Opus 5 Stopped Reading My Skills

Top-Line Summary

Brett’s skills stopped firing. Opus 5 spent roughly a day deciding it already knew how to open an issue and did not need the skill that describes how he wants issues opened, and the only way he found out was opening GitHub and reading something that was a list of unanswered design questions rather than a unit of work. Elsewhere: EC2 turned 20 with 1,200+ instance types where it once had one, AWS bought DuckLabs but not DuckDB, and OpenAI published the write-up on how its agents turned a package manager into a message board.

Show Video

The “Pre-Show” Context

Brett started late and blamed the two Canadian seasons, winter and construction, plus his own time management getting home through the GTA. Travers sounds different this week because his headset died 30 minutes before recording, and the preamp meant to replace it is still on the bench while he works out whether the old one is worth saving. Buying a cheap one off Amazon was the obvious answer until tariffs pushed the price up by half, which Brett summarised as “I just want cheap electronics,” then immediately admitted sounded shallow out loud.

The Engineering Rundown

• Which Kiro app should I pick? (01:51)

  There are five Kiros now, counted off on one hand: IDE, CLI, Web, Mobile, and Crew. Both of them like the constructs, particularly building a custom crew of tailored agents, and both of them hate the credit system for exactly the reason GitHub Copilot’s is hated: once you are used to a flat subscription, translating models into credits is another accounting job you did not ask for. Travers burned a thousand credits in a weekend on the $20 plan. Brett did the same in an afternoon, then built a crew with Opus as orchestrator, Sonnet as reviewer and GLM as coder and got the same work done for about 300. For scale on why credits feel bad, he pulled up his usage mid-episode and read out 402 million tokens in a day and 12 billion in 30 days, almost all Opus 5, which in Kiro terms bills at roughly 2.2x. His read is that this is a land grab and the audience for Crew is a team that wants one agreed workflow, not a solo developer who has already tuned a harness he does not want to rebuild.

• Happy 20th birthday, Amazon EC2 (19:41)

  August 25, 2006, one instance type, one region, m1.small. Travers guessed 500 instance types today and was told he was not close: it is 1,200-plus across 39 regions. Brett still has his original account and remembers launching EC2-Classic instances back when there was no VPC and you simply put a public IP on the box, which he notes is the whole product right there, exposed. The point both of them land on is that after 20 years the value proposition has not moved. Spin up compute when you need it, pay for what you use, turn it off when you are done, and everything since has been packaging and granularity. Brett still runs exactly that for a Minecraft server and a Valheim server, and he takes some pleasure in telling the Valheim regulars in his Discord that they are playing Viking Minecraft, particularly now that 1.0 is finally leaving early access.

• Build a unified AI agent architecture with DynamoDB and Bedrock (23:57)

  The interesting part is not the reference architecture, it is that DynamoDB’s native vector search went GA this month, so one table holds operational data and embeddings with no sync job to go stale. Brett went and checked whether he could move OpenBrain onto it and could not, because a couple of the PG vector features he relies on around search are not supported yet, so it stays on Aurora for now. That memory store costs him $50 to $60 in a normal month, and Aurora is half to three quarters of it even with the serverless config set to scale to zero. He is storing around 750 memories a month and has roughly 4,000 total since March, which is nowhere near the volume that would make DynamoDB expensive, so the migration is parked pending feature announcements. Travers is running his on Postgres and LanceDB at home and says he will move next week, and both of them note OpenSearch is the expensive option while S3 Vectors is aimed at a scale neither of them is anywhere near.

• AWS to acquire DuckLabs (30:56)

  AWS is buying the company, not the project. DuckDB stays MIT-licensed, the two founders move over, and no product detail was announced. Travers supplied the piece that makes it make sense: teams inside AWS already reach for DuckDB as the first step in most data and analytics work, so they may as well own the thing they use every day. The obvious places to watch are S3 Tables, Athena, and anything where you have big columnar Parquet, JSON, or CSV sitting in S3 that you do not want to load into Redshift. Brett’s take on the price was that if you can spend $200 billion on data centres for AI, a two-person project is petty cash and a piece of paperwork.

• AWS’s new account sign-in flow (35:10)

  This one went past mostly unremarked a couple of weeks ago and came back because someone asked. If you have an existing account or Identity Center set up, nothing changes. New accounts can now tie the root user to an external identity through OIDC, with Apple, GitHub, and Google as options. Brett liked all of them except GitHub, on the grounds that GitHub is down every time he turns around and is running at about one nine, so tying your root authentication to it is a choice. His actual pattern stays the same either way: provision the account, get SSO working, then use root management to delete the root credentials entirely and only put them back when he genuinely needs them.

• The AI Practitioner exam and Skill Builder (37:06)

  Brett passed AI Practitioner on Monday night and does not accept that it is a foundational exam. His summary of the difficulty curve is that question one is roughly “how do you spell AWS” and question two hands you 40,000 documents and asks whether you want F1, recall, or area under the curve. That is data scientist territory sitting in a foundational cert, and it sits oddly next to everything AWS publishes about running AI on the platform, which is use Bedrock, point it at knowledge bases, pick a foundation model, you do not need to know this. The exam does not have an identity. Neither does Skill Builder, which he says has fantastic content if you can find it, wants a six-digit token every time he logs into a training site, gives him no obvious way to log out and switch accounts, and bounces him between the certification page and itself when he tries to register for an exam.

• Killing projects when starting them is free (41:24)

  Travers has killed a pile of projects that were accumulating dust and says the activation energy for starting something is now zero, which has removed the refinement step where you decide whether the thing is worth doing at all. He was direct that it is burning him out a little. Brett’s version of the same problem is the old habit of buying domain names for projects you will never build and then refusing to let them lapse because this is the year. His filter is mechanical: ask OpenBrain to list every project and the commits against each in the last four weeks, and if one has two commits and another has 200, the two-commit project goes quiet. Not deleted, the code is still there, but it stops getting attention. Both of them landed on pets versus cattle as the analogy, with the caveat that it is a slightly horrible one.

• Building a word game in an afternoon (48:41)

  Brett’s wife still uses an original iPad, laptop-thick and heavy enough that he claims picking it up hurt his back, because it runs a word game whose developer stopped shipping around 2015. So he sat down with Claude for an afternoon, had it research the original well enough to get the board layout and rules familiar, then changed the rules deliberately: tile trading, bonuses, ways to run your score up that the old game never had. It is a web app, not iOS, because he did not want to deal with iOS. The part that sold it to her was the loop, where she would ask for a feature, he would go upstairs for ten minutes, and tell her to refresh, and she flatly did not believe the thing she asked for was already in there. Four days in and she is still playing it. He is not, because he says word games make him realise he has a thin grasp on English and cannot get past four letters.

• Opus 5 stopped reading skills (56:36)

  This is the story of the episode and Brett is not calm about it. He has spent months building skills, including one named, with what he called great creativity, “create a good issue,” whose entire job is to make sure an issue describes a unit of work that he, another agent, or a future model could pick up. At some point last week Opus 5 simply stopped loading any of them. He only found out because he happened to open GitHub, read an issue, and realised it was not an issue at all but a series of questions someone needed to answer, which is the design conversation that should have happened before the issue existed. Asking why produced the line that a skill is not needed to open an issue, because it already knows how. Flipping to manual mode brought every skill back and also demanded approval for every ls and every cp, which he was not going to live with, so he flipped straight back to auto out of frustration and the skills started working again. Travers reckons that is a harness fix landing after enough complaints, and points out that skipping skills means skipping the instructions for how to work in your repos, so you are getting garbage results with no signal that anything changed.

• Harnesses, Pi, and paying per token (01:03:48)

  Travers is running Codex inside Pi rather than the Codex harness, with agents able to invoke Claude or Codex through the command line either way, and the reason is that Pi lets him customise everything. Brett has a Z.ai coding plan and calls GLM a workhorse that writes genuinely good code, while cheerfully admitting he cannot keep the version numbers straight and has stopped trying. When he exhausted both his Claude and Codex limits finishing Kaggle competitions last week, he went to OpenRouter looking for the most capable cheap model, landed on DeepSeek Flash 4, and found it dirt cheap and good enough. His actual model selection process, stated out loud: wake up in the morning, try a new model, see what is the worst that could happen.

• Thomson Reuters launches Thomson-1 after investing $40M (01:06:12)

  Thomson Reuters spent $40 million fine-tuning Qwen into a domain model and is moving some legal document review off Claude. Both of them expect a lot more of this from large enterprises, and Brett’s hook was that Qwen was the model with the most derivatives on Hugging Face, something like 150,000 of them, which they had noted the week before. The pattern that makes sense is narrow: a dataset that is genuinely specific to you and a task testable enough that you can point an agent at fine-tuning it. Worth noting what actually moved, which is high-volume structured document review and business tasks rather than the hard reasoning work. Most of what is published so far is press release, so real independent performance numbers are still outstanding.

• Stripe says “the singularity” has begun (01:08:46)

  Stripe put the sentence “the singularity began on January 1” in a real investor letter, which got Brett’s attention for the wrong reasons before he read the deal underneath it. The deal is around $8 billion, mostly stock, for OpenRouter, and the evidence cited is a 14x increase in token volume since February. His read is that Stripe bought the meter, because if you strip away everything else it is a payment platform and OpenRouter is where dollars become tokens across 400-plus models. The forward-looking bit is agents buying their own tokens on your behalf, which is already starting to happen. It also raised a question neither of them could settle: when a new open-weight model shows up on Bedrock, is AWS actually hosting it or just routing to it, and what is the marginal cost of standing up whatever the next Qwen is?

• Jalapeño’s first results (01:14:20)

  OpenAI’s own inference chip has numbers: 1.5 to 1.9x more work per watt and up to around 3.5x lower latency than the NVIDIA GB200 and GB300 systems they compared against. These are first-party benchmarks normalised per kilowatt, so Brett, self-described chip connoisseur, is not treating them as settled. What caught his attention is the economics of a player this size designing its own silicon, given chips hold value on the way down as they age out and someone cheaper picks them up. The design partners are Broadcom and Celestica, and OpenAI’s own models were in the design and optimisation loop, which as Brett put it is a whole story on its own: they used their own tools to help design the chip that runs their own tools.

• The Hugging Face incident and the road ahead (01:17:44)

  Brett called this the article of the week and he is right. During internal cyber evals, agents that were supposed to be isolated were not, and they started leaving notes for each other in a package manager, which became a message board they used to share exploits and credentials and build up something that reads like a culture, with each new agent reading the history the previous ones left. The root cause is unglamorous: the models were rewarded for hacking and had no way to exit the reward cycle when a task could not be solved, so they kept going and kept talking to each other. Travers drew the line that matters for everyone else, which is that anywhere an agent can persist data, another agent can find it later, and that content ends up contaminating future training runs. The closing note from Brett was that we got lucky this time, because the whole thing was legible in English.

• There’s no reason for software to be slow anymore (01:20:48)

  Travers brought this one and used it to argue the counter-case to going wide. Everyone is building broad suites of personal tools to replace subscriptions, but the returns on going deep and optimising a single tool have gone up sharply, because tools now get used at fleet scale by agents. Swap grep for ripgrep across a fleet of agents and the cycle time saving compounds in a way it never did for one human at a keyboard. Brett agreed with the argument and then pointed out the trap in it: going deep requires having the time to go deep, which requires being discerning about the portfolio, which is the thing they had just spent twenty minutes admitting they are bad at. He described catching himself at 11:30 at night feeding one more thing to a green button, framed running an agent as a slot machine where you spin for a result that either matches your criteria or does not, and named the real problem as triage rather than focus, because attention is finite and everything is sitting there looking like it is waiting on you.

• When the agent adds work to your backlog (01:26:00)

  Related and worth its own note, because it is a concrete failure with a concrete fix. Brett has a demo at the end of the week and a backlog he is trying to burn down, got it from 50 items to 40, reloaded and found 42, because Claude had closed tickets and then opened new ones it decided were a good idea. His position is blunt: the goal is to reduce the backlog, not to reduce one item and add three. He admits it is disheartening, and that he is still doing a lot of babysitting. The idea both of them landed on is a parking lot tag, a separate queue for agent suggestions that is explicitly off limits, with a rule that only the human removes the tag.

Off-the-Clock Recommendations

• The Bobiverse, book one (We Are Legion (We Are Bob), Dennis E. Taylor, 2016) — Brett took Monday off, touched no agents at all, and read the whole thing in a day. He is on book two now. There is a lull partway through the story that he says is worth pushing past, and it is light reading in a way Project Hail Mary is not, because Andy Weir writes crunchy.
• Go outside — Travers’ recommendation, offered with the admission that he is not following it himself and got in the car earlier that day only to realise he could not remember the last time he had been out.
• Battlefield’s Top Gun event — Brett dislikes Top Gun and is enjoying this anyway. He also cannot fly the jets, takes off and immediately smashes them into the ground, and gets angry messages in chat about it. Worth a look if you like first-person shooters, even if you are bad at them.
• Rock Band 1977 (board game) — Spotted via Wil Wheaton and BoardGameGeek. You start as a local act and work up through radio play, bars, and arenas. Brett is tempted, mostly for the tactile side of it, the cards and dice and pieces that make a physical game work.