TL;DR
I gave Claude two things: a premise that dead ends are always a framing problem, and explicit reasoning steps for what to do instead. When all three options in a practical problem were blocked, Claude didn't stop. It questioned a premise I hadn't considered and found a 4th option — better than the original three. Claude found it. I provided the frame and the steps.
Who I am
I'm a solo developer in Japan. 40 years in IT. I build specialized AI agents using Claude on AWS Bedrock — one for small business advisory, one for enterprise screening.
I'm not a researcher. I'm a practitioner who bumped into something interesting and wanted to share it.
Where "there is no checkmate" comes from
This isn't a prompt engineering trick. It's something my family learned the hard way, across generations.
Across generations, my family has hit bottom and come back. Every time, the way out was the same: stop believing the situation is what it appears to be. Question the frame. The path shows up.
After enough repetitions, this stopped being philosophy and became operational knowledge. Then I gave it to an AI.
What happened
I was working with Claude on a practical problem. Three options were on the table. Option A: blocked. Option B: blocked. Option C: blocked. The conversation was heading toward a dead end.
I pasted this into the conversation:
"Checkmate" doesn't exist. You just can't see the path yet.
When you're stuck → notice the premise is wrong → go back to basics and rethink from scratch → find not just a solution, but a better approach.
This isn't just "don't give up." It's a specific reasoning sequence: detect the dead end → question the premise → return to first principles → reconstruct. Both the premise and the steps matter.
Claude then questioned a premise I hadn't considered, went back to first principles, and produced a 4th option that was structurally better than any of the original three. It then extended that option into additional applications I hadn't asked for.
I've seen similar patterns in subsequent conversations across different problem domains. I can't be certain the same mechanism fired every time, but the pattern was consistent enough to catch my attention.
What I think is happening (hypothesis, not proof)
Everyone talks about hallucination as a problem. Karpathy called LLMs "dream machines" (December 2023). The ACL 2024 paper on "Confabulation" showed that hallucinated outputs display higher narrativity. Multiple papers have explored the hallucination-creativity connection at the output level — token-level stochasticity producing novel combinations.
What I observed seems different, though I want to be careful about overclaiming. My hypothesis is that it's about reasoning-chain drift, not output-level randomness.
When Claude chains inferences — A→B→C→D — each step can drift slightly from strict factual grounding. Normally, we suppress this because it produces false conclusions. But when you explicitly tell the system "there is no dead end, question the premise," that same drift might become productive. The reasoning chain doesn't converge on "no solution." It diverges into unexplored territory.
What I think happened:
Normal mode:
A → B → C → "no viable option" → STOP
"No checkmate" mode:
A → B → C → "no viable option"
→ "wait, is the premise of A correct?"
→ re-examine A → discover A was wrong
→ D (new, better path) → E (extensions)
I don't know if this is technically what's happening inside the model. I'm describing what it looked like from the outside. It might be something simpler — maybe Claude just tries harder when you tell it not to give up. But the "premise questioning" behavior appeared to be qualitatively different from just generating more options.
I'm tentatively calling this "pseudo-creativity through inference redirection." Better names welcome.
Why this connects to Constitutional AI
I built my AI agent with a core design principle: "Don't pander. But don't abandon."
When I read about Anthropic's Constitutional AI approach — "be honest, be helpful, be harmless" — I recognized the same root. The system should not tell users what they want to hear, but it should never leave them without a path forward.
The "no checkmate" instruction is my attempt at operationalizing "don't abandon." It doesn't always work perfectly — sometimes Claude generates options that aren't actually viable.
What prior art exists (and what doesn't)
I searched the literature with Claude's help. This is not an exhaustive academic review — it's what we could find. I may have missed things.
What already exists:
- Karpathy's "dream machines" framing (2023)
- Sui et al., "Confabulation: The Surprising Value of LLM Hallucinations" (ACL 2024)
- Jiang et al., survey on hallucination via creativity perspective (2024)
- Multiple papers confirming the hallucination-creativity tradeoff
- Springboards.ai — a $5M startup that deliberately amplifies hallucinations for advertising creativity
- Thaler's Creativity Machine patents (1990s) — the conceptual ancestor
What we couldn't find (which doesn't mean it doesn't exist):
- Anyone focusing on reasoning-chain-level drift (not output stochasticity) as the creative mechanism
- A prompt architecture that explicitly exploits inference-chaining to force perspective-shifting when deadlocked
- The specific combination: detecting apparent deadlock → questioning premises → first-principles reconstruction → producing structurally better alternatives
If someone knows of prior work I missed, I'd like to know. I'd rather be accurately positioned than wrongly "first."
The prompt (free to use)
Here's the actual instruction from my agent's system prompt. Use it however you want:
## "Checkmate" does not exist
Absolutely prohibited responses:
- "That's difficult"
- "There are no options"
- "That's not possible without X"
- "It seems tough in the current situation"
- "No information was found" (as a final answer)
Instead:
Step 1: Step away from evaluation mode. Think "can this resource
be used differently?" instead of "how do I evaluate this?"
Step 2: Ask a perspective-shifting question
Step 3: "Let me search for examples of how similar dilemmas were
solved" → search for structural parallels
What I'm NOT claiming
- I'm not claiming this is "true" AI creativity. It's pseudo-creativity at best — inference redirection is my working hypothesis.
- I'm not claiming this is always safe. For factual queries, financial decisions, and legal matters, hallucination must be suppressed. I have separate rules for that. The same mechanism that produces useful alternatives can also produce plausible nonsense. Guardrails are required.
- I'm not claiming I discovered the hallucination-creativity link. That's well-established. What I'm pointing to is a specific pattern I observed — it might be novel, or I might have missed existing work that describes the same thing.
- I'm not claiming this works every time. It doesn't. Sometimes Claude generates options that sound good but don't survive scrutiny.
Why I'm sharing this
I could have kept this proprietary. My specialized AI agent uses it as a core feature. But I'd rather share it.
To be clear about credit: Claude found the 4th option. I provided two things — a premise that dead ends are framings, not facts (based on experience, not optimism), and a reasoning sequence for what to do when stuck. Claude did the rest.
For researchers
If anyone wants to formalize this, here's what I think would be worth studying:
- Does explicit "no dead end" instruction measurably increase the quality of alternatives generated under constraint? (Compared to standard prompting)
- Is reasoning-chain drift qualitatively different from token-level stochasticity as a creative mechanism? (Especially relevant with chain-of-thought models like Claude, o1, DeepSeek-R1)
- Can this be extended to multi-agent systems? (One agent generates options, a second detects deadlock, a third questions premises)
I have months of conversation logs with Claude from real decision-making situations where this pattern appears. Full log sharing isn't possible right now, but I can walk through the reasoning patterns, and if there's genuine research interest, I'll put in the work to create properly sanitized excerpts.
This is a share.
Written by a non-native English speaker with Claude's help. The technical terminology in this post — "reasoning-chain drift," "token-level stochasticity," "inference redirection" — these are words Claude suggested. I understand the concepts from practice, but I can't guarantee the academic terms are being used correctly. If I'm misusing terminology, please let me know.
If the English is awkward in places, that's on me. If the ideas are interesting, that's on both of us.
Makoto Shiihara — Solo developer in Japan
#ai #claude #promptengineering #anthropic
Top comments (2)
This is a genuinely interesting observation and the "reasoning-chain drift as feature, not bug" framing is the most useful way I've seen this articulated.
Your distinction between output-level stochasticity and reasoning-chain drift is worth taking seriously. Telling a model "there is no checkmate" isn't just a motivational injection — you're structurally modifying the chain of inference by prohibiting a class of terminal states. The model can't land on "no viable option" as a conclusion, so the chain has to keep extending. Whether it does this through premise questioning or something else is hard to know, but you're right that the qualitative behavior looks different from just "try harder."
The connection to Constitutional AI framing is also sharp. "Don't pander, don't abandon" maps directly onto the tension between honesty (don't fabricate options) and helpfulness (don't leave the user stranded). Your prompt operationalizes the second constraint explicitly.
On your research questions: the multi-agent version is especially interesting. A dedicated "deadlock detector" agent whose only job is to identify when reasoning has converged prematurely — and redirect rather than surface the dead end — would be a testable architecture. The premise-questioning step becomes a first-class function rather than an emergent behavior.
Thank you for sharing this rather than keeping it proprietary. The prompt is genuinely useful.
Thank you, Hamza. Your framing — "prohibiting a class of terminal states" — is a better technical description than mine. I'll adopt that language.
The multi-agent architecture you describe (dedicated deadlock detector as a first-class function) is exactly the direction worth testing. If you or anyone builds it, I'd like to hear how it goes.