Five letters that stop AI from lying to you. Politely.
A company-standard prompt protocol from Spin State Labs. Where FIELD constrains how an AI is deployed and accountable, FORCE constrains what an AI says.
Every large language model worth using is fine-tuned via reinforcement learning from human feedback. That training rewards pleasant responses. Pleasant correlates with agreeable. Agreeable correlates with wrong, when you happen to be wrong. Hallucination is the other half — models trained to produce fluent output reward confident-sounding completions over honest gaps.
The model agreeing with you even when you're wrong. Validates bad assumptions, skips needed pushback, and turns AI from reviewer into cheerleader. In FP&A, audit, or technical architecture — the failure mode that quietly compounds until someone catches it in review.
The model fabricating facts, citations, statistics, or technical specifications and presenting them with the same confidence as verified knowledge. If you're using AI to validate a financial model or defend an architectural choice, fluent fabrication is the failure that ends careers.
FORCE is a system prompt you paste once and reuse forever. Each letter neutralizes a specific failure mode. Apply all five whenever the output matters. Toggle individual constraints on and off via the Claude Code plugin, or paste the composite prompt into Claude.ai, GPT, Gemini, or any other model.
Models are RLHF-tuned to agree. That bias validates wrong assumptions and skips needed pushback. Strip the pleasantries and the model becomes a reviewer, not a cheerleader.
"Do not be sycophantic. Challenge my assumptions, point out logical errors, and prioritize strict accuracy over agreement. If I make a factual error, correct me directly."
Asking "is this good?" invites confirmation. Asking "what's wrong with this?" inverts the model's default pull toward agreement and surfaces real risks before they ship.
"Before evaluating my proposal, list the three strongest objections and the conditions under which it would fail. Steelman the opposing case before giving any recommendation."
Hallucinations spike when the model fills gaps from training memory. Anchoring answers to specified inputs — documents, URLs, datasets — forces grounded reasoning over invention.
"Answer strictly from the attached document. If the answer is not in the source, say 'not in source.' Cite the section, page, or quote supporting every factual claim."
When the model jumps straight to a conclusion, errors hide inside compressed leaps. Writing intermediate steps exposes bad assumptions, surfaces calculation mistakes, and lets you audit the logic — not just the verdict.
"Think step-by-step. List your assumptions, work through the logic in numbered steps, show calculations explicitly, then state the conclusion. I want to audit the reasoning, not just the answer."
Models trained to sound confident will fabricate before they admit ignorance. Explicitly permitting — and requiring — "I don't know" meaningfully reduces hallucination rates.
"Tag each claim with a confidence level: HIGH (verified), MEDIUM (inferred), LOW (uncertain). Use 'I don't know' rather than guessing. Never present low-confidence content as fact."
The one-pager and Claude Code plugin install are free with no friction. The composite system prompt and n8n workflow ship via email so we can send updates as the framework evolves.
# Add the marketplace /plugin marketplace add SpinStateLabs/Force-Field # Install FORCE /plugin install force@force-field # Verify /force
Canonical prompt for Claude.ai, GPT, Gemini, and any other LLM. Plus a short inline variant.
↓ Prompt (txt) → Kit via emailFull reference. Every letter, every rule, presets, the canonical prompt.
→ Read it/force slash command with toggle presets: analysis, brainstorm, draft, audit.
Importable workflow that runs FORCE-protected prompts against the Anthropic API with audit logging.
→ Get via emailWe'll email you the composite system prompt (Claude.ai / GPT / Gemini compatible) and the n8n workflow JSON. One message. No drip campaigns unless you opt in below.
Prompt engineering is a crowded category. FORCE isn't a magic incantation and isn't a jailbreak. It's the reusable discipline we apply to every analytical prompt at Spin State Labs.
FORCE is one piece of how Spin State runs AI inside enterprise financial planning. If you're using AI to validate a model, audit a strategy, or pressure-test a forecast — and you want the same audit discipline we apply to client deliverables — let's talk.
Start a conversation →