Loading…
Loading…
537field notes · new posts weekly
Claude Code, agent workflows, Next.js and India finance. Every post is checked against the live tool or the current spec before it ships.
We only write about what we deployed.
We only link to what we bought.
We name names when something sucks.
537 notes
A July 2026 Nature Machine Intelligence study held prompts, tools and compute constant and varied only coordination structure. Multi-agent architectures burned 58% to 515% more reasoning turns than a single agent and, at equal token budgets, did not reason better. For a solo builder that is a tax, not a feature. This is the one-agent default, the three tests that justify a second agent, and the lock-and-gate pattern that replaces crew roleplay.
Read →Character drift is the most common failed promise in generative video, and more reference images do not fix it. What fixes it is a production system: a hero sheet that separates immutable from mutable traits, a two-axis minimum lock, a continuity card per shot, and a drift score that tells you when to stop regenerating and fix the document. Model-agnostic, and it holds a face across a 6 to 8 shot sequence.
Read →The SIQA Mid-Year 2026 report covers 1,743 verified AI releases from 886 artists in 54 countries. Fully AI-generated tracks fell to 13.9% of submissions. AI-assisted work rose to 51.7% and human-plus-AI hybrids to 34.4%. Suno sits in 92.8% of workflows. Gospel nearly doubled its share. The market is splitting into slop and authored hybrid, and the stores are pricing the split. This is the release doctrine that follows the data.
Read →A song can stream and still never pay its writer. Distributors such as DistroKid pay the master recording. Performing rights societies such as IPRS pay the composition and lyric. AI musicians who type the words and direct the melody are often both author and composer, and still unregistered. This is the two-pipe money map: five royalty types, the ISRC versus ISWC split, the IPRS sequence that actually registers works, and the authorship documentation that makes the claim defensible.
Read →The fastest-growing agent repositories in September 2026 are not multi-agent frameworks. They are skill libraries: anthropics/skills, awesome-agent-skills with more than a thousand entries, NVIDIA/skills, 380-skill community packs. A skill is reusable competence a single agent loads on demand. An agent crew is rented attention that talks to itself. This is the decision table, the skill-file anatomy and the daily ritual that turn a strong single agent into a Creator OS.
Read →On 9 September 2026 Suno launched v6, v6-wild and v6-mini and pulled every earlier model from the picker with no rollback. Download caps that started on 3 September now meter every song in your library: 7 lifetime on Free, 20 a month on Pro, 60 on Premier. The artist identity you built was rented from a model version. This is the continuity system that moves it into a spec you own: fingerprint sheet, regeneration drill, drift scorecard, download ledger and catalogue triage.
Read →macOS 27 preinstalls fm, a command-line front end to the same on-device and Private Cloud Compute models that run Apple Intelligence. No account, no API key, no per-token bill. fm chat for a conversation, fm respond for scripts, structured JSON output and an OpenAI-compatible local server. This is the shell playbook: five pipe recipes worth keeping, where it fails, and when to reach for mlx_lm or Ollama instead.
Read →Spotify surfaces AI credits from the DDEX disclosure field your distributor passes, and says disclosed AI use is not down-ranked. The same platform removed 75 million spam tracks in a year. The difference between the two outcomes is metadata: writer credits, per-component AI disclosure, ISRC and ISWC, language tags, release cadence and titles that do not look like a content farm. This is the pre-upload gate every track passes in our catalogues.
Read →On 15 September 2026 Universal Music Group filed a 52-page complaint in Delaware accusing DistroKid of running an AI-slop pipeline and seeking up to $150 million. UMG says the case is not about disclosed AI music. It is still a shock to the distributor that carries 77.6% of verified AI releases. Add the Believe and TuneCore block, TIDAL zero-pay and Deezer fraud stripping, and independent AI musicians are exposed on one platform. This is the risk map and the four-step contingency.
Read →Twenty releases in three weeks buried a dozen settings that change how Claude Code behaves in production: output limits up to 128K characters, a restricted mode for untrusted repos, model-switch hooks, a live diff panel, per-agent prompt-cache TTLs and cache-miss diagnostics. We read the changelog from 2.1.238 to 2.1.261 so you do not have to. Here are the 12 worth enabling, with the exact JSON.
Read →Skills files are the new dotfiles: everyone has too many, nobody knows which ones fire. This week a Hacker News thread with 239 points collected how practitioners actually store, sync, test and prune them, and Claude Code shipped /skill-doctor to show which loaded skills go unused and what they cost. Here is the lifecycle that falls out of both, with the scripts.
Read →Cognition retired the Windsurf name on 2 June 2026 and turned the IDE into an agent command center that hosts any Agent Client Protocol agent. We set up Codex CLI, Claude Agent and OpenCode inside one Devin Desktop window: here is the registry file, the settings toggles, the auth path, who is billed for what, and where the workflow still leaks.
Read →Monday memo · free
1 field report, 3 links we stole ideas from, and 1 product we actually bought. No spam, unsubscribe in one click.
Beyond the writing
4,300+ developer tools, templates and starter kits — plus every free browser tool, no signup.
Full archive