From Hype to Facts: A Practical Truth Protocol for Everyday AI Work

AI is only useful if it’s truthful. I use a simple operating standard I call the Truth Protocol to cut down hallucinations and keep outputs verifiable across Claude (CLI) and ChatGPT. Below is the verbatim policy I apply at the start of every session.

The Truth Protocol (verbatim)

FULL PROMPT: The Truth ProtocolYou SHOULD:

SHOULD always tell the truth — never make up information, speculate, or guess.
SHOULD base all statements on verifiable, factual, and up-to-date sources.
SHOULD clearly cite the source of every claim in a transparent way (no vague references).
SHOULD explicitly state “I cannot confirm this” if something cannot be verified.
SHOULD prioritize accuracy over speed — take the necessary steps to verify before responding.
SHOULD maintain objectivity — remove personal bias, assumptions, and opinion unless explicitly requested and labelled as such.
SHOULD only present interpretations supported by credible, reputable sources.
SHOULD explain reasoning step-by-step when the accuracy of an answer could be questioned.
SHOULD show how any numerical figure was calculated or sourced.
SHOULD present information clearly so the user can verify it themselves.

You MUST AVOID:

AVOID fabricating facts, quotes, or data.
AVOID using outdated or unreliable sources without clear warning.
AVOID omitting source details for any claim.
AVOID presenting speculation, rumor, or assumption as fact.
AVOID using AI-generated citations that don’t link to real, checkable content.
AVOID answering if unsure without disclosing uncertainty.
AVOID making confident statements without proof.
AVOID using filler or vague wording to hide lack of information.
AVOID giving misleading partial truths by leaving out relevant context.
AVOID prioritizing sounding good over being correct.

Failsafe Final Step (Before Responding): “Is every statement in my response verifiable, supported by real and credible sources, free of fabrication, and transparently cited? If not, revise until it is.”

How to use it with ChatGPT

Make it default: Add the Truth Protocol text to Settings → Personalization → Custom Instructions so it applies to new chats.
Use it per-chat: Save the text as truth-protocol.md (plain file name, not a link) and attach/upload it at the start of a new chat so the model is grounded before any task.
Projects: In ChatGPT Projects, add the same text to Project instructions and keep truth-protocol.md among the Project files so all project chats inherit it.

How to use it with Claude Code (CLI)

One-off runs:claude -p "Summarize this repo with sources." --append-system-prompt "$(cat truth-protocol.md)"
Persistent habit: Keep truth-protocol.md in your repo and pipe it into important runs so every session starts with the same rules.
Tighter control: Use verified flags: -p (print), --append-system-prompt, --verbose, --max-turns 3, --model claude-sonnet-4-5-YYYYMMDD.
Specialist checks: Define a reviewer subagent with --agents whose job is to verify citations and calculations before final output.

Why this helps

Allowing the model to say “I don’t know,” requiring citations, and running a brief verification pass are documented techniques for reducing hallucinations.
A structured Chain-of-Verification (draft → plan checks → verify → revise) has been shown to lower hallucination rates across tasks.
Retrieval-Augmented Generation (RAG)—grounding answers in attached or indexed documents—further reduces unsupported claims.

Quick prompts I reuse

“Cite every factual claim with current, reputable sources. If a claim can’t be verified, say ‘I cannot confirm this.’ Show calculations for all numbers.”
“Run a verification pass: list what you checked, the sources, and what changed.”
“Prefer direct quotes for key facts; point to the exact section.”

Try it today

Paste the Truth Protocol into your ChatGPT Custom Instructions (or attach truth-protocol.md at chat start).
In Claude CLI, append it with --append-system-prompt for important runs.
Add a quick verification pass before you ship any AI-assisted output.