i-have-adhd Tops GitHub Trending: New Trend in Agent Output Shape Control

i-have-adhd: Curing Agent Verbosity

ayghri/i-have-adhd is an Agent Skill that topped the GitHub daily trending list for multiple consecutive days. It enforces 10 rules and 4 pre-send deletion checks, with a single core mandate: the first line must state "what to do next." Multi-step tasks are auto-numbered and capped at 5 steps. All background introductions, repetitive summaries, and sign-offs are stripped.

Run npm install jsonwebtoken@latest, then edit src/auth.ts:42. Open src/auth.ts Replace verifyToken (lines 42-58) with the code below Run npm test -- auth.spec.ts next: if any test fails, paste the first failing line.

The project has 42k stars, uses the MIT license, and supports Claude Code, Codex, Cursor, OpenCode, and Gemini CLI. The name "ADHD" is a metaphor — no diagnosis required — highlighting that the real problem is the "shape" of agent responses.

humanizer: Curing AI-Like Writing

blader/humanizer addresses a different issue: AI-generated text that is instantly recognizable as machine-written. Its detection checklist is not heuristic; it is based on the Wikipedia "Signs of AI writing" guide maintained by WikiProject AI Cleanup, derived from observation of thousands of AI text instances. Version 3.0.0 (released September 6) consolidated 33 patterns into 25, ordered by intensity.

The rewrite runs in two passes: first removing AI patterns, then performing an "anti-AI audit" to catch residual traces. A no-fabrication rule prevents the addition of facts, names, dates, or citations not present in the source text — solving a common flaw where anti-AI tools invent content to sound human. A voice calibration feature analyzes the user's own writing samples to apply their sentence rhythm, vocabulary habits, and punctuation preferences, producing output that matches personal style rather than a generic "clean" voice. The project has 47k stars with a weekly growth of 4,243.

Three Projects, One Emerging Track

Together with Nutlope/hallmark (57 slop-tests for AI design taste, from July), a complete "output quality control" track has formed:

hallmark — checks UI design taste

humanizer — checks text AI taste

i-have-adhd — checks response shape

The trend is traceable: May's taste-skill defined design taste rules, July's hallmark checked AI design taste, August's ponytail reduced agent over-engineering of code. This week i-have-adhd and humanizer push control granularity down to "how every sentence is written."

archify: Three Consecutive Weeks, Sustained 10k+ Heat

tt-a1i/archify appeared for the third consecutive week. Weekly star growth sequence: 8,800 → 26,626 → 10,493. The drop from 26k to 10k is normal — last week's spike came from the v2.16.0 release; this week returns to steady-state growth. Sustained 10k+ heat across three weeks signals persistent demand.

archify generates architecture diagrams from a codebase where every node traces back to specific files and Git commits. Identical input yields identical output, enabling architecture snapshots to be diffed for PR reviews. This week cathrynlavery/diagram-design also returned (weekly growth 7,564, total 38.5k). The two projects illustrate a clear divide: diagram-design prioritizes "brand fit" — reading website colors in 60 seconds to make charts look good; archify prioritizes "verifiability" — tracing nodes to files to make charts accurate. Both "pretty" and "accurate" are real needs, but archify's consistently higher heat suggests "accurate" is closer to the core engineering pain point.

Weekly Outlier: Writing Video with HTML

heygen-com/hyperframes is HeyGen's open-source HTML video rendering framework. Developers write HTML/CSS and the framework deterministically renders to MP4. Weekly growth 5,029, total 49k stars. It takes a different path from calesthio/OpenMontage (which topped the list for three consecutive weeks in June). OpenMontage orchestrates multiple AI vendors for end-to-end video production. hyperframes reduces video creation to "writing HTML" — agents use existing web development knowledge to author frames, while the framework uses headless Chrome for frame-by-frame rendering and FFmpeg for encoding.

The key differentiator is determinism. OpenMontage produces different results each run; hyperframes yields pixel-perfect consistency for identical inputs. This has practical value for CI/CD regression testing and batch production. Twenty Skills cover the full pipeline from planning to rendering, compatible with Claude Code, Cursor, and Codex.

VoiceStudio: Local-First Voice Workstation

debpalash/VoiceStudio brings the "local-first" wave to the voice domain. It bundles 16 TTS engines and 11 ASR engines supporting 646 languages, all running locally. Weekly growth 7,513, total 20k stars. Over recent weeks "local-first" has appeared in video (OpenMontage), coding (magnitude), and inference (colibri); VoiceStudio extends that line into speech. The AGPL license and "active testing" tag restrict commercial use, but individual developers can use it directly.

What This Week Reveals

The agent capability layer is already sufficient. Models are strong enough, tools are plentiful, and Skills are rich. The bottleneck has shifted — from "can the agent do it?" to "does the agent do it right, say it well, and can it be verified?"

i-have-adhd's #1 ranking isn't because it makes agents smarter. On the contrary, it makes them say less. But every sentence is useful. That is what developers actually want.

References

ayghri/i-have-adhd

blader/humanizer

Nutlope/hallmark

tt-a1i/archify

cathrynlavery/diagram-design

heygen-com/hyperframes

calesthio/OpenMontage

debpalash/VoiceStudio