Twitter AI Coding - 2026-05-22¶
1. What People Are Talking About¶
1.1 Skills, playbooks, and workflow telemetry became first-class assets π‘¶
The strongest cluster of posts treated AI-coding leverage as something to package, inspect, and reuse, not just something to prompt ad hoc. The highest-signal evidence combined Codex skills distribution, a new local workflow-audit dashboard, and GitHub opening more of Copilot's agent stack to public inspection. This theme was supported by at least three substantial items.
@jxnlco said (186 likes, 16 replies, 6,824 views, 128 bookmarks) that Codex users can ask Codex to inspect the openai/skills installer and recommend skills to install. The linked repo describes itself as a skills catalog for Codex, says .system skills are auto-installed, and explains how curated and experimental skills can be installed with $skill-installer, which turns skills from an abstract concept into a concrete distribution mechanism.
@akshay_pachaar reported (33 likes, 12 replies, 2,948 views, 59 bookmarks) that Microsoft had open-sourced AI Engineer Coach. The repo says it reads local AI session logs, tracks practice scores and weekly trends, detects 45 anti-pattern rules, measures output by harness and model, discovers repeated prompts as skills, and scores context health; the attached screenshot shows those ideas as a working dashboard rather than a concept.
@GHchangelog announced (27 likes, 1,522 views) that GitHub Copilot for Eclipse is now open source under MIT. The linked changelog post and repository say developers can now inspect chat, completions, agent mode, MCP integration, skills, BYOK, custom agents, isolated subagents, and plan-agent behavior directly in code.
Discussion insight: Replies under the skills post asked for a shared marketplace, better access to computer use and Chrome in Europe, and easier ways to push meeting notes into project context, which suggests distribution and durable memory are still the missing layer around these packaged capabilities.
Comparison to prior day: Compared with May 21, where first-party Codex rollout notes dominated, May 22 shifted toward user-facing ways to install, inspect, and audit agent workflows.
1.2 Quotas and usage controls became part of the product story π‘¶
The second major theme was that usage limits, reset timing, and model controls were no longer background billing details. They were central enough to generate status-page screenshots, cancellation threats, and new tools built specifically around usage visibility. This theme was supported by at least four high-signal items.
@TimJayas said (75 likes, 11 replies, 3,674 views) that two Gemini Pro questions consumed most of the day's allowance even after Google had increased weekly Antigravity and Gemini quotas. The attached screenshot shows 83% used on current usage with only 4% used on the weekly limit, and replies called that ratio "nasty" for agent workflows and said they were canceling subscriptions.
@weswinder noted (68 likes, 10 replies, 3,360 views) that OpenAI appeared aware of Codex limit issues and that a reset might be coming. The attached screenshot from the status site shows an "Increase in users hitting Codex rate limits" incident marked Investigating and Degraded performance, which turned anecdotal complaints into a public operations issue.
@marouane53 complained (73 likes, 7 replies, 2,559 views) that Codex had removed per-model thinking controls and heavy mode, and said that kind of quiet change undermined trust in the workflow. The screenshot shows a simplified intelligence picker with Instant, Medium, High, and Pro under GPT-5.5, which gives the complaint a concrete UI target rather than leaving it as a vague grievance.
Discussion insight: Not every reply confirmed the same severity - one response to weswinder said Codex still felt nearly unlimited for hours-long runs - but the volume of screenshots and vendor-facing complaints shows that people were now debugging plans and controls as much as model outputs.
Comparison to prior day: May 21 already featured Google's 3x quota-reset announcement; May 22 showed that the usage issue had not cooled down and had expanded into public Codex incident tracking too.
1.3 Builders kept wrapping existing agents with team, search, and messaging layers π‘¶
The most concrete build activity did not try to replace Codex, Claude Code, Copilot, or OpenCode. Instead, builders added shared servers, semantic indexing, chat bridges, and local routing layers around them. This theme was supported by at least four project-sharing items.
@thdxr showed (90 likes, 21 replies, 16,907 views) a gangprompt.opencode.ai setup behind Cloudflare Access that runs an OpenCode server on a fast machine with all repos cloned so the whole team can prompt against shared context and see one another's sessions. One reply said they had built something similar and another said Microsoft still had not added an equivalent for private GitHub organizations, which makes the shared-harness gap explicit.
@Suryanshti777 shared (13 likes, 8 replies, 348 views) CodeGraph as a local knowledge graph for Claude Code, Cursor, Codex CLI, OpenCode, and Hermes Agent, then linked the colbymchenry/codegraph repo. The repo says it is a pre-indexed code knowledge graph and publishes benchmark results across seven open-source codebases showing an average 35% cheaper, 59% fewer tokens, 49% faster, and 70% fewer tool calls.
@zarazhangrui introduced (7 likes, 3 replies, 391 views) an open-source Feishu / Lark bridge for Claude Code. The repo says it forwards Lark messages into the local claude CLI, supports per-chat sessions, multiple workspaces, images and files, and interactive cards; the screenshot shows those capabilities exposed inside Lark as a persistent colleague-style chat surface.
@BeauJohnson89 argued (29 views) that 0xSero/codex-shim gives Codex Desktop a local model escape hatch. The repo says it is a local Responses API shim that exposes Factory BYOK models and optional ChatGPT GPT-5.5 passthrough to Codex Desktop without recompiling the app.
Discussion insight: Replies around CodeGraph and the shared OpenCode server focused on tradeoffs - how much setup the graph requires, whether shared context creates new failure modes, and whether these layers reduce noise or add maintenance - but the build pattern itself was consistent.
Comparison to prior day: This extends May 21's interoperability theme, but today's examples leaned more toward team control planes, messaging surfaces, and local routing than PR-review plugins alone.
1.4 Antigravity stayed polarizing after the 2.0 rollout π‘¶
Google's Antigravity remained one of the busiest names in the feed, but not in one direction. Some users described the new setup as clearly better than rival models, while others focused on limits, rollback instructions, and churn in the product surface. The result was a polarized theme rather than a simple up-or-down story.
@haider1 said (29 likes, 10 replies, 2,043 views) that Antigravity 2.0 was "far better than opus 4.7" for cases where a solution looked correct but was actually hacky or broken. That is the strongest direct performance praise in the day's feed.
@TimJayas said (75 likes, 11 replies, 3,674 views) that quota burn was bad enough to make normal Gemini use feel unrealistic, and replies said they were canceling subscriptions after short sessions. @maloymediika said (2 likes, 2 replies, 104 views) Antigravity 2.0 had split the IDE into separate apps and wiped settings, leading some users to roll back or uninstall.
@ash_twtz asked (38 likes, 20 replies, 706 views) what the problem with the new Google Antigravity was, and the attached screenshot shows the linked video description framed as "How to go back to old Google Antigravity IDE," which is a clear piece of rollback evidence even though the tweet itself is just a prompt for opinions.
Discussion insight: The split between "better than Opus 4.7" and "I canceled my subscription" suggests that model quality and product policy were being evaluated separately.
Comparison to prior day: May 21's quota reset did not settle the Antigravity debate; May 22 still mixed fresh praise, rollback curiosity, and hard complaints about everyday usage.
2. What Frustrates People¶
Opaque quota math and sudden rate-limit burn¶
The clearest frustration was that people could not predict how fast premium capacity would disappear. @TimJayas said (75 likes, 11 replies, 3,674 views) that two Gemini Pro questions burned through most of the day's allowance, and the attached usage screen shows 83% used for the current window despite only 4% used of the weekly limit. Replies under that post said agents would "bankrupt that plan before lunch," that one user had canceled a subscription, and that another switched an image-generation task to ChatGPT after Gemini consumed most of the limit without satisfying the prompt. Severity: High.
Codex users were voicing the same complaint from the OpenAI side. @weswinder said (68 likes, 10 replies, 3,360 views) OpenAI was aware of accelerated Codex limit burn, and the screenshot showed a public rate-limit incident on the status page. @ziwenxu_ said (15 likes, 8 replies, 443 views) the status page confirmed users had been "screaming about for days" before a possible reset. Worth building for: Yes. The feed points to a direct need for live usage visibility, reset timing, and clearer explanations of why one short session can consume a day's budget.
Workflow controls keep changing underneath users¶
People were also frustrated when important controls disappeared or stayed server-gated. @marouane53 complained (73 likes, 7 replies, 2,559 views) that Codex had quietly removed the ability to choose thinking level per model and that heavy mode no longer existed, which made him question whether the vendor was saving tokens at the user's expense. The screenshot of the simplified GPT-5.5 intelligence picker made the complaint concrete, and replies said other users were seeing the same rollout. Severity: High.
The other version of the same frustration is server-side gating. @BeauJohnson89 argued (29 views) that Codex Desktop only exposes models on a server allowlist and pointed to 0xSero/codex-shim as a local workaround. The repo says the shim exposes Factory BYOK models and optional ChatGPT GPT-5.5 passthrough to Codex Desktop, which means users are building local escape hatches rather than waiting for the official picker to open up. Worth building for: Yes. People want stable control surfaces and explicit routing, not silent UI changes.
Team context is still awkward to share safely¶
Shared context remains hard enough that teams are building their own infrastructure. @thdxr showed (90 likes, 21 replies, 16,907 views) a private OpenCode server with all repos cloned and Cloudflare Access in front so the whole team can prompt against one environment and see ongoing work. One reply said they had built something similar, while another said Microsoft still had not shipped an equivalent for private GitHub organizations. Severity: Medium.
Replies on the skills and bridge posts pointed to the same gap from a different angle. Under @jxnlco sharing (186 likes, 16 replies, 6,824 views, 128 bookmarks) the skills installer flow, one reply asked for meeting-note recording that automatically updates project context and another asked for easier access to computer use and Chrome in Europe. @zarazhangrui showed (7 likes, 3 replies, 391 views) one workaround: moving Claude Code into Lark with per-chat sessions and document actions. Worth building for: Yes. The demand is for controlled, shared, durable context that works across teams and devices.
Regional access restrictions are creating a proxy economy¶
The most structural frustration came from access itself. @Akasheth_ said (24 likes, 15 replies, 196 views) an Oxford researcher had exposed the underground market for cheap Claude and OpenAI access in China, then linked the article "How to Buy Cheap Claude Tokens in China". The article says API proxies, transfer stations, and related evasion infrastructure let Chinese developers access Anthropic models at as low as 10% of the official price, and that the market is visible on GitHub, Taobao, Twitter, and Telegram. Severity: High.
This is not a complaint about model quality; it is a complaint about getting compliant access at all. The coping mechanism is already visible in public: proxy layers, regional routing, and grey-market resellers. Worth building for: Yes, if the product is compliant and traceable. The current evidence shows people will route around official access barriers when the official path is absent or too expensive.
3. What People Wish Existed¶
Usage visibility that lives inside the workflow¶
The most practical wish in today's feed was not for a better model but for better visibility into what the current models were costing and when limits would reset. @_brian_johnson said (207 views) that this was exactly the pain TokenBar was built for: live Codex and Claude usage, cost, reset timers, and model mix in the Mac menu bar. The TokenBar site says it keeps usage, credits, resets, and alerts visible across OpenAI, Claude, Gemini, OpenCode, and other supported tools, while AI Engineer Coach adds local charts for progress, anti-patterns, and output. Opportunity rating: direct.
Reusable project memory and skill distribution¶
@jxnlco showed (186 likes, 16 replies, 6,824 views, 128 bookmarks) that Codex can inspect the openai/skills installer and suggest skills to install, but the replies immediately asked for the next layer: a shared marketplace, easier access to browser and computer-use capabilities, and the ability to record meetings so project context stays updated. That is a clear request for durable operating memory that teams can load rather than retype. Opportunity rating: direct.
Remote and mobile control over local coding agents¶
The feed also showed demand for a way to use local agents from outside the terminal. @zarazhangrui introduced (7 likes, 3 replies, 391 views) a Feishu / Lark bridge that lets Claude Code run through chat with per-session threads, document actions, images, files, and interactive cards. The screenshot matters because it shows a persistent chat surface with slash-command style workflows and document publishing, not just a toy bot. Opportunity rating: competitive.
Shared team workspaces for promptable repos¶
@thdxr showed (90 likes, 21 replies, 16,907 views) one version of this wish already being built: a shared OpenCode server with repos cloned centrally so teammates can prompt against the same environment and inspect one another's work. The reply saying Microsoft had not added something like that to private GitHub organizations is the unmet-need statement in plain language. Opportunity rating: direct.
Official model routing instead of local escape hatches and grey markets¶
Two different posts pointed at the same missing feature: officially supported routing. @BeauJohnson89 pointed (29 views) to 0xSero/codex-shim because Codex Desktop's picker stays constrained by a server allowlist, while the ChinaTalk article linked by @Akasheth_ sharing (24 likes, 15 replies, 196 views) the proxy economy shows what happens when official access is too rigid or unavailable. The wish here is not abstract openness; it is supported, compliant routing for the models and regions people already use. Opportunity rating: competitive.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| OpenAI Codex | Coding agent | (+/-) | Skills catalog, strong interface mindshare, active community workflows around it | Rate-limit incident, disappearing thinking controls, server-side model allowlists |
| Google Antigravity | Agent IDE / workspace | (+/-) | Some users say 2.0 beats Opus 4.7 and works well with Gemini tiers | Quota burn, rollback sentiment, settings churn, split-app confusion |
| GitHub Copilot | IDE / coding agent | (+) | Eclipse plugin is now open source with agent mode, MCP, BYOK, custom agents, and skills | Usage-based billing transition still hangs over the ecosystem |
| Claude Code | Terminal agent | (+) | Remains a favored base runtime for bridges, skills discussion, and comparison testing | Users keep building extra layers for messaging, shared context, and remote control |
| AI Engineer Coach | Workflow observability | (+) | Local-only session analysis, 45 anti-pattern rules, skill discovery, context-health scoring | Requires local log collection and is still early enough to ship as a VS Code extension package |
| CodeGraph | Semantic code index | (+) | Local knowledge graph, published repo benchmarks, broad support across Claude Code, Codex, Cursor, OpenCode, and Hermes Agent | Adds indexing and maintenance overhead that reply threads explicitly questioned |
| TokenBar | Usage monitor | (+) | Glanceable menu-bar view of usage, credits, resets, and alerts across providers | Mac utility rather than a full cross-platform control plane |
| OpenCode | Agent server / shared harness | (+) | Can be centralized behind SSO so a team can work against shared repos and visible sessions | Needs hosted infrastructure and stronger guardrails when many people share one environment |
The overall satisfaction spectrum was pragmatic rather than loyalist. @haider1 said (29 likes, 10 replies, 2,043 views) Antigravity 2.0 was better than Opus 4.7 on some broken-solution cases, while @TimJayas said (75 likes, 11 replies, 3,674 views) the same Google stack could burn through a day's limit in two questions. On the Codex side, @jxnlco showed (186 likes, 16 replies, 6,824 views, 128 bookmarks) practical upside through installable skills, while @weswinder showed (68 likes, 10 replies, 3,360 views) that rate-limit behavior had escalated to a public incident.
The main workaround pattern was to add another layer instead of replacing the base tool. @Suryanshti777 shared (13 likes, 8 replies, 348 views) a local semantic graph for faster repo navigation, @zarazhangrui introduced (7 likes, 3 replies, 391 views) a messaging bridge for Claude Code, @thdxr showed (90 likes, 21 replies, 16,907 views) a shared OpenCode server for team use, and @BeauJohnson89 pointed (29 views) to a local Codex model shim. The competitive dynamic is no longer just which model is best. It is which runtime becomes easiest to observe, route, and extend.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| OpenAI Skills | OpenAI | Catalog of installable Codex skills with system, curated, and experimental packs | Makes repeatable agent capabilities easier to distribute and reuse | Python, Markdown skill packs, Codex installer | Shipped | GitHub |
| AI Engineer Coach | Microsoft | Local dashboard that analyzes AI coding sessions across harnesses | Gives teams observability into prompt quality, context health, anti-patterns, and code-review habits | TypeScript, VS Code extension, local session logs | Beta | GitHub |
| Gangprompt OpenCode server | @thdxr | Shared OpenCode server with repos cloned centrally and sessions visible to the team | Solves team context sharing and lets multiple people work from the same repo state | OpenCode server, Cloudflare Access, centralized repos | Beta | tweet |
| CodeGraph | Colby McHenry | Pre-indexed semantic code graph that agents query instead of scanning files | Reduces token, time, and tool-call overhead on large repositories | TypeScript, local index, MCP-style agent integration | Shipped | GitHub |
| Feishu / Lark bridge for Claude Code | @zarazhangrui | Runs Claude Code through Lark chat with per-chat sessions, files, and cards | Gives users remote and mobile access to local coding sessions | TypeScript, Node.js, Claude CLI, Lark app | Beta | GitHub |
| codex-shim | 0xSero | Local Responses API shim that exposes BYOK and passthrough models inside Codex Desktop | Works around server-side model allowlists and routing limits | Python, local API shim, Factory settings | Alpha | GitHub |
@jxnlco pointed (186 likes, 16 replies, 6,824 views, 128 bookmarks) to the openai/skills repo as a practical Codex workflow, and the repo itself says .system skills ship automatically while curated and experimental skills can be installed with $skill-installer. That makes skills less like prompt snippets and more like distributable capability packs.
@akshay_pachaar reported (33 likes, 12 replies, 2,948 views, 59 bookmarks) AI Engineer Coach as an open-source dashboard for local session logs, and the repo is unusually explicit about what it measures: practice scores, weekly trends, AI-generated code volume, anti-pattern detection, skill discovery, and context-readiness checks. That is a distinct builder pattern from last week because it instruments the operator, not just the model.
@thdxr showed (90 likes, 21 replies, 16,907 views) a shared OpenCode environment, while @zarazhangrui showed (7 likes, 3 replies, 391 views) a chat bridge for Claude Code. Both projects wrap an existing agent with another operating surface instead of replacing the underlying model.
@Suryanshti777 shared (13 likes, 8 replies, 348 views) CodeGraph with repo benchmarks claiming median savings of 35% cheaper, 59% fewer tokens, 49% faster, and 70% fewer tool calls, while @BeauJohnson89 pointed (29 views) to codex-shim for routing BYOK models into Codex Desktop. Repeated builder pattern: today's projects mostly sit around the agent - indexing, routing, observing, or sharing it - rather than inventing a brand-new assistant.
6. New and Notable¶
Codex skills became the day's highest-signal practical workflow¶
@jxnlco said (186 likes, 16 replies, 6,824 views, 128 bookmarks) Codex users should ask Codex about the openai/skills repo itself. That mattered because it reframed skills as a working installer path with curated and experimental packs, not as an abstract future feature.
Microsoft opened the tooling layer around AI-coding behavior¶
@akshay_pachaar reported (33 likes, 12 replies, 2,948 views, 59 bookmarks) AI Engineer Coach as an open-source workflow-audit tool, and the repo says it reads local session logs, flags 45 anti-patterns, and discovers reusable skills. That is notable because it treats prompt hygiene and context management as something measurable.
Copilot for Eclipse moved more of GitHub's agent stack into public view¶
@GHchangelog announced (27 likes, 1,522 views) Copilot for Eclipse is now open source. The linked changelog and repo say the public code includes chat, code completion, agent mode, MCP integration, skills, BYOK, custom agents, isolated subagents, and plan-agent behavior.
The Claude and Codex access story widened beyond normal pricing complaints¶
@Akasheth_ linked (24 likes, 15 replies, 196 views) the ChinaTalk article on cheap Claude tokens in China, and the article says API proxies can bring prices down to around 10% of official Anthropic pricing while creating a broader transfer-station economy. That was notable because it moved the conversation from plan frustration into public evidence of a parallel access market.
Codex limit pain escalated into a visible status-page incident¶
@weswinder showed (68 likes, 10 replies, 3,360 views) that OpenAI had posted a Codex rate-limit incident, and later screenshots in the feed showed the same incident moving to monitoring. The notable part was not only the outage itself; it was that multiple users had already been treating the behavior as product degradation before the status page caught up.
7. Where the Opportunities Are¶
[+++] Usage and spend control inside the agent workflow β Evidence came from both sides of the market: Gemini users hitting day limits after a few prompts, Codex users watching rate-limit incidents, TokenBar pitching live resets and credits in the menu bar, and AI Engineer Coach turning local logs into consumption and anti-pattern views. This is strong because it spans product complaints, public incidents, and builder responses.
[+++] Shared context and team control planes β Gangprompt's shared OpenCode server, replies asking for meeting-note ingestion and browser access in Codex, and the Lark bridge for Claude Code all point to the same opening: teams want one promptable environment with durable memory, controlled access, and remote surfaces. This is strong because the builds and the requests align closely.
[++] Semantic context layers for large repositories β CodeGraph is the clearest evidence here: public repo benchmarks, broad harness support, and replies agreeing that most agent time is spent hunting context. This is moderate because the benefit is obvious on bigger repos, but reply threads still worry about setup and maintenance costs.
[++] Official routing and model choice without hacks β codex-shim exists because users want Codex Desktop to expose BYOK and passthrough models as first-class options, while Copilot for Eclipse already documents BYOK and extensible agent surfaces. This is moderate because the need is concrete, but today's strongest implementation is still an early local shim.
[+] Compliant regional access alternatives β The ChinaTalk proxy-economy article and the Europe access request under the skills thread both show unmet demand for official access paths that do not force people into resellers or unsupported workarounds. The signal is emerging because the need is real, but today's evidence is stronger on pain than on credible solutions.
8. Takeaways¶
- Packaging workflow knowledge outperformed generic prompting discourse. @jxnlco showed (186 likes, 16 replies, 6,824 views, 128 bookmarks) a practical Codex skill-installation path, while @akshay_pachaar reported (33 likes, 12 replies, 2,948 views, 59 bookmarks) a tool for auditing how people actually use coding agents.
- Usage controls and reset timing are now product-critical. @TimJayas said (75 likes, 11 replies, 3,674 views) Gemini limits felt unrealistic after two questions, and @weswinder showed (68 likes, 10 replies, 3,360 views) that Codex limit pain had escalated to a public incident.
- Builders are still adding layers around the agent rather than replacing it. @thdxr showed (90 likes, 21 replies, 16,907 views), @Suryanshti777 shared (13 likes, 8 replies, 348 views), and @zarazhangrui introduced (7 likes, 3 replies, 391 views) team, search, and messaging layers on top of existing harnesses.
- Antigravity remained split between performance praise and product frustration. @haider1 said (29 likes, 10 replies, 2,043 views) it beat Opus 4.7 on broken-solution cases, while @TimJayas said (75 likes, 11 replies, 3,674 views) the same stack could burn through a day of quota almost immediately.
- Access friction is spawning parallel infrastructure. @Akasheth_ linked (24 likes, 15 replies, 196 views) public reporting on cheap Claude-token transfer stations in China, and @BeauJohnson89 argued (29 views) for a local Codex shim so users can route their own models through the official interface.