Claude Code

Anthropic's coding agent. Originated in Anthropic Labs (late 2024) under Boris Cherny. The agent currently maintaining this Second Brain.

Origin & inflection (per Boris Cherny on Coding Is Solved (Sequoia AI Ascent))

Built pre-PMF for ~6 months because Boris bet the next model would unlock agentic coding
State of the art at launch was tab-complete (Sonnet 3.5)
PMF inflection: Opus 4 in May 2025, kept inflecting through 4.5 → 4.6 → 4.7
Codebase: TypeScript + React (chosen because most on-distribution for the model in late 2024)

Role in this wiki

The maintainer. This vault's CLAUDE.md is the schema Claude Code reads when ingesting/querying/linting. See LLM Wiki Pattern and Harness (LLM Agents).
Karpathy's gist (LLM Wiki (Karpathy gist)) lists Claude Code as one of the canonical agents this pattern is designed for.

Surfaces

CLI — original surface; majority of Boris's audience uses it
Desktop app — secondary
iOS app — Boris's primary surface in 2026 (5–10 sessions, hundreds of agents)
VS Code / JetBrains plugins — minority usage per Boris's poll
Routines — server-side scheduled agents (don't need laptop open)

Extension surfaces relevant to this wiki

CLAUDE.md — per-project schema/instructions; the load-bearing config for this vault
Skills — invokable workflows (e.g. the Printing Press factory ships as a skill); self-improvable via auto-research loops over Binary Eval Assertions — see Skills (Claude Code)
MCP servers — alternative tool-loading mechanism; criticized in CLI vs API vs MCP for context bloat
/loop — schedule a recurring cron job (Boris Cherny: "loops are the future")
/batch — parallel agents
Sub-agents — agent-spawned-by-agent
Workflows / Ultracode — a planning agent that fans out to hundreds of parallel sub-agents (e.g. one per file for a whole-codebase review). MYTHOS MYTHOS MYTHOS (Matthew Berman) watched 63 agents run in parallel for a single build and 100+ for a codebase review; Fable 5 is described as especially good at exploiting it. The product-level instance of the multi-agent / sub-agent loop.
Hooks, slash commands, settings.json — other knobs (not yet used in this vault)

Boris's workflow (canonical example of agentic engineering at scale)

Most work from his phone via the Claude iOS app
5–10 sessions, several hundred agents running at any time
Overnight: a few thousand agents on deeper work
Dozens of /loop jobs running: babysitting PRs (auto-rebase, fix CI), keeping CI healthy, clustering Twitter feedback every 30 min

"There was a day last week I did like 150 PRs in a day. That was a record."

Cross-source observations

All five ingested sources treat Claude Code as the default agent — it's the hub entity in this wiki. Both Andrej Karpathy (calls it the canonical Software 3.0 agent) and Praveen Akkiraju (calls it a "very well-defined harness") reference it explicitly. The product-vs-model attribution Boris discusses (50/50 in the early days, model-leaning over time) is a useful frame for any agent-product analysis.

The IC / non-engineer view (Sandeep)

How To Use Claude Better Than 99% Of People (theMITmonk) reframes Claude as a five-surface product stack rather than just the CLI tool, aimed at the consultant / business-user / IC operator audience (the complement to Boris's CLI-engineer audience):

Surface	Activity	Sandeep's pitch
Claude Chat	Think	Messy problems, fuzzy ideas
Claude Projects	Remember	Persistent context across sessions; user customizes at project level, not account
Claude Cowork (desktop)	Execute	Multi-step tasks with local-disk output; sandbox to a specific folder
Claude Code	Build	"If you can type in English, you don't need a CS background" — vibe-coding for the non-engineer
Claude Chrome (extension)	Browse	Acts inside the browser workflow — job listings, industry reports, forms

Plus Skills (repeatable workflows) and MCP connectors for fan-out to other tools. Worked example in the video: Cowork + MCP → Higgsfield for ad-creative generation.

This view is structurally distinct from Boris's "I do 150 PRs a day from my phone" CLI worldview — but both are valid; the stack frame lands for users whose work doesn't fit a code-PR shape.

As an operating system (Nate Herk)

I Turned Claude Opus 4.8 Into My Entire AI Operating System (Nate Herk) pushes the surface-collapse one step further than Sandeep: instead of a 5-surface stack, work out of Claude Code by default for everything (the Default Shift) so context compounds in one place. The OS is just files/folders → tool-agnostic across Claude Code / Codex / OpenClaw. Built via the Four C's Framework; governed via the Bike Method and Capabilities vs Instructions (Agent Keys). See AI Operating System (AIOS).

Counterpart at the other lab

Codex (OpenAI) is the structural twin of Claude Code. Ryan Lopopolo's Harness Engineering (Ryan Lopopolo, AI Engineer) talk describes a workflow nearly identical to Boris's — full-send agentic engineering, no editor use, hundreds of agents running, harness as the leverage point. The choice of lab matters less than the discipline around it. Lopopolo's prescription to depend on first-party harnesses directly via SDK or CLI (to ride the post-training leverage the labs invest in their own apply-patch / bash-invocation semantics) applies equally to Claude Code and Codex.

2026-06-13 — writes most of Anthropic's own code (Economist)

Per How AI Got Better at Building Itself (Economist), Claude Code now writes >80% of Anthropic's own published code (up from low single digits). This makes it the most direct corporate instance in the Recursive Self-Improvement argument — the lab's own development loop is now majority-agent-written, the scaffolding building the next generation of the scaffolding.

2026-07-24 — the new rules of context engineering (Shihipar's official Anthropic blog)

Per The New Rules of Context Engineering for Claude 5 (Thariq Shihipar), Anthropic published the official rulebook for how to author system prompts, CLAUDE.md files, and skills for the Claude 5 generation. Same 80%-reduction claim as the WF talk below, now with a qualifier — "with no measurable loss on our coding evaluations" — and formalised as six named pattern shifts:

Rules → Judgment — let Claude use contextual judgment rather than explicit worst-case rules.
Examples → Interface design — expressive tool interfaces (parameters, enums, hints) replace tool-usage examples; examples now "constrain them to a certain exploration space."
Upfront context → Progressive disclosure — load context selectively via skills, deferred tool definitions, and ToolSearch; avoid "making these a central repository for every known practice."
Repetition → Simple descriptions — put tool instructions only in tool descriptions; delete redundancy across system prompt + tool description.
Manual memory → Auto-memory — Claude auto-saves relevant memories to work context rather than users pre-writing CLAUDE.md.
Simple specs → Rich references — HTML artefacts, code samples, test suites, and function implementations beat markdown specs. "An HTML mockup will generally produce better results than a description of the design or a screenshot."

Anthropic's operational answer to the ongoing rightsizing problem: claude doctor / /doctor — a first-party command inside Claude Code that helps rightsize existing skills and CLAUDE.md files for the newer model class. First-party version of Pocock's deletion test from the Skill Checklist (Pocock). See Context Rightsizing for the vault's concept page synthesising the discipline.

Direct-bearing artefact rules from the blog:

System prompt: keep product context; delete the obvious things Claude can see in the file system; spend most authoring effort here for custom agents.
CLAUDE.md: lightweight repo description + non-obvious gotchas only; split verification/validation into separate skills; use progressive disclosure.
Skills: lightweight retrieval guides; divide long skills across multiple files.
References (@mentions): prefer code / test suites / HTML mockups over prose descriptions.

2026-07-07 — 80% of the system prompt removed for the Fable-class model

Per Field Guide to Fable (Thariq Shihipar, Anthropic), Thariq Shihipar (MTS on Claude Code) revealed that Claude Code removed 80% of its system prompt for the new Mythos-class tier. His diagnosis: examples that used to help older models now constrain Fable, which is more imaginative than the examples given. Give the model context, not constraints; avoid "do not do this" phrasing that was necessary for previous generations.

Direction of travel: small system prompt + many tools + fewer examples (Fable-class), reversing the Sonnet-3.5-era practice of small system prompt + few tools + many examples, and the intermediate large-system-prompt-with-examples-and-many-tools generation.
Product-level trace: the ask-user-a-question tool (Shihipar built it) went from barely working under Opus 4 → interviewing users with 40+ questions under Opus 4.5 → building embedded-HTML questionnaires under Opus 4.8 / Fable. A concrete instance of the Capability Overhang closing when the harness catches up to the model.
Practitioner heuristic Shihipar advises: find your unknowns before you plan. Claude Code's default plan mode "is extremely eager to create an asset" — for Fable-class work, prefer an interview / grill-me / references-first flow that surfaces unknowns before the agent starts fanning out. Direct heir to Matt Pocock's Grill Me skill.

Related upstream methodology: Design Concept (Brooks) and Ubiquitous Language — two-source convergence between Pocock's Software Fundamentals talk and Shihipar's Fable talk on the same "front-load the alignment, don't front-load the artifact" prescription.

2026-06-27 — separated from Cowork in Anthropic's own usage data

Anthropic Economic Index Cadences Report (June 2026) is the first time Anthropic publishes usage data with Claude Code split out as its own cohort (alongside Claude Cowork and 1P API). The artifact-mix difference is the headline:

Cowork (with chat) — dominated by explanations and guidance (~⅓ of artifacts)
Claude Code — much more code and technical artifacts

This is the data-side confirmation of the surface-split this page has tracked since Sandeep's 5-surface view — and it now shows up in Anthropic's own analytics, not just in the product UI. See Claude Cowork for the sibling page; see Artifacts (Claude Output) for the classification primitive Anthropic uses to make the split visible.

2026-07-13 — "context rot" and the fresh-instance-per-task pattern (GSD)

Per Spec-Driven AI — BMAD SpecKit GSD Superpowers (Vimal Dwarampudi), Lex Christopherson's GSD (Get Shit Done) framework is built on the claim that "Claude's output quality degrades measurably as the context window fills — and the fix is to never let it fill." The named phenomenon: context rot. His prescription: every task gets a fresh 200k-token Claude instance; independent tasks run in parallel waves; the main session stays lean while subagents do the heavy lifting.

This is the same pattern the vault already tracks under Loops as Core Primitive (Cherny's "5–10 Claudes in parallel"; Karpathy's fresh subagent per eval), now given a name and packaged as a methodology.
Discipline complement — Superpowers (Jesse Vincent): enforces TDD at the moment of implementation — any code written before a failing test is literally deleted. 170k+ GitHub stars in three months as an adoption signal.
Constitution complement — SpecKit (Microsoft): constitution-first pipeline; the constitution rail governs downstream interactions once set. Structurally akin to CLAUDE.md / SKILL.md at the project level.
These stack: BMAD + SpecKit for planning; GSD + Superpowers for runtime discipline. See Spec-Driven Development (SDD) for the umbrella and how it composes with Pocock's Specs-to-Code critique.
Note the technical distinction: context rot names within-session quality degradation, not the 5-minute prompt-cache TTL Anthropic documents for pricing/latency. Independent benchmark of context degradation vs fresh-instance quality is still a follow-up.

2026-07-06 — Claude Tag: Claude Code from anywhere in Slack

Per The Big Ways AI Just Changed (AI Daily Brief), June 2026 brought Claude Tag — not another Claude-in-Slack chat integration, but a way for anyone, from any part of Slack, to invoke the full power of Claude Code. Three effects Whittemore highlights:

Democratization — non-terminal users get Claude Code's technical capability without the CLI.
Persistent group context — the Slack thread/channel becomes the context substrate.
AI shifts from individual to group experience — the headline claim, now traced to the primary (Anthropic, "Introducing Claude Tag", June 23 2026): "Today, 65% of our product team's code is created by our internal version of Claude Tag." Note the exact wording — the stat is attributed to the internal Claude Tag product (described as an evolution of Claude Code that starts on Slack), which Whittemore's "initiated from Slack" paraphrase compresses slightly. Covered same-day by VentureBeat and Fortune; launched as a beta for Enterprise/Team customers.

This is a fourth invocation-surface story to add to the CLI / desktop / iOS / plugins list above — and evidence that the "surface" that wins may be wherever the team conversation already lives. (Still first-party with no named speaker; independent corroboration of the 65% would firm it up.)

2026-08-09 — action-time supervision ("approve forming / review forming") named as shipped

Nate B. Jones in Your Chatbot Hallucinated in 2024 Your Agent Lies in 2026 (Nate B Jones) names Claude Code (alongside Codex) as shipping an action-time LLM as Judge variant — a separate agent inspects tool-call requests during execution for intent-alignment. Jones's transcript wording is "approve forming or review forming" — likely a mishearing of the shipped feature's actual name; the primitive is real. Jones positions it as the minimum-viable defense against the agent-lying failure mode. Composes with the batch-review reviewer-agent pattern (persona reviewers on push) as different insertion points of the same primitive.

Sources

Boris Cherny on Coding Is Solved (Sequoia AI Ascent) (origin & internals)
LLM Wiki (Karpathy gist)
Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent)
Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk)
Printing Press (Nate Herk video)
How To Use Claude Better Than 99% Of People (theMITmonk) (5-surface stack view; non-engineer audience)