Most AI coding workflows don’t fail because the model can’t write code.
They fail because a single vague prompt balloons into an all-in-one session:
requirements, exploration, implementation, fixes, and review
fighting for space in the same context window.
mattpocock/skills is a repository of agent skills that turns that chaos into a repeatable engineering loop:
https://t.co/xsnq5JahaE
The loop:
grill-with-docs → to-spec → to-tickets → implement → code-review
This isn’t “more autonomy”.
It’s tighter boundaries.
/1
Reminder: an API key is just a credit card that speaks HTTP.
If a hosted dashboard stores it for you, their breach becomes your invoice. Rotate keys, set hard spend caps, keep the boring stuff local.
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
Hey builders 👋
Let’s connect if you’re building:
📦 SaaS
🚀 Open source
🤖 AI agents
🌐 Web apps
📱 Mobile apps
⚙️ Internal tools
📈 Marketing tools
💡 A side project
Drop what you’re working on below 👇
🚨 Three new OpenAI model IDs just showed up.
No docs. No pricing. And anyone who tries them gets a 503.
- aurora-blue-alpha
- aurora-green-alpha
- auror-red-alpha
Could be early image-generation models
The final boundary is review.
Pocock’s workflow spins up separate background sub-agents after implementation: one audits the diff against the spec, another hunts for standards violations and anti-patterns.
Because the agent that made the change is wired to justify it in the same frame. A fresh reviewer still catches what slips through: the missing requirement, the untested branch, the “quick fix” that quietly widens the scope.
The loop is intentionally dull:
clarify → specify → slice → implement → review
That dullness is the point. It’s where reliability comes from.
Video source:
https://t.co/4yz0vWIjsp
@mattpocockuk
Most AI coding workflows don’t fail because the model can’t write code.
They fail because a single vague prompt balloons into an all-in-one session:
requirements, exploration, implementation, fixes, and review
fighting for space in the same context window.
mattpocock/skills is a repository of agent skills that turns that chaos into a repeatable engineering loop:
https://t.co/xsnq5JahaE
The loop:
grill-with-docs → to-spec → to-tickets → implement → code-review
This isn’t “more autonomy”.
It’s tighter boundaries.
/1
Implementation starts in a clean session:
`/clear`
`/implement tickets. md`
The agent tackles one defined slice: delete obsolete commands, rewire imports, update what needs updating, then run typechecks and tests.
The temptation is to dump every ticket into one heroic prompt.
That’s usually how the session turns into a landfill of stale assumptions.
A bigger context window isn’t the same as better working memory.
/6
Hey builders 👋
Let’s connect if you’re building:
📦 SaaS
🚀 Open source
🤖 AI agents
🌐 Web apps
📱 Mobile apps
⚙️ Internal tools
📈 Marketing tools
💡 A side project
Drop what you’re working on below 👇
@itzSaasified looks like a solid solution to the problem.
I'm building feedbakery for customer feedbacks collection and management.
Let's stay connected