Open Source

10 GitHub Repos for Clearer, Leaner AI Coding

Direct Answer

The strongest repositories in this roundup do not make the model smarter. They make the working environment clearer. I Have ADHD changes how an agent communicates. ECC and HumanLayer encode engineering routines. Ponytail pushes toward smaller implementations. Archify exposes architecture visually. Context Mode keeps bulky tool output outside the conversation. Matt Pocock's skills turn planning and review into repeatable verbs.

That distinction matters. Many agent failures begin as workflow failures: vague requirements, excessive output, hidden architecture, duplicated code, lost context, or a missing review step. A small instruction file can help when the same failure repeats. A plugin becomes more useful when the workflow also needs integrations, commands, scripts, or packaged tools.

Do not install all ten. Pick one observable failure, test one repository on three comparable tasks, and keep it only if accepted work improves without creating new permission or maintenance problems.

Watch the GitHub Roundup

Credit and evidence note: the demonstrations and commentary come from Andrew Warner and Adam Brakhane's episode, published on 11 September 2026. Repository links and supplemental media were organized from the episode's GitHub Show research page. Current repository status was checked on 13 September 2026.

The Repository Map

RepositoryJobBest first testMain caveat
I Have ADHDConcise, action-first outputCompare one troubleshooting answerCommunication preference, not medical guidance
ECCStructured coding-agent processRun one bounded feature ticketLarge packs can add unused instructions
PonytailReduce over-engineeringReview one diff for deletionsSimplicity can become under-building
HumanLayer SkillsReusable engineering loopsImprove one project instruction fileChoose only the relevant skill
OpenAI PluginsPackage skills with toolsInspect one official pluginMore capability means more review
ArchifyInteractive architecture diagramsMap one important code pathValidate every claimed connection
Context ModeExternalize large tool resultsCompare one log-analysis taskBad retrieval can erase savings
Matt Pocock SkillsPlan, spec, ticket, build, reviewRun Grill Me before one featureThe mental model may be enough
HumanizerRemove recurring AI-writing patternsEdit one draft against a voice sampleNeutral prose is not your voice
OpenAI SkillsHistorical standalone catalogUse as migration contextDeprecated; use Plugins

1. Clearer Output: I Have ADHD

I Have ADHD is a compact skill for changing response shape. It asks the agent to lead with the next action, number steps, cap lists, remove preambles and recaps, and pause when an assumption may be wrong. The before-and-after is the product: less prose between the user and the work.

This is useful when a technically correct agent still creates cognitive overhead. It is also easy to fork because the behavior is visible in one Markdown file. Rename it for a team communication standard if the original framing is not appropriate. The repository does not diagnose ADHD and should not be treated as health guidance.

Related media: see the X demonstration collected by GitHub Show or watch AICodeKing's installation walkthrough.

2. Engineering Discipline: ECC, Ponytail, and HumanLayer

ECC: a broad operating pack

ECC packages agents, commands, skills, rules, and hooks into a fuller engineering environment. That breadth is its advantage and its risk. A new project can gain a shared process quickly, but every extra instruction competes for attention and can age. Audit which components actually run, remove overlaps, and version the chosen configuration with the codebase.

Ponytail: strategic laziness

Ponytail applies a senior-engineer bias toward the smallest adequate change: use what exists, avoid speculative abstractions, and review diffs for code that can disappear. That is valuable because agents are often rewarded for producing visible volume rather than reducing future maintenance.

Better Stack explains Ponytail's decision ladder, caveats, and a side-by-side test. The accompanying Miles Cranmer X post shows why the idea spread. Treat benchmark percentages as claims to reproduce, not guarantees.

HumanLayer: reusable control loops

HumanLayer's skills cover focused jobs such as improving project instructions, narrowing React prop types, building iterated agent loops, and explaining code visually. The useful pattern is not a huge universal prompt. It is a narrow routine with a clear trigger and output. Matt Pocock's show-me recommendation on X is a practical example.

3. OpenAI Plugins Are the Packaging Layer

The episode calls OpenAI Plugins an app store for Codex. The important technical distinction is packaging. A skill can teach a method; a plugin can bundle that method with app connectors, MCP configuration, commands, scripts, agents, hooks, and supporting assets.

OpenAI did not remove skills from Codex. It deprecated the standalone openai/skills catalog and directs users to the plugins repository. Skills remain part of the plugin format. The change says that instructions alone are often not enough for a complete reusable workflow.

Owain Lewis installs and uses a Linear plugin from the Codex plugin marketplace starting at 15:05. The video demonstrates the feature rather than documenting the repository itself.

Permission check: before installing a plugin, inspect every skill, command, executable script, hook, dependency, MCP server, app connection, and requested credential. Test it in a disposable workspace before using private or production data.

4. Make Architecture Visible and Context Retrievable

Archify: force a testable picture

Archify turns architecture and workflows into interactive diagrams. A diagram is useful precisely because it is less forgiving than a long explanation: missing boundaries, impossible arrows, and unclear ownership become visible. The repository can map architecture, workflows, sequence, data flow, and lifecycle views.

The demo installs Archify, maps a real code path, inspects architecture diffs, and compares it with other diagram approaches. Also see Voxyz's prompt and install example on X.

Context Mode: move bulk out of the conversation

Context Mode keeps raw tool output in a local searchable layer and returns a smaller result to the agent. The mechanism can help with logs, searches, and large command output. It is not free compression: if the retained summary omits a needed detail, the agent may retrieve the source again and pay for the growing conversation history.

The walkthrough covers context virtualization, session checkpoints, installation, log analysis, and a cost review. Measure the repository against an unchanged control task before accepting headline savings.

5. Matt Pocock Skills and Humanizer

Learn the verb before installing the skill

Matt Pocock's repository turns a development sequence into shared vocabulary: grill the idea, write a spec, create tickets, implement, and review. Grill Me works because it exposes assumptions before code makes them expensive. In many cases, saying "grill me about this plan" is enough. The repository is still useful as a library of mental models and as a consistent team interface.

The Theo recommendation on X focuses on alignment: the agent asks until it has extracted decisions the initial prompt omitted. For larger work, specs and tickets also survive context-window boundaries and allow separate agents to work from the same contract.

Humanizer removes patterns; your voice still needs evidence

Humanizer targets repeated AI-writing habits such as theatrical framing, forced triads, inflated significance, decorative formatting, and chat leftovers. That can make a draft cleaner, but passing an AI detector is not a meaningful editorial standard. Detection tools are unreliable, and generic neutral prose is not the same as a person's voice.

Use the skill as one editing pass after the writer has supplied original ideas, examples, claims, and a real voice sample. The X example discussed in the episode demonstrates the pattern-removal claim; it does not prove authorship.

6. Four Audience Projects Worth Separating From the Top Ten

ProjectWhat it doesUseful first test
ReckonerCombines AI-provider balances and limits in one dashboardAdd one provider and verify warning accuracy
Tim Harris SkillsPackages agent workflows for users who do not read codeRun one skill and inspect every generated artifact
ClodexUses GPT models beneath a Claude Code-style terminalCompare it with the native Codex harness on one task
AI Marketing PanelUses simulated customer personas to critique copyCompare predictions with five real customer interviews

These projects solve different problems, so star counts do not make a useful ranking. Reckoner is operational visibility. Tim Harris's collection is accessibility. Clodex is model-harness experimentation. AI Marketing Panel is synthetic research. The last category needs the strongest caveat: simulated personas can generate hypotheses, but they cannot replace observations from real customers.

A Seven-Day Repository Test

  1. Name one failure. Choose verbosity, vague requirements, bloated code, invisible architecture, context overflow, or weak prose.
  2. Save a control task. Keep the prompt, repository state, model, reasoning level, and acceptance criteria fixed.
  3. Inspect before installing. Read instructions and scripts; list network, file, shell, credential, and external-app access.
  4. Run the baseline. Record tokens, elapsed time, retries, changed lines, test results, and reviewer corrections.
  5. Add one repository. Do not combine several packs while trying to identify the source of improvement.
  6. Repeat three times. One impressive output may be luck. Look for a stable reduction in rework.
  7. Keep, fork, or remove. Keep proven behavior, fork what needs tailoring, and delete instructions that do not change outcomes.

Video Chapters

TimeTopicTimeTopic
00:00I Have ADHD10:17Archify
01:40ECC engineering process11:36Context Mode
03:02Ponytail13:23Matt Pocock Skills
05:20Zapier MCP15:29Learning new AI workflows
05:55HumanLayer Skills16:07Humanizer
08:22Should the show skip skills?17:59OpenAI Skills deprecation
08:53OpenAI Plugins18:54Audience submissions
19:28Reckoner21:34AI Marketing Panel
20:05Tim Harris Skills22:41Wrap up
20:36Clodex

Verdict

The useful unit is not the repository. It is the repeated behavior that survives a real test. I Have ADHD is a response contract. Ponytail is a simplicity constraint. Grill Me is a requirements interview. Archify is a visual verification surface. Context Mode is a retrieval strategy. Plugins package those behaviors with the tools needed to execute them.

The episode's debate about whether skills deserve a weekly ranking is healthy. A popular skill may teach a valuable mental model without deserving permanent installation. Start by using the idea in plain language. Package it only when repetition, team consistency, or tool access makes the file valuable.

Source Links

Episode and research page

Repositories

Repository behavior, installation methods, licenses, dependencies, and security posture can change. Recheck the current README and source before installation.

Common questions

What is the best first repository from this roundup?
Ponytail is useful when agents over-engineer; I Have ADHD is useful when their explanations are too long; and Grill Me from Matt Pocock is useful when requirements are unclear. Choose the friction you can measure instead of installing every repository.
Did OpenAI deprecate skills?
OpenAI deprecated the standalone openai/skills catalog in favor of openai/plugins. Skills still exist inside plugins alongside connectors, commands, scripts, agents, and other resources.
Can I use the I Have ADHD skill as medical advice?
No. It is a set of communication preferences for concise, action-first agent output. It is not a diagnostic or treatment tool and should not be presented as one.
Do context-compression tools always save tokens?
No. They can reduce repeated tool output, but a poor summary may force the agent to retrieve the source again. Measure total tokens, retries, accepted results, and time to completion on your own tasks.
Should I install third-party agent skills without reviewing them?
No. Inspect instructions, scripts, dependencies, hooks, network access, and requested permissions. Test in a disposable project before granting access to credentials, production systems, or private data.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call