Matt Shumer
Matt Shumer (@mattshumer_) is an AI product builder and investor, previously CEO of HyperWrite. In July 2026 he popularized and named the “Gauntlet Loop,” a builder–critic prompting pattern in which lead agents decompose ambitious work, independent fresh-context critics compare real outputs with concrete external quality bars, and failed parts return for another round. [Source: https://x.com/mattshumer_/status/2081830214384886228; https://somethingbig.ai/gauntlet-loop]
Why track
Shumer is a useful practitioner signal for long-running agent work, sparse goal-oriented prompting, external verifiers, fresh-context criticism, observable live workbenches, and multi-agent coordination. His public Claude of Duty repository is unusually valuable because it includes not only a viral result but also harness tooling, performance data, failed approaches, and an explicit statement that the artifact did not meet its aspirational benchmark. [Source: https://github.com/mshumer/Claude-of-Duty]
His broader method is documented in How I Prompt Fable: give the model goals rather than implementation recipes, fence autonomy with house rules, define “done” through real evidence, separate builders from graders, persist progress externally, and provide budgets rather than interrupting every tool use.
Timeline
- 2026-07-03 | Published How I Prompt Fable, describing hard quality bars, separate graders, long-running loops, and workbench-based coordination.
- 2026-07-25 | Published the Claude of Duty demo, followed by the repository and complete prompt.
- 2026-07-27 | Publicly named and defined the Gauntlet Loop and published a detailed guide.
- 2026-07-28 | Demonstrated a live Gauntlet run for a horror novel, extending the pattern beyond games and code.