Loopy

Loopy is Forward Future's loop library plus an installable loopy skill for finding, adapting, crafting, running, and debriefing repeatable AI-agent workflows. It should be Kevin's default route when a prompt says "keep improving" but lacks a check or stop condition.

Routing Summary

Use loopy when the task is to create or run a bounded feedback loop: code improvement, product evaluation, documentation freshness, skill quality, bookmark absorption, content quality, or other recurring agent work. Pair it with Agent Self-Improvement Eval Library when the loop's judge needs to be benchmarked or evolved.

Do not use Loopy as permission for an agent to run forever. A usable loop needs a goal, check, next action, stop condition, and handoff condition. Source: Forward-Future/loopy README, 2026-06-29

What It Adds

Loopy separates the library from the skill:

This maps directly to Kevin's No One-Off Work rule: when a workflow repeats, first run it manually, then turn the pattern into a bounded loop, then codify it as a skill or automation when it proves useful.

The library launched as a community catalog with roughly 50 loops and later reached 70. It exposes catalog.json, catalog.md, catalog.txt, llms.txt, RSS, and an /agents/ surface. The launch also used here.now as an agent-oriented publishing/storage companion. These are discovery surfaces, not quality guarantees: community loops remain reference candidates until Kevin's evaluator and stop-condition checks pass. Source: X/@MatthewBerman, 2026-06-18; Loop Library live catalog, 2026-07-03

Version Snapshot

Checked 2026-07-03:

• GitHub repo: Forward-Future/loopy
• Default branch: main
• HEAD: cc7cb1712e104e895660524410035c7b4fa4b44c
• Tags: no tag refs returned by git ls-remote
• License: MIT
• Stars: 2,304
• Forks: 201
• Live Loop Library catalog: 70 loops, updated 2026-06-26, with catalog.json, catalog.md, catalog.txt, llms.txt, and /agents/ surfaces
• npm package @forward-future/loopy: not published; install route is npx skills add Forward-Future/loopy --skill loopy ...

Source: GitHub CLI, git ls-remote, live catalog API/markdown, and npm registry, 2026-07-03

Kevin-Stack Use

• Skill creation: use local loopy before skill-creator when the requested skill is a repeated loop rather than a static procedure.
• Agent self-improvement: use Agent Self-Improvement Eval Library for the evaluator, then Loopy for the improvement loop around that evaluator.
• Bookmark artifacts: when a bookmark describes a workflow, promote the thought into a loop skill, playbook, or automation instead of leaving a hub note.
• Product/content no-slop loops: the reviewed marketing-agent diagram is a Loopy-shaped workflow: context gate, freshness gate, two-model review, taste/proof, then a loop back to research. Route the executable content rule through Writing and Content Skills and Social Content OS. Source: X/@EXM7777 and local artifact review, 2026-07-04

The current Loopy source at 75966cbd makes its three useful operator stages explicit: interview to design a loop, run/monitor the loop, then review failed or surprising runs. Kevin's durable loop manifest should therefore record the goal and owner; source/input cursor; step and state schema; evaluator and acceptance digest; time/token/cost/attempt budgets; no-progress and repeated- error counters; human pause/cancel/escalation; outputs and side effects; and the exact condition for completion, exhaustion, expiry, stall, failure, or handoff. /goal is one runtime interface over this state machine. An evaluator that sees only the transcript cannot prove hidden external state, and a prompt cannot raise its own budget or weaken its stop rule. Source: X 2069729160600203595, 2070876072619765805; recovered /goal X Article 2054806436883070977; Forward-Future/loopy@75966cbd, reviewed 2026-08-12

Claude Of Duty: What “Until 100” Actually Requires

Claude of Duty is a strong negative and positive fixture for ambitious Loopy prompts. Its harness used named subsystem ownership, deterministic RNG, reproducible isolated screenshots, pixel diffs, scripted playtests, and gameplay frame-time distributions instead of a static median. But the repo's own report also found that broad parallel fan-out made coupled rendering concerns fight one another. Sequential ownership per coupled concern plus independent critics improved the score most. Source: mshumer/Claude-of-Duty README and ARCHITECTURE.md, commit d9b237b, checked 2026-08-10

The honest result matters: eleven critics moved the visual score from 3.59 to 5.05/10, and every blind comparison still chose the real Call of Duty frame. Therefore “loop until 100” means 100% of a frozen, inspectable gate set—not an unbounded instruction to self-declare perfection. A strong loop stops on all gates passing, a demonstrated hard ceiling, or repeated no-improvement, and records the remaining gap. Source: X/@mattshumer_, 2026-07-25

---

Timeline

• 2026-08-12 | Refreshed Loopy to 75966cbd and turned its interview, run/monitor, and review stages into a durable loop-manifest contract with harness-owned budgets, stop states, interrupts, outputs, and side effects. Source: current Loopy source and recovered /goal article

• 2026-08-10 | Added Claude of Duty as the ambitious-quality loop fixture: serialize coupled systems, parallelize independent owners, use reproducible comparative proof and percentile performance, question the requested fix, and stop with an honest ceiling rather than a fake 100. Source: User request; X/@mattshumer_; mshumer/Claude-of-Duty

• 2026-07-15 | Folded the redundant Loop Library article into this canonical Loopy page, preserving the launch artifact, catalog history, agent-readable surfaces, here.now context, community-quality caveat, citations, and old slug through a redirect. Source: Wiki Corpus Refactor Program; source-preserving consolidation review, 2026-07-15

• 2026-07-04 | Attached the reviewed agentic marketing no-slop workflow as content-loop evidence: context, freshness, review, taste/proof, and loopback are the minimum bounded-loop shape. Source: X/@EXM7777, 2026-06-28; local artifact review, 2026-07-04

• 2026-07-03 | Refreshed Loopy and Loop Library source status: repo still untagged at HEAD cc7cb17, 2,304 stars, 201 forks, MIT; live catalog now exposes 70 published loops and agent-readable catalog files. Source: GitHub and Loop Library catalog, 2026-07-03

• 2026-06-29 | Promoted from hub artifact to first-class tool page and local loopy skill after Kevin requested using Loopy to craft loop skills. Source: User request, 2026-06-29; raw/x-bookmarks/enriched/2070876072619765805.json