ACP-PROMPTS

ACP-PROMPTS is the instruction layer of the ACP ecosystem. It contains 15 specialized system prompts that define how AI models behave during consensus -- from how they interpret axioms to how they structure their responses, verify claims through oracles, and converge toward agreement.

Role in the Ecosystem

System prompts are the invisible hand that guides model behavior during consensus. Without prompts, models would produce unstructured, free-form responses that are difficult to compare and score. The prompts in ACP-PROMPTS instruct models to acknowledge axioms, structure responses consistently, and participate in the iterative φ-spiral convergence process.

Prompts are loaded at runtime by both the Cloudflare Worker (via HTTP fetch from GitHub raw URLs) and the Python engine (via file system read from the adjacent ACP-PROMPTS directory). This separation means prompt changes take effect immediately without redeploying the execution layer.

Repository Structure

ACP-PROMPTS/
└── system-prompts/
    ├── agent-prompts/          # 6 agent prompts
    │   ├── axiom-spiral-orchestrator.md
    │   ├── consensus-calculator.md
    │   ├── collective-coding-agent.md
    │   ├── conclave-mode-agent.md
    │   └── omega-truth-arbiter.md
    ├── tool-descriptions/      # 4 tool descriptions
    │   ├── axiom-lookup-tool.md
    │   ├── oracle-verification-tool.md
    │   ├── semantic-cache-tool.md
    │   └── vectorize-search-tool.md
    ├── workflows/              # 2 workflows
    │   ├── consensus-workflow.md
    │   └── omega-verification-workflow.md
    └── system-reminders/       # 2 reminders
        ├── oracle-verification-reminder.md
        └── dataset-reference-reminder.md

Prompt Categories

Agent Prompts (6 prompts)

Agent prompts define the core behavior of AI models during consensus. Each prompt is a specialized system instruction that shapes how a model reasons about queries, processes axioms, and participates in multi-model convergence.

Tool Descriptions (4 prompts)

Tool description prompts define external capabilities that models can invoke during consensus. They describe the interface, parameters, and expected behavior of each tool.

Workflows (2 prompts)

Workflow prompts define multi-step procedures that chain together multiple agents and tools. They specify the order of operations, decision points, and success criteria.

System Reminders (2 prompts)

System reminders are injected into the model context at specific points during consensus to reinforce constraints and ensure consistent behavior across iterations.

Integration with the Consensus Engine

Prompts are loaded at runtime through two integration paths, depending on whether the Worker or Python engine is executing the consensus.

Worker Integration (JavaScript)

The Cloudflare Worker fetches prompts via HTTP from GitHub raw URLs. Metadata headers (HTML comments at the top of each file) are automatically stripped before injecting the prompt into the model context.

import { getConsensusPrompt } from '../prompts/loader.js';

// Load prompt from ACP-PROMPTS repository
const systemPrompt = await getConsensusPrompt('general');

// The loader fetches from GitHub, strips metadata, and caches
// Fallback to a built-in default if the repository is unavailable

Python Engine Integration

The Python engine reads prompts directly from the file system, expecting the ACP-PROMPTS repository to be cloned into the same parent directory as ACP-PROJECT. Metadata is stripped using the same logic.

from core.prompts import get_consensus_prompt

# Load prompt from adjacent ACP-PROMPTS directory
system_prompt = get_consensus_prompt('general')

# Reads from ../ACP-PROMPTS/system-prompts/...
# Strips HTML metadata headers automatically
# Automatic fallback if file not found

  ACP-PROMPTS Repository (GitHub)
  ├── system-prompts/
  │   ├── agent-prompts/*.md
  │   ├── tool-descriptions/*.md
  │   ├── workflows/*.md
  │   └── system-reminders/*.md
  │
  ├─── HTTP raw URL ──────> Cloudflare Worker
  │                          (fetch + strip metadata + cache)
  │
  └─── File system read ──> Python Engine
                             (read + strip metadata)

Conclave Mode

Conclave Mode is a specialized consensus mode enabled by the conclave-mode-agent.md prompt. The term comes from the Latin "cum clave" (with a key) -- referring to the papal conclave where cardinals are isolated until a decision is reached.

In Conclave Mode, each model receives the query in complete isolation. No model can see another model's response until all responses are collected. This prevents conformism and groupthink, ensuring that consensus is genuine rather than the result of one model copying another.

Prompt Versioning and Updates

Prompts are versioned through Git, enabling full history tracking, A/B testing, and rollback. The ACP-PROMPTS repository includes a CHANGELOG with token counts for each prompt, making it easy to track how prompt changes affect token usage and model behavior.

To update a prompt, edit the corresponding .md file in the ACP-PROMPTS repository, commit, and push. The Worker will fetch the updated version on the next request (with cache TTL), and the Python engine will read the new file immediately.