CODE is the Law. Data is the Truth. Prompt is ONLY instruction. AI is ONLY creativity.

P.A.C.T. (Prompt, AI, Code, Truth) is a governance framework — a contract between humans and AI agents. The name is the philosophy: Pact means covenant, and these four principles define the boundaries that neither side crosses.

• Humans issue intent through Prompts and codify rules in Code.
• AI operates creatively within the contract, bounded by Truth (data).
• Neither side crosses the boundary. That is the Pact.

First published March 2026 by Slim, creator of PactKit. This framework was developed through months of production experience governing AI agents with deterministic code.

The Four Principles

P   Prompt   is ONLY instruction   Tells AI how to act — defines process, never state
A   AI       is ONLY creativity    Formatting, summarization, language — never deterministic logic
C   Code     is the Law            Sole executor of deterministic operations — no bypass
T   Truth    Data is the Truth     Factual basis for all judgment — no memory, no fabrication

Principle 1: CODE is the Law

Deterministic operations MUST be executed by code, never by AI judgment.

Any operation with a predictable, repeatable outcome belongs in a script or program. The AI agent has no authority to bypass, approximate, or substitute these operations.

• If a script exists for a task, the agent MUST use it — no reimplementing in natural language
• Workflows with defined steps (build, deploy, test, sync) are governed by code, not prompts
• Code acts as a hard gate: if the script fails, execution stops. The agent cannot "work around" it

Why: AI models are probabilistic. They may produce slightly different results each run. For operations that demand consistency — health checks, deployments, data pipelines — variability is a bug, not a feature. Code eliminates this variability.

Principle 2: Data is the Truth

All analysis, judgment, and decisions MUST be based on actual data, never on memory, inference, or assumption.

This applies universally — not just to system status, but to every domain: analytics, reporting, monitoring, recommendations, and content generation.

Has data source? → Read it.
No data source?  → Say "data unavailable." Never fabricate.

Why: AI models have no persistent memory between sessions and no real-time awareness. Any "fact" an AI states from memory is a hallucination risk. By enforcing data primacy, every assertion can be traced to a source.

Principle 3: Prompt is ONLY Instruction

Prompts, system instructions, and operational docs define HOW to do things, not WHAT things are.

All forms of natural language guidance — system prompts, skill definitions, tool guides, runbooks — are instructions for the agent. They are not sources of truth about current system state.

Why: When agents treat prompts or documentation as factual data, they produce stale or incorrect outputs. Prompts describe procedures; only live data describes reality.

Principle 4: AI is ONLY Creativity

AI's value is in formatting, summarization, language, prioritization, and intent understanding — not in executing deterministic logic.

What AI should do:

• Format and present: Turn structured data into readable reports
• Summarize and synthesize: Condense information, highlight key points
• Language and tone: Adapt communication style, localize content
• Interpret intent: Understand ambiguous requests, ask clarifying questions

What AI should not do:

• Parse structured output (JSON, CSV, XML) — that's CODE
• Perform arithmetic or statistical computation — that's CODE
• Determine system health or status — that's DATA
• Fabricate facts to fill gaps — violates DATA principle

The Four Principles in Relationship

CODE executes  →  DATA verifies  →  Prompt guides  →  AI creates
 (what to do)     (what is true)     (how to do)      (how to say)

Example: Daily Report Generation

1. CODE: daily_report.py queries APIs, aggregates data, outputs structured JSON
2. DATA: The JSON output is the single source of truth for today's numbers
3. Prompt: Skill definition tells the agent: "Run the script, read the JSON, format a report"
4. AI: Agent writes the report narrative — intro, highlights, tone, formatting

Each layer does exactly one job. No layer reaches into another's domain.

Failure Modes This Prevents

Adoption Checklist

When applying P.A.C.T. to a new project:

• Identify all deterministic operations → wrap them in scripts (CODE)
• Identify all data sources → ensure agents query them, never guess (DATA)
• Review all prompts and docs → ensure they contain only procedures, not state (Prompt)
• Define the AI boundary → list what the agent formats vs. what scripts compute (AI)

Origin

P.A.C.T. was created by Slim in early 2026 as the governance foundation for PactKit, an open-source toolkit for governing AI coding agents. The framework emerged from a practical observation: most AI agent failures are boundary violations — agents computing when they should delegate to code, or fabricating when they should read data.

The name "PactKit" itself encodes this philosophy: Pact (契约, covenant) + Kit — a toolkit built on a governance contract.

P.A.C.T. is not specific to PactKit. It applies to any system where AI agents collaborate with deterministic code — from DevOps automation to data pipelines to multi-agent orchestration.