San Francisco Bay Area
- Tweets 507
- Following 333
- Followers 89
- Likes 427
A tweet mocking GPT's habit of piling on disclaimers prompted me to revisit my global agent rules. I wanted direct, useful answers without having to ask for them every time.
I had my agent raise the bar for caveats and carry the same principle into code and tests. I added multilingual behavior checks and made running them part of future rule changes. The checks target the same defensive habit even when the wording changes.
Here's a prompt you can use in your own setup:
“Review and improve my persistent agent instructions to reduce AI slop, excessive hedging, and defensive rhetoric. Make direct, useful answers the default across languages and tasks. Set a high bar for qualifications: include one only when a concrete fact in the task makes omission likely to cause a material misunderstanding or wrong action. Remove unsolicited rebuttals, repeated limitations, formulaic contrasts, and disclaimers that merely sound cautious, including paraphrases. Preserve factual accuracy, honest status reporting, and faithful quotations. Apply the same proportionality to code and tests: extra checks, fallbacks, and complexity need a concrete requirement or a failure supported by the actual task. Make these expectations durable in my existing setup. Assess them against varied scenarios, including cases where a qualification is necessary, and make relevant future rule changes trigger appropriate checks. Keep the guidance portable across models and usable without repeated reminders.”
ChatGPT 已经彻底失去了说人话的能力：
你说今天天气真好。它说：确实，但今天晴朗不代表明天晴朗，更不足以证明当地气候正在改善。
大哥，我是想出门散步，不是让你给 ICLR 审稿。
ChatGPT 三句话里有两句在叠甲：不能推出、并不代表、尚不足以确认、没有证据表明……
于是我们愤怒骂他，不要再防御性回复了，少说废话。
它诚恳道歉，然后补一句：但个别回答出现防御性表达，并不足以证明模型整体存在这个问题。
ChatGPT 现在已经习惯了给每句话都套了十层 try/except...
1
114
Well, it appears that switching the gpt-6-astra reasoning effort mid-conversation from medium to xhigh still incur a cache miss. Anyone else has the same experience?
Day 19/ Small updates. Calls connect much faster for dots, changing reasoning in codex & work mid conversation does not invalidate cache anymore and a we've also updated the codex cli.
25
Reading this discussion about unnecessary subagents prompted me to revisit how I ask AI agents to work. I had been encouraging delegation whenever it might help. Looking at that, I realized how easily "might help" could become a reason to split almost anything.
I've tightened that default. I want the lead agent to stay involved in doing the work, with delegation serving a clear purpose.
For me, the useful question is what a separate agent adds. Independent research can benefit from parallel exploration. A fresh reviewer can catch something the original agent missed. But when several steps depend on the same evolving context, handing them around can mean more explanation, lost details, and coordination.
The Anthropic article behind this discussion gave me a practical way to think about it: keep useful knowledge available through skills and tools, and test whether an extra agent improves the whole task.
I haven't measured the impact of my changes yet. My aim is clearer ownership, fewer unnecessary handoffs, and better use of the context already available.
If you're building a similar workflow, I'd start with one task and compare the results before adding another agent. Keep the delegation that earns its place in practice.
x.com/pvncher/status/2114109…
1
1
133
GPT sensei estimates that Jacob Coxon gave up ~$3M equities when resigning from Anthropic:
chatgpt.com/share/6aa64a03-7…
56
I love the idea of comparing paid GPT plans to long Call options: you exchange $20 premium for the right to "exercise" your Plus plan into a Pro $200 plan with accumulating banked resets, and hope for delayed Astra rollout to your account. I created two new Plus accounts, each of which has accumulated 2 resets as of now, and look for the right time to "exercise" my options.
好多人没看明白这个逻辑，这就是买个期权。理论上付费用户都能获得可存储的重置卡，每天一张（价值约 50-80 $），成本 Go 8 $/月，Plus 20 $/月。行权的时候花 200 $ 升级成 Pro，再使用重置卡就行了。相当于用 8$ 成本锁定了 n x 50-80$ 的价值，n 为 Astra 滚动到你的天数。
767
pluoki retweeted
BREAKING: It’s been revealed NVIDIA's $12,930,300,000.00 acquisition of Hugging Face contains an easter egg.
The first 6 numbers of the acquisition price represent the decimal conversion of Unicode character U+1F917.
The 🤗 emoji.
115
306
6,118
2,057,764
GPT-6 Astra prompting guidance & best practices are released. Happy to see that its prompting advice on intent following, subagent delegations, plain prose writing, and proportional testing closely resembles what have already been established in my agent rules suite.
developers.openai.com/api/do…
181
Version 26.727.51351 (build 6119) of Unified ChatGPT/Codex, released on 2026-07-31, is the last confirmed Codex build working on macOS Monterey 12, whereas the next release 26.730.61309 as well as all subsequent releases breaks on Monterey and requires macOS 13 or above. I figured this out thru my own testing (thanks to Time Machine), and confirmed via other reports (see GitHub openai/codex/issues/37019 & 37070).
This is unfortunate, because many old macs are stuck with Monterey. To make things worse, ChatGPT/Codex auto-upgrades itself beyond 26.730.61309 on Monterey, entering itself into a state of being unable to start up. See screenshot no.1 for what happened to my 2016 x64 MBP. I am sure many others had a similar experience.
The long term solution seems using Codex CLI on Monterey and forgetting about the desktop app.
In the short term, however, if you are like me, who want to revert to the last working release 26.727.51351 on your Monterey macs and pin its version for as long as possible, you may download this build from the CDN link posted in the codex issue below, and extract it into your /Applications to overwrite the incompatible build:
github.com/openai/codex/issu…
and then prompt your agent for a durable version pin - see the gist below for my prompt:
gist.github.com/pluo/406ae8a…
After prompt execution, you may quit and restart ChatGPT/Codex, and visually confirm that the "Check for Updates..." item disappears from the menu, as shown in screenshot no.2.
1
292
Thanks Tibo for confirming the context window setup. I just prompted codex to "update my config to use a 500K-token context window with 10% compaction headroom."
I considered going up to 1M tokens, but 500K feels like a better starting point for my workflow. My sessions often include large instruction sets, tool outputs, source material, and subagent results. A larger window can help preserve context longer, but it increases cost and latency. Bigger is not automatically better.
The 10% buffer is intentional. A smaller buffer like 5% could be consumed quickly by a large tool response or agent handoff. The extra headroom improves compaction reliability.
My approach is to pick the smallest window that fits the work, with enough space for compaction to operate safely. I prefer tuning based on real usage rather than defaulting to maximum context sizes.
I’ll monitor `/status`, compaction behavior, latency, and whether long-running tasks retain the right details. If 500K turns out to be misaligned, I’ll adjust it.
Curious how others are sizing theirs.
x.com/thsottiaux/status/2089…
Here is how to enable a 1M-token context window in Codex for GPT-5.6 Sol.
Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.
A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. And GPT-5.6 Sol, for example, has a documented 1,050,000-token window.
Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:
```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom. Restart Codex client and start a new session after saving.
To try the configuration for a single CLI session without changing your defaults:
```
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000
```
Have fun, but also know that we tuned the default carefully!
3
2
317
Details I had to sort out before changing this setting:
The 1M figure is not exclusive to GPT-5.6 Sol. OpenAI documents a 1,050,000-token limit for Sol, Terra, and Luna on Amazon Bedrock. However, my Codex 0.147 catalog lists all three at 272,000 tokens, with 95% available before its compaction margin. The model's maximum and Codex's configured working window are separate numbers.
model_context_window tells Codex how much context to budget. It cannot increase a provider's hard model limit. Setting 500K is an override of Codex's current 272K catalog value, not a guaranteed unlock on every product surface. I am treating it as an experiment and checking /status, latency, usage, and compaction behavior.
model_auto_compact_token_limit controls when Codex starts summarizing older history. At 450K, a 500K window keeps 50K tokens, or 10%, in reserve.
Subagents inherit the top-level configuration by default. A role-specific TOML config can override context settings, while an explicit spawn override can select another model. Each subagent still has its own separate context; agents do not pool one shared 500K window.
So Tibo's idea uses supported keys. The judgment call is how far beyond Codex's current catalog value you want to go, and whether your endpoint honors it reliably.
x.com/thsottiaux/status/2076…
(1) is not correct, it is not due to 2X charging after 272k context, we don't charge for longer context on the subscription for GPT 5.6 Sol as we control all the settings. It is due to something else, which I will attempt to explain below.
The overall trajectory length, which is the total some of all context windows across compactions, changes little based on the reasoning effort. Similarly the quality of the overall output is similar across context lengths above 272k. The benefits of higher context lengths are mostly overall speed (as you don't wait for compaction), ability to deal with humongously large input and potentially cost if the system is well tuned and you hit your cache perfectly.
The actual reason is the what you can see depicted in the chart below, which is the difference in the orange line and the blue line. It is caused by overall cost of cache reads going up with the size of the context being shuffled back and forth between toolcalls. The sweet spot in terms of cost is therefore not necessarily to use the maximum possible context length.
What we're working on is tuning the system differently so that we can go back higher without it resulting in higher usage being charged. Hope this clarifies a bit.
1
1
218
I reduced my Codex context default from 500K/450K to 320K/270K. The 500K setup worked, but consumes quota sooner than I expected, and ordinary tasks rarely need that much room.
model_auto_compact_token_limit, which I now set to 270k, is when Codex starts summarizing older history. I want this below 272K, the threshold for a different pricing tier listed on the public API pricing page. Codex usage does not follow that pricing directly, but its internal pricing and token model are opaque, so compacting before the threshold is a prudent safeguard.
model_context_window is the budget I ask Codex to use. Mine is now 320K tokens, which becomes a 304K usable/effective window after Codex’s 5% safety margin. That leaves about 34K tokens between compaction and the effective limit. It gives large tool responses or subagent handoffs ample room while remaining much lighter than 500K.
This seems like a better everyday default for coding, research, long instructions, tool output, and subagents. I can still use a larger setting for unusually long investigations. I’ll continue monitoring /status, compaction timing, latency, and whether summaries preserve the details I need.
331
I asked Codex to update a scheduled task. The `automation.toml` changed, but the Scheduled tab later showed an old prompt. After comparing them, I filed a GitHub issue and sent in-app feedback. Reproducible reports beat guessing. github.com/openai/codex/issu…
2
87
Could Codex expose a per-run usage ledger?
I use Codex with large contexts, auto-compaction, different reasoning levels, and subagents. When the usage meter drops, I cannot tell how much came from the selected model, context processing, reasoning, reviews, or child agents.
That breakdown would help users diagnose unexpected usage and help OpenAI distinguish normal heavy workflows from unsupported sharing or abuse.
x.com/thsottiaux/status/2090…
We've investigated a few messages about codex usage limits being different. That's not something we change without engaging the community and being transparent.
What we did see is that when talking to affected users many were using sub2api. Converting a subscription into api traffic to then re-serve or share across many users is not something we support and this type of usage gets flagged by our fraud-prevention systems.
You are completely fine if you use your subscription through Sign in With ChatGPT, either through the official clients or through one of the many OSS clients (Pi, OpenCode, ...) that support signing in with your account and using your included usage.
359
I impulsively started considering an Apple Vision Pro to replace my monitor, then looked closely at its sales, durability, ownership costs, and future roadmap. I am passing for now. It is still too expensive and too heavy. The headset cannot run without a wired external battery pack that must stay attached and be carried around, which feels awkward. Apple offers no trade-in value, resale looks weak, and I am not optimistic about Apple's roadmap for it. Ownership is hard to justify over the lifecycle that matters to me.
1
72
I recently noticed the same writing pattern across many online posts I was reading. Their subjects and intended readers varied, yet the pages often looked strikingly similar. Many paragraphs contained only one sentence, and some contained only a short phrase. Large areas of white space separated ideas that I would normally expect to see developed together.
That format caught my attention because it looked fragmented to me. I could understand each sentence, but I often had to reconstruct the relationship between one block and the next. At the same time, I kept seeing this style across online posts and articles. I wanted to know whether I was looking at a recognized writing technique, a habit created by mobile screens, or a style designed for a particular kind of reader.
I opened a Codex session and gave it a narrow assignment. I asked it to use the built-in browser to inspect how a sample of those posts was structured. I specifically told it to ignore the main ideas of the posts and study how their paragraphs were composed. Then I kept pushing the analysis further: What is this style called? Why has it become so common online? What reading situation does it serve? When does it help, when does it weaken the writing, and what would be a better default for my own public work?
The inspection confirmed that the visual pattern was intentional. Across the sample, most content blocks contained one sentence or a sentence-like unit. The neutral name is single-sentence paragraphing, while descriptions such as short-paragraph web writing and staccato prose refer to its broader design and rhythm. Its target is less a demographic than a reading situation: someone scanning a phone, moving through a feed, getting interrupted, and deciding whether to give the next few seconds of attention. That reader can still be highly technical or deeply informed.
The analysis also helped me separate a useful technique from a mechanical habit. A one-sentence paragraph can work when it marks a real turn, isolates a question, emphasizes a verified result, or gives the reader a needed pause. It becomes weaker when every sentence gets the same dramatic space. Too many isolated sentences can hide the connection between a claim, its evidence, and its consequence. The page becomes easier to scan while the reasoning becomes harder to follow.
I decided to use compact, developed paragraphs as my default for durable public writing. I want most paragraphs to carry one controlling idea and enough explanation, evidence, interpretation, or consequence to complete that thought. Two to four sentences is a useful starting point for me, although it is not a quota. This structure works for the broad readers I want to reach, including experienced engineers and builders, because it gives me enough room to show how I reached a conclusion, what I verified, and why the result matters.
I will still use shorter blocks for platform-native posts, and I will keep an occasional one-sentence paragraph when the pause has earned its place. I chose this balance because I want my writing to remain easy to read on a screen without turning every idea into an isolated performance. White space should help readers see the structure. It should not carry reasoning that I have left unwritten.
I did not want this conclusion to remain a preference that my agents might forget. I asked Codex to turn it into a one-shot implementation prompt, check that the new wording preserved my existing rules, update my version-controlled public-writing guidance, test the behavior in fresh contexts, and commit the verified change. The rules now give long-form writing the compact, developed default, preserve flexibility for short social formats, treat four consecutive one-sentence paragraphs as a review trigger, preserve user and platform overrides, and let bilingual paragraph pairs differ in sentence count while keeping their meaning, order, and tone aligned.
If this approach sounds useful, you do not need to copy my rules word for word. Give this post to your writing or coding agent and say: "Read this post, compare its paragraphing approach with my existing writing style and rules, preserve my voice and current exceptions, and adapt only the parts that improve how I communicate with my readers." Then review the proposed change before letting the agent update anything. Your best default may differ from mine, but it should be a decision you made rather than a format you absorbed from the feed.
1
73
I have a similar ad hoc prompt, which I typically use as a follow-up (not a skill):
“
Stress-test this decision.
Reframe it as a neutral question and state any material assumptions. Identify the realistic options, including the status quo or a third option when relevant. Steelman my current view and the strongest alternative using the best available evidence. Separate facts, inferences, and value judgments. Identify the evidence, variables, or thresholds most likely to reverse the recommendation.
Ask at most three questions in one turn only if their answers could materially change the recommendation; otherwise proceed. Then give me a clear conditional recommendation, the key uncertainty, and the next reversible action.
My situation: [paste it here]
”
用“钢人论证法”做了一个底层很有趣的Prompt，可能会让你让AI帮你做决策的时候效果更好更加全面。
先别急着回答，也别默认我已经把问题想清楚。请先对这个问题做一次详细思考的“双向钢人论证”：
1.用最完整、最有力的方式，重述我真正想解决的问题。
2.使用钢人论证法分别给出支持我当前想法、以及反对它的最强论证。
3.找出双方真正的分歧，以及最可能改变结论的关键变量。
4.只问我一个最关键的问题。等我回答后，再给出明确判断、理由和下一步行动。
我的问题是：
[粘贴你的问题]
如果对这个Prompt的底层感兴趣，可以看看这个更详细的文章～
61