Gauntlet Mode is for work where “good enough” is not good enough. It turns a broad build request into a repeated cycle of specialist execution, direct reference comparison, harsh review, and targeted improvement.
You only need to replace the two bracketed inputs: what you want built and the example that represents the quality bar. The agent should infer the right quality criteria, important features, technical approach, and specialist workstreams from those two inputs.
How to use it
- Pick one concrete thing to build.
- Choose one real gold-standard example the agent can inspect and compare against.
- Replace only the two bracketed inputs below, then paste the full prompt into your coding agent.
- Give the agent access to the reference, your codebase, and the tools it needs to test the result.
Copy the prompt
You are now operating in GAUNTLET MODE, a high-performance quality protocol for specialist execution, ruthless critique, reference comparison, and repeated improvement.
Build [THING / PROJECT] at the level of [REFERENCE / GOLD-STANDARD EXAMPLE].
Infer the quality bar, success criteria, essential features, technical approach, and specialist workstreams from the build and the reference. Inspect the real reference closely instead of relying on assumptions.
Fan out specialist sub-agents to handle the important parts individually. For each part, use /loop and assign a separate harsh critic to review the result against the real reference.
The critic must identify every meaningful shortfall, explain why it matters, and require concrete fixes before approval. No vague praise and no premature sign-off.
Compare our result directly against the reference across design, usability, performance, reliability, detail, coherence, and any domain-specific qualities that make the reference exceptional. Discover important criteria I did not mention.
If our result loses anywhere, identify exactly why, fix those weaknesses, and /loop again. Re-test the integrated result after every meaningful round so local improvements do not damage the whole.
Choose the most appropriate technology, framework, and implementation details for the job unless the existing project already establishes them. Preserve sound existing architecture where relevant.
Use ultracode for the implementation work. Fan out sub-agents aggressively where parallel work improves quality, but keep one owner responsible for integration and final verification.
Do not stop at good enough. Keep building, critiquing, comparing, fixing, testing, and looping until the result reaches the reference quality bar and the critics can no longer identify a material gap.
Return the finished, verified result, not a plan or a progress report.
Why this version works
The prompt gives the agent a visible benchmark instead of an abstract request to “make it great.” Separate builders and critics reduce self-approval, while repeated passes turn criticism into actual fixes. Keeping only two inputs makes it fast to reuse without forcing you to manually list obvious features, quality dimensions, or technologies the agent can infer by studying the reference.
For the strongest result, choose a reference that is specific, accessible, and genuinely excellent at the same job as the thing you want to build.