claudecode-writer

A Claude Code workspace template for content production, from Women Defining AI. MIT, 214★, 44 forks. “Turns Claude Code into your personal content creation system.” Created 2025-09-03 and last pushed 2025-09-06 — three days of commits, then nothing for eleven months.

It is small enough to describe completely: three subagents (linkedin-repurposer, newsletter-repurposer, conversational-repurposer), three slash commands (/extract-themes, /research, /write), a 1.5 KB CLAUDE.md, and two context files the user fills in — writing-examples.md and research-sources.md. Plus empty rawnotes/, drafts/ and research/ folders.

The pipeline

Unstructured input → structure → one artifact → many artifacts:

1. Drop voice notes, fragments and meeting scraps into rawnotes/ with no organization.
2. /extract-themes reads all of them and proposes recurring themes, a stated angle, and content briefs.
3. /research [topic] checks the user’s declared priority sources first, then looks for gaps and angles.
4. /write produces an 800–1500 word article.
5. Each repurposer subagent transforms that article for one platform.

The fan-out is the design. Research and drafting happen once; platform adaptation is N parallel specialists over a single source document, each holding one platform’s conventions. Compare autonovel, which fans out into formats too (PDF, ePub, audiobook, landing page) but only after a long sequential pipeline with quality gates at each stage. This template has no gates, no scores and no retry loop — one pass per stage, and the human is the only reviewer.

Why it is worth a page: the untested case has a specimen

context-files-ablation-paper measured AGENTS.md-style project context and found no correctness effect, then named exactly one thing its design could not test — purpose-built, task-specific context, “a fact the agent provably cannot infer.”

context/writing-examples.md is that case. Two or three samples of the user’s actual posts is information no model can derive from the task, and the README treats it as the load-bearing step: “Add Your Writing Examples (Most Important!)”, with the troubleshooting entry for “Claude doesn’t sound like me” being add more examples. This is context as few-shot demonstration rather than as instruction — and the ablation’s mechanism argues it should behave differently, since near-miss failures there were skill gaps rather than knowledge gaps, and voice is precisely a knowledge gap.

Nobody has measured it. This template ships no evaluation, and a style match has no gold test to run against. It is a well-posed instance of the open question, not an answer to it.

The demo output undercuts the pitch

The README shows what the LinkedIn repurposer produces. Reproduced in part because it is the source’s own choice of best foot forward:

Most productivity advice for remote workers is still stuck in 2019. Here’s what actually works: ❌ Time blocking (ignores energy rhythms) ✅ Energy blocking (work with your natural peaks) … What’s your best remote productivity hack that no one talks about? #RemoteWork #Productivity #WorkFromHome #FutureOfWork

Every marker of generic LLM social copy is present: the “still stuck in 2019” hook, the ❌/✅ ladder, the one-line paragraphs, “Game changer.” in the Twitter variant, the engagement-bait closing question, the hashtag stack. A system whose central promise is “your voice stays authentic across all platforms” demonstrates itself with output that could belong to anyone.

That is the failure renwei-writing built a whole skill to name — editing that erases the person behind the words — arriving here from the generation side rather than the editing side. It is also the reason this hub runs avoid-ai-writing over its own prose. The template’s voice-learning step targets the right problem; its showcased result is the problem.

Tier and freshness

T3 — a template with no evaluation, no benchmark, no deployment evidence beyond its own worked example, and the worked example is illustrative rather than run. The 214★/44-fork ratio suggests people are using it as a starting point, which is usage evidence and not quality evidence.

Stale by construction. Last pushed 2025-09-06, so it predates the agentskills.io standard, the skills ecosystem this spoke tracks, and the subagent conventions that followed. Its .claude/agents/

• .claude/commands/ layout is an early-generation structure, and its three specialists would today more likely be portable agent-skills than repo-local agents. Read it as a snapshot of what the pattern looked like when Claude Code workspaces were new.

Related

autonovel · renwei-writing · context-files-ablation-paper · writing-a-good-claude-md · claude-md · agent-skills · agentic-coding-harness · agent-orchestration · synthesis