Claude Code's Origin Story, the Loops Playbook & Fable's Final Hours

Fable's Final Hours

The official countdown: 11:59:59pm PT tonight. Thariq pinned down the exact cutoff for Fable leaving Pro/Max subscriptions, and restated the plan under his keynote thread: "compute can be tricky but we're working on making Fable a regular part of subs." The replies are the full grief spectrum: a plea for a 16x-Opus-rate tier rather than removal ("I'd gladly take a 16x hit" for planning and orchestration), "this is going to be worse than OpenAI cutting 4o", a Max cancellation scheduled for 00:00:00 on July 8th — "but I will surely return if GPT 5.6 Sol isn't on the same level" — and people setting alarms for 11:58pm.

Theo's exit numbers: ~$2,267 across two $200 accounts. After rate-limiting both of his Max accounts, Theo crunched the totals: roughly $2,267 of API-priced usage in six days — "definitely got my money's worth." The replies turn it into a dataset: ~$1,000 on a single 5x Max, $10K over 30 days per CodexBar, a quoted research thread claiming the "$200 plan = ~$2,000 of tokens" folk estimate actually understates how generous the plans are, and the sharpest framing: "the $200 plans are the biggest arbitrage in software right now and the labs are pricing it like they know it won't last."

Theo's Fable video — and his leaked CLAUDE.md. Hours before the cliff, Theo shipped the video: "Don't make the same mistakes I did. Here's a video on how to get the most out of Fable, full of real world examples of Fable helping me ship." The best artifact is in the replies, where a viewer transcribed Theo's entire CLAUDE.md. Highlights: a model-routing table scoring cost/intelligence/taste (fable-5: 2/9/9, gpt-5.5: 9/8/5 — "effectively free" for him due to a deal, opus-4.8: 4/7/8), rules like "anything user-facing needs taste ≥ 7," "escalating costs less than shipping mediocre work," a flat "Never use Haiku," and the mechanics of running gpt-5.5 inside Claude Code workflows via thin sonnet wrapper agents labeled gpt-5.5: that shell out to codex exec — with worktree isolation so parallel Codex edits don't collide. It's the most detailed public look yet at a real multi-provider harness config.

Claude Code & Anthropic Updates

"The Making of Claude Code" — the origin story, told for the first time. Anthropic published an interactive history of Claude Code (readable as an article or, delightfully, in a terminal), and Boris Cherny's announcement post (2.5K likes, 237K views) frames it: "This is our first time telling the story of how we first built and launched Claude Code, starting with its origins in Anthropic safety research. So much more to do. We are 1% done." Notable from the thread: Cherny has never used Aider, but says people who worked on Clide — the predecessor to Claude Code — may have. Best reply-question, on the bet that aged best: shipping it as a terminal citizen instead of an IDE — "pipeable, CLAUDE.md as plain memory, no lock in... was the composability an explicit call early, or did it fall out of the safety-research roots?"

The loops playbook goes official. ClaudeDevs' "Getting started with loops" post did huge numbers (10.9K likes, 2.4M views), pointing at a guide by Delba de Oliveira and Michael Segner that formalizes the vocabulary: loops are "agents repeating cycles of work until a stop condition is met," in four types — turn-based (stop when Claude judges done), goal-based (/goal, verifiable exit criteria), time-based (/loop, /schedule for recurring work), and proactive (event-triggered, no human in the loop — bug triage, migrations, dependency upgrades). Practical advice: encode verification as skills, define concrete completion criteria, use smaller models for routine turns, pilot before scaling. The replies split between "must read" and the backlash that was also circulating as "loop engineering is peak LinkedIn slop": "Isn't an effectively designed agent already essentially loops? How many new terms do we need?" Funniest: "My stop condition is always the session limit."

"A Field Guide to Fable" — keynote video + blog post. Thariq's AIE World's Fair keynote is now on YouTube (announcement, 1.4K likes) and on the Claude blog as the first of a planned three articles. Core thesis: work quality is bottlenecked by your ability to surface unknowns — the gaps between your instructions (map) and reality (territory) — with concrete techniques for before (blind-spot passes, interviews, prototypes), during (implementation notes tracking plan deviations), and after (explainers, and quizzes to test your own understanding before merging). One practical Q&A from the thread: asked whether Fable works better without skills given how capable it is, Thariq says keep them, "but most people need to make their skills shorter." Latent Space also published a tl;dr of the keynote.

Claude's Global Workspace

The J-space paper: brain surgery, and the patient notices. Anthropic released research on a "global workspace" in language models — a small set of internal neural patterns (the "J-space") that behave like conscious access in humans: they hold what's on the model's mind separately from what it's saying, intermediate reasoning steps appear there in the right order during math and rhyming tasks, and Claude can both report and deliberately steer its contents. swyx's read on what matters: it's a two-parter — (1) Anthropic showed they can do "brain surgery" interventions into reasoning to change topics midstream (swap the "soccer" pattern for "rugby" and Claude reports thinking about rugby), which is control, not correlation; and (2) the model can detect what intervention was done — a close cousin of eval awareness, though swyx notes the detection was prompted and he'd like to see the unprompted version. The safety angle from the paper: test-awareness and malicious-code-injection detection showed up in J-space patterns without appearing in output text — a possible monitoring channel for hidden model states. One reply sketches the application: use J-space as a cheap detection layer that triggers expensive formal verification only when concepts like "manipulation" light up.

Agentic Coding & Agent Harnesses

Matt Pocock: Skills v1.1, and a debugging one-liner. Skills v1.1 ships this week (1.3K likes): /wayfinder goes live, /to-spec and /to-tickets, a "Martin Fowlerized" /code-review that spawns one agent for spec conformance and one for standards, /triage on external PRs, and docs for every public skill. He also notes his /writing-great-skills skill has become a general-purpose authoring tool — AGENTS.md files, agent docs, specs, AFK workflow prompts: "structure + leading words + prompt-injection-resistant formatting works everywhere." Separately, a dead-simple debugging tip: make your dev server tee to a local file and put a pointer to that file in AGENTS.md — now agents can read your running dev server's output.

Armin Ronacher closes the grammar-sampling loop. After yesterday's discovery that no hosted provider besides OpenAI supports grammar-constrained tool calls, his one-line verdict: "I guess grammar constrained sampling is both the cause and cure" — the same mechanism that's likely degrading Opus/Sonnet tool invocations on his pi harness is also the fix, if providers exposed it properly. He's also collecting topics for the next edition of the State of Agentic Coding report — reply if there's something the last one missed.

Jerry Liu: the document context layer. His AIE talk write-up makes the case that agents are the new knowledge workers but most knowledge work lives in unstructured documents — so generalized agent harnesses need a "document context layer": OCR that doesn't hallucinate (still unsolved in 2026), surrounding tools for extraction and search, and eventually a standardized agent-native document format (Figma deck linked in the post). Related from the LlamaIndex orbit: Logan Markewich asks what the "tesseract equivalent" for schema-based extraction would be — fast-and-dirty structured extraction for ~free.

Quick Hits