a single generic brainstorm. In later loops, those diverse outputs are compared and synthesized to generate more extreme follow-on ideas.

Why do the loops include both expansion and selection steps?

Expansion increases creative coverage, while selection prevents the idea set from becoming unmanageable. The transcript shows an expansion phase where three initial ideas lead to five new ones after correlation-based brainstorming. It then describes a selection phase where a larger pool (e.g., 20 IDs) is narrowed to a top three, saved to a new file. This alternating pattern helps the system keep variety while still converging toward the “best ranked” outputs.

What does the system produce at the end, and how is it formatted?

The final prompt asks for a well-structured list of the five most crazy, best-ranked out-of-the-box ideas. When run, the output is described as five ideas with a “crazy scale,” a title, and a TL;DR for each concept. Examples mentioned include an Earth-to-space elevator, an instant food replicator, future map brain implants, and a biofeedback theme park.

What practical constraints are mentioned for running OracleAGI?

The transcript suggests running on “chapter turbo” for testing, while ChatGPT-4 is possible but slower and more expensive. An estimate is given: about 40 minutes runtime and roughly $6–$7 cost due to 25+ chain prompts. The creator also notes the system hasn’t been fully explored yet, implying performance and behavior may vary with further tuning.

Review Questions

If you wanted to change the creative direction of OracleAGI, which parts would you modify: the seed ID, the persona prompts, or the loop prompts—and why?
How does saving intermediate outputs to text files affect the system’s ability to correlate ideas across loops?
What trade-off does the system manage by alternating between generating more ideas (e.g., 5) and selecting fewer (e.g., 3)?

Key Points

1
OracleAGI uses 25+ chained prompts and multiple loop cycles to iteratively generate and refine unusual business ideas.
2
Eight expert agent personas (business, marketing, and software-focused) contribute ideas from the same seed “ID.”
3
Each loop writes outputs to text files, enabling later loops to reuse and correlate improved idea sets.
4
The workflow alternates between expansion (e.g., generating five new ideas from three) and selection (narrowing a larger pool to top candidates).
5
The final step ranks and formats the five best ideas with titles and TL;DR summaries.
6
Running on ChatGPT-4 is described as slower and costlier (~40 minutes, ~$6–$7) due to the number of chained prompts.
7
The system is presented as simple to set up despite the large prompt library, and further experimentation is encouraged.

Highlights

OracleAGI treats creativity as an iterative pipeline: seed one idea, generate multiple candidates, correlate across candidates, then narrow to top picks.

Persona diversity is central—business strategy, marketing niches, and software perspectives all shape the idea outputs.

Intermediate results persist to files, letting later loops synthesize across earlier generations rather than starting over each time.

The end product is a ranked, structured list of five “crazy” ideas with TL;DR summaries.

Topics

Multi-Agent Prompting
Iterative Ideation
Prompt Engineering
Business Idea Generation
ChatGPT Workflows