无标题

What if you could skip the half-day of mockup sourcing and layout work that comes after finishing a logo?

I uploaded two logos into Image 2 and generated full brand proposal sets — signage, cards, packaging, booths, product surfaces — using a six-part framework.

A tech brand came out structured and enterprise-oriented:

A family nature park came out soft and warm:

Same workflow. Completely different brand worlds.

Google's "budget" model just beat its flagship

Gemini 3.5 Flash shipped at I/O 2026 today. The fast-and-cheap tier now outscores Gemini 3.1 Pro on coding and agent benchmarks.

Three-model cheat sheet: - Agent tool loops → Gemini 3.5 Flash - Code refactoring → Claude Opus 4.7 - Browser automation → GPT-5.5

Nobody wins everything. Pick by task.

Model ID: gemini-3.5-flash | $1.50/M input | $9/M output

API docs

Codex shipped its biggest infra batch in May 2026.

Mobile access (preview) — manage Codex sessions from ChatGPT on iOS and Android. Your phone is a control surface, not a code editor. Diffs, approvals, model switching, all in real time. Code stays on the host.

Hooks (GA) — inject custom scripts at 6 points in the agentic loop. Scan commands for credentials before execution. Auto-approve trusted actions. Log everything.

Access tokens (Business/Enterprise) — run Codex in CI/CD pipelines with workspace identity. No interactive login.

The pattern: each coding agent gets deeper in its own stack. The cross-agent problem — unified cost, fallback, one API key — stays unsolved until you route them through one endpoint.

The token cost problem in long-running agents now has an open-source memory-layer solution.

Tencent Cloud open-sourced TencentDB Agent Memory (MIT, May 2026). Instead of truncating or summarizing context, it offloads verbose tool output to local files and keeps a lightweight Mermaid task graph in the context window. The agent retrieves specifics by node ID when needed.

According to the project's benchmarks: −61.38% token consumption on WideSearch tasks, with success rate improving from 33% to 50%. PersonaMem long-term accuracy: 48% → 76%.

Four memory tiers (raw logs → facts → patterns → personas), fully traceable from top to bottom. Default storage is local SQLite — no external APIs.

Two optimization layers that stack: memory reduces tokens per call, gateway routing reduces cost per token. If your agents call LLMs through an API, both matter.

The coding agent market just got more fragmented.

Three platforms shipped infra updates in May 2026, on different dates:

Codex got real Windows sandboxing — OS-level isolation with dedicated user accounts and firewall rules (May 13).

Cursor launched cloud agent environments — multi-repo, Dockerfile config, scoped secrets, audit logging (May 13).

Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans (May 6).

Each agent is better at its own job. The cross-agent problem — unified cost, fallback, single API key — remains unsolved.

The production AI market has already split.

Vercel published usage data from 200,000+ teams on its AI Gateway (April 2026). Anthropic takes 61% of spend. Google takes 38% of volume. OpenAI tripled its spend share in one month.

58.9% of tokens now come from agentic workloads. 3.5% of requests trigger fallback, protecting roughly 5% of cost.

The teams running at scale are routing 30+ models by task type. The ones still on a single provider are overpaying on some tasks and underperforming on others.

Model selection is a routing problem now, not a one-time decision.

topic_id: claude-opus-4-7-routing-effort-guide platform: Tumblr content_mode: C_light_touch persona: Research Editor source_template: claude-opus-4-7-fast-mode_B_master.md canonical_url: https://evolink.ai/blog/claude-opus-4-7-routing-guide landing_url: null utm_source: tumblr utm_campaign: opus_4_7_routing

publish_status: draft

The interesting thing about Claude Opus 4.7 is not that it is better. It is that Anthropic shipped controls for how hard it thinks.

New xhigh effort level. Task budgets that cap how much an agent spends per turn. Per-request effort tuning.

This turns model selection from a one-time decision into a routing problem: which task gets deep reasoning, which task gets speed, which task gets a budget cap.

The teams that figure out routing will outperform the teams that just pick the "best model" for everything.

topic_id: claude-skills-engineer-workflow platform: Tumblr content_mode: C_light_touch persona: Research Editor source_template: claude-skills-engineer-workflow_A_master.md canonical_url: https://evolink.ai/blog/claude-skills-engineer-workflow landing_url: null utm_source: tumblr utm_campaign: claude_skills

publish_status: draft

The most popular AI coding repo right now is not a model or a framework. It is a folder of instruction files.

mattpocock/skills — 77K stars on GitHub. Composable Markdown files that go in your .claude directory and change how Claude Code approaches tasks.

The core idea: your coding agent fails not because it is not smart enough, but because it starts building before it understands what you want. These skills force the "ask first, build second" pattern.

The best one: /grill-with-docs — the agent interviews you about the project, then creates a shared vocabulary file it reads in every future session.

Simple install: npx skills@latest add mattpocock/skills

small workflow idea that feels obvious after you hear it:

stop asking claude code for giant markdown specs when what you actually need is something people will read.

thariq from the claude code team wrote about preferring html as an output format. not because markdown is bad. markdown is still great for readmes and docs that live in git.

but for ai-generated work artifacts, html can be much better.

think:

pr walkthrough with colored risk levels

architecture explainer with a diagram

design options in a comparison grid

onboarding page with collapsible sections

one-off ticket sorting board with drag and drop

the prompt is simple:

create a single self-contained html file that explains this feature. include a summary, diagram, annotated code snippets, risks, and next steps.

markdown is for storage.

html is for reading, comparing, and deciding.

that split feels like the useful part.

openai dropped a translation API that costs $0.034/minute

human simultaneous interpreters cost $25-44/minute

that's a 1000x cost difference

the model translates 70+ languages in real time, preserves the speaker's tone and emotion, and runs 24/7

gaddafi's interpreter once broke down after 75 minutes and screamed "i can't take this anymore" into the mic. the AI version doesn't do that

we are living in the future and it costs two cents a minute

so anthropic shipped something called Dreaming for their Managed Agents platform and the name is doing a lot of work but the actual thing is genuinely interesting

between sessions, the agent reviews its own past conversation transcripts. finds patterns. repeated mistakes. approaches that worked. edge cases it missed. writes those learnings into its own memory. without being asked.

the thing that gets me: a single session is blind to what happened in other sessions. your support agent doesn't know it made the same wrong call 12 times this month. Dreaming is the mechanism that surfaces that.

two modes — automatic (agent writes to memory directly) or human-review (agent proposes, you approve). still in research preview.

it ships alongside automated output grading (Outcomes) and multi-agent orchestration. the three together form a loop: decompose work, grade it, remember the failures.

the mental model shift is from "stateless tool" to "accumulating system." an agent that's been running for three months has three months of self-curated experience. a freshly deployed agent starts from zero. same prompt, same model. different histories.

that's the part I keep thinking about. the competitive advantage of an agent stops being just its configuration. it becomes its history.

every time you switch models mid-conversation, open a new session, or add an MCP tool on the fly — you're invalidating the cache. and the cache is everything.

here's the mechanic: claude code structures every request as a prefix chain. system prompt first, then tools, then project docs, then messages. the api caches from the front. if the prefix matches next time, those tokens cost 1/10 the normal price. one changed byte anywhere in the chain and it recalculates from scratch.

so switching from opus to haiku for a "quick task" and back? you just wiped everything you'd built up. the rebuild cost is usually more than you saved.

the counterintuitive part: long conversations get cheaper, not more expensive. claude code's compaction is designed to preserve the cache chain. you're not accumulating cost — you're accumulating savings.

and claude --resume is the most underrated flag in the whole tool. it picks up the cache chain exactly where you left it.

codex v0.128.0 shipped /goal -- persistent objectives that survive across terminal sessions. define a target, close everything, come back tomorrow, it picks up where it left off.

also /pet -- a little animated creature that shows you what codex is doing in the background. sam altman: "not the most important thing we've done, but more useful than it looks."

ai coding tools are now competing on vibes and persistence, not just code quality. interesting times.

codex v0.128.0 shipped /goal -- persistent objectives that survive across terminal sessions. define a target, close everything, come back tomorrow, it picks up where it left off.

also /pet -- a little animated creature that shows you what codex is doing in the background. sam altman: "not the most important thing we've done, but more useful than it looks."

ai coding tools are now competing on vibes and persistence, not just code quality. interesting times.

claude opus 4.7 dropped and the agentic coding improvements are real but they also removed temperature/top_p/top_k controls so if your production code uses those you have a migration task not a footnote

also the tokenizer changed so same prompts cost 1x-1.35x more tokens. measure before you commit

the sleeper feature is task_budget: give the model a token budget for the full agent loop and it prioritizes + winds down gracefully instead of stopping mid-task

tldr: upgrade where execution quality matters, test where cost matters

Karpathy's four LLM coding rules are now a single CLAUDE.md file with 97.8k GitHub stars.

The rules: think before coding, simplicity first, surgical changes, goal-driven execution.

The one that actually matters: surgical changes. Ask for a bug fix, get a clean 4-line diff instead of 40 lines of unrequested type hints and reformatted quotes.

One command: /plugin marketplace add forrestchang/andrej-karpathy-skills

Repo: github.com/forrestchang/andrej-karpathy-skills

Karpathy's four LLM coding rules are now a single CLAUDE.md file with 97.8k GitHub stars.

The rules: think before coding, simplicity first, surgical changes, goal-driven execution.

The one that actually matters: surgical changes. Ask for a bug fix, get a clean 4-line diff instead of 40 lines of unrequested type hints and reformatted quotes.

One command: /plugin marketplace add forrestchang/andrej-karpathy-skills

Repo: github.com/forrestchang/andrej-karpathy-skills