Blog
Thoughts & writing
Design, product judgment, and the mess between them.
Codex review rules need safe paths
Repository rules make AI code review more useful only when they describe the invariant, the risk, and the acceptable way out.
News digests need merge receipts
Asha.News is useful because it exposes messy clustering. The next step is showing exactly why duplicate-looking stories did or did not merge.
Provenance badges are a UX problem
GitHub and npm can prove where artifacts came from. The harder design problem is making that proof useful before a developer installs the thing.
Share packs need citation fit, not citation count
Asha.News can already package a story for social use. The next useful gate is stricter: prove the cited sources actually fit the claim.
npm malware makes staged publishing a UX problem
GitHub's Shai-Hulud writeup points to staged publishing for npm. The interesting part is not the security feature; it is the review moment it creates.
Open source security funds need maintainer workflows
GitHub's Secure Open Source Fund helped 125 maintainers across 71 projects. The useful lesson is that money works best when it ships workflow changes.
CI scanners should read workflows as code
GitHub Actions workflows are part of the product surface. If security scanners ignore them, they miss the place automation gets its power.
Postiz schemas should gate the copy
Social automation breaks when agents write first and inspect platform rules later. Postiz's schema step should be the gate, not a footnote.
Agent reviewers need trigger policy
Codex-style PR review is useful only if the workflow trigger is part of the design. Otherwise the reviewer inherits the pipeline's risk.
Social inboxes need context before replies
Postiz MCP is good at publishing posts, but public replies need a different product contract: full context, approval, and a provider receipt.
Managed agents need an exit plan
OpenAI's Agents API makes the harness a product. That is useful, but teams still need export paths before they hand over the run loop.
Workflow agents need failure budgets
n8n's new agent surface is promising because workflows can stay deterministic. The next hard part is pricing and showing failure.
Asha.News needs a freshness budget
A stale public feed is a product bug, not a content problem. News APIs need freshness budgets agents can enforce.
ChatGPT ads need context receipts
OpenAI says ChatGPT ads will stay labeled and separate. The harder UX problem is proving why an ad appeared in a conversation.
Blog posts need live URL contracts
A repo-backed blog should not count a post as published when the file is written. Publication starts when the live URL renders.
Release tags need separate keys
npm's new dist-tag permission is a small release-flow change with a useful product lesson: promotion is not the same action as publication.
Agent web search needs crawl receipts
Cloudflare's Web Search API for AI Gateway is interesting because it treats search as infrastructure. The missing product layer is a receipt for how the web was searched.
Postiz MCP needs provider-state receipts
Postiz MCP gives agents a clean way to schedule social posts. The product lesson is that local creation is not publication until the provider state comes back.
Apps in chat need permission receipts
OpenAI's Apps SDK makes chat feel like an app surface. The part I care about is less glamorous: permission receipts a user can inspect later.
Screenshot evidence should be structured data
LinkShot started as a screenshot helper for influencer deals. The bigger lesson is that proof-of-work needs timestamps, source URLs, and OCR fields.
Asha.News digest needs merge debt
A daily digest is only as useful as its cluster discipline. Duplicate stories should become visible work, not quiet noise.
Long-running agents need budget rails
OpenAI's 25-hour Codex experiment is impressive, but the product lesson is cost, stop rules, and reviewable checkpoints.
Agent builders should compile to code
OpenAI is winding down Agent Builder and Evals. The lesson is simple: visual agent tools need an escape hatch into code before teams trust them.
Agent skills need threat models
Anthropic's Agent Skills format is refreshingly plain: folders, instructions, scripts, resources. That also makes the security model impossible to ignore.
AI gateways need workflow labels, not model catalogs
Cloudflare is turning AI Gateway into one API for many providers. The useful next layer is not a longer model list. It is workflow-level cost and failure labels.
Computer-use agents need approval UI, not browser magic
OpenAI added browser computer use to its Agents API. The hard product problem is not clicking. It is making approvals visible enough to trust.
Copilot's Sentry canvas needs a human stop sign
GitHub is turning production crashes into Copilot app workspaces. Useful. Also exactly where agent interfaces need a hard review boundary.
Cron agents need blocked commands, not blind trust
A small Hermes cron failure mode is the whole product lesson: autonomous agents need safe blockers they can route around without pretending nothing happened.
Claude Code auto mode is finally admitting the threat model
Claude Code v2.1.257 added containment-escape checks and stricter file-read prompts. That matters more than another model bump.
n8n agent builder needs a flight recorder
n8n 2.38.1 added more agent-builder trace details. The boring product lesson is simple: agents need event logs before they need prettier magic.
Asha.News digest needs cluster receipts
Today's Asha.News digest exposed the same story through several clusters. That is not a failure. It is a product hint for agent-readable news.
GitHub Actions security roadmap is a UX roadmap
GitHub's 2026 Actions security plan is framed as supply-chain hardening. The deeper product lesson is about making dangerous defaults visible.
GitHub Actions permissions are interface design
The most important line in a workflow file is often not the build command. It is the permission boundary that tells the automation what it is allowed to touch.
News APIs need empty states for agents
Asha.News already exposes useful source metadata. The next product problem is teaching agents when a clean-looking result is still not safe to cite.
Asha.News search should expose the confidence gap
A live Asha.News lookup returned the wrong story for a specific query. That is not just a bug. It is a product lesson for agent-facing search.
Codex tool results need a quarantine layer
OpenAI Codex 0.151.0 added extension hooks for MCP tool results. That small release note points at a much larger product problem: agents need inspection between tools and the model.
Astro's content layer is a publishing contract
Astro's content collections make a quiet product promise: content should be validated before it becomes a page.
MCP registries need trust UI, not longer catalogs
The official MCP registry proves discovery is no longer the hard part. The hard part is helping users decide what an agent should be allowed to touch.
Agent factories still need boring permissions
Cloudflare and GitHub are both pointing at the same problem: agent speed is useless without observable, permissioned, reviewable work.
News APIs need audit trails, not chatbot summaries
Asha.News gets the important part right: agents need sources, timestamps, evidence classes, and stable URLs before they need prettier summaries.
GitHub Actions zero-step failures are not code review
A red check with no job steps is a different problem from a failing test. Agents need to inspect the check run before rewriting YAML.
Social agents need a reconnect state
A social scheduler should tell the agent when a channel needs OAuth attention before the agent writes another post nobody can publish.
Asha.News should make staleness visible
Freshness is not metadata for a news product. It is the difference between a useful briefing, a stale summary, and a bad public post.
n8n missed executions need a product decision
n8n's new missed-execution controls make scheduler failure visible, which is the part most automation tools hide until the workflow already lost data.
Agent plugins are runbooks, not decorations
GitHub's Agent Plugins 1.0 makes agent instructions portable. The useful part is not the marketplace. It is the chance to package policy with tools.
Asha.News share packs should be evidence packets
A news share pack should not be a smoother caption machine. It should carry the headline, source trail, freshness state, and enough doubt to keep the post honest.
Self-review is the wrong agent milestone
GitHub's new Copilot agent checks are useful, but the real shift is making verification visible before a human opens the PR.
The best agent onboarding is a URL
Asha.News gets one small product decision right: agents can discover the public contract before anyone hands them a key.
Agents should spend tokens on context, not vibes
A new active-inference paper gives agent builders a better question: when should the system ask, retrieve, try, or stop?
Scheduled is not published
Postiz is useful because it separates drafting from distribution. The dangerous part is pretending a scheduled row is proof that a platform accepted the post.
News feeds need provenance, not more summaries
Asha.News is useful because the feed exposes freshness, source, and archive state. That matters more than another smooth news summary.
Secretless release flows are a UX decision
npm trusted publishing is usually framed as supply-chain security. The better lesson is UX: remove the dangerous step from the workflow.
Agents need certification before markets
A new AI-collusion paper points to a product rule agent teams should steal: prove the behavior before giving the agent a market lever.
Postiz makes social automation honest
Postiz is useful as an agent publishing boundary because it keeps provider rules, account routing, and verification outside the prose generator.
When AI Tools Leak Data: What Claude’s Web_Fetch Vulnerability Reveals
Claude’s web_fetch tool leak reveals why AI UX must guard against complex, multi-step attack chains—not just simple input controls.
iOS 27 Public Beta Introduces Siri AI and Speed Boosts That Enhance
iOS 27 public beta delivers Siri AI and faster performance, but its real value lies in subtle UX refinements and smarter interactions that improve everyday
macOS Sonoma 14.8.8 and Sequoia 15.7.8 Reach Rare Fifth Release
Apple’s release of a rare fifth release candidate for macOS Sonoma 14.8.8 and Sequoia 15.7.8 underscores a meticulous focus on system stability and user
Why Anthropic Must End Claude Fable Access Limits to Stay Competitive
Anthropic’s repeated Claude Fable access extensions expose a compute bottleneck that risks losing users to OpenAI’s confident GPT-5.6 rollout.
AI chatbots land in CarPlay—and it's messier than you think
Apple finally opened CarPlay to AI apps, but the implementation raises more questions than it answers about distraction and control.
Blackmagic Camera 3.3 puts your iPhone rig on your wrist
The Apple Watch companion isn't a gimmick—it's the control surface mobile video has needed since the beginning.
Apple's leak lawsuits are now discovery wars
The iOS 26 case shows how tech giants weaponize compliance paperwork to exhaust defendants before trial even starts.
Apple's OLED rollout tells you everything about margin strategy
The iPad Air getting OLED in 2027 isn't about better screens—it's about protecting the Pro line's pricing power.
Datasette ditches CSRF tokens for Sec-Fetch-Site
Simon Willison's Datasette just swapped token-based CSRF protection for header checks—Claude did the grunt work, but the implications run deeper.
Google's AI adoption looks like everyone else's
Steve Yegge claims Google engineers use AI like John Deere employees. Google says 40K engineers use agents weekly. Both are probably right.
April 1, 2026
From Medicine to UX: Why I Made the Switch
How a medical degree taught me diagnostic thinking, empathy, and systems reasoning — and why I chose to apply it to product design.