Claude Code Ships AGENTS.md, Skills Prove Their Lift, and Open-Source Tools Pile Up

Today's findings

1. #1 Claude Code adds native AGENTS.md supporttool

   Claude Code now reads AGENTS.md as a fallback when no CLAUDE.md exists, via a new built-in "mod" system.

   Claude Code · agent instructions

   One instruction file now feeds every coding agent

   Before: a file per tool

   • CLAUDE.md — Claude Code
   • AGENTS.md — Codex
   • Own rules file — Cursor
   • Same rules, kept in sync by hand

   → Claude Code 2.1.277+

   Now: AGENTS.md fallback

   • AGENTS.md — shared by all
   • Claude reads it when CLAUDE.md is absent
   • CLAUDE.md still wins if present
   • No extra file to create

   New built-in "mod" system falls back to AGENTS.md — existing projects need no change.

   Why it matters: Teams juggling multiple coding agents (Claude Code, Codex, Cursor) can maintain one shared instruction file instead of duplicating project rules per tool.

   How to apply: Update to Claude Code 2.1.277+; if a project already has an AGENTS.md for other agents, no extra file is needed — Claude will pick it up automatically when CLAUDE.md is absent.

   claude-codeagentstooling

   Read more: Quoting Thariq Shihipar · Claude Code is getting native AGENTS.md support!

2. #2 Nested CLAUDE.md files silently skip loading outside the Read tooltip

   In a monorepo, service-level CLAUDE.md files only load when Claude uses its native Read tool on that folder — Bash, cat, sed, grep, or heredoc edits don't trigger it.

   Claude Code · monorepo context

   Nested CLAUDE.md loads on the Read path only

   Native Read tool

   • Opens the folder's CLAUDE.md
   • Service rules enter context
   • Visible as a Read call in the transcript

   vs

   Bash · cat · grep · sed

   • File is read, loader never fires
   • Service rules silently absent
   • Common in auto/agentic shell edits

   Same file, same folder — only one path enforces the rules.

   Consolidate into root CLAUDE.md, or tell agents to Read the nested file first.

   Why it matters: Teams relying on nested CLAUDE.md for service-specific rules may be getting none of that context enforced, especially in auto/agentic modes that favor shell commands over Read.

   How to apply: Audit session transcripts for Read tool calls on nested CLAUDE.md paths; if agents are editing via Bash/grep without ever Reading those files, consider consolidating rules into the root CLAUDE.md or explicitly instructing agents to Read the nested file first.

   claude-codecontextmonorepo

   Read more: In auto mode, your nested CLAUDE.md files are probably not being loaded

3. #3 ossrules.md catalogs real-world AGENTS.md and CLAUDE.md filesrepo

   A new index collects production AGENTS.md/CLAUDE.md files from ~100 open-source repos (Next.js, Ollama, Transformers, etc.) to show how maintainers actually prompt coding agents.

   ossrules.md

   A corpus of how maintainers actually prompt coding agents

   ~100

   open-source repos indexed

   Next.js · Ollama · Transformers

   2 files

   AGENTS.md and CLAUDE.md per repo

   Conditional

   tests run only when relevant, not always

   Concrete

   scoped prohibitions, not open-ended bans

   Browse repos near your stack, adapt their rule patterns.

   Why it matters: Seeing patterns like conditional test execution and concrete (not open-ended) prohibition rules gives a template for writing more effective, less contradictory project instructions.

   How to apply: Browse ossrules.md for repos similar to your stack and adapt their conditional-test and rule-scoping patterns into your own CLAUDE.md/AGENTS.md.

   claude-codeprompt-engineeringagents

   Read more: Cool resource: ossrules.md catalogs production AGENTS.md and CLAUDE.md files across open-source projects

4. #4 Prompt-cache writes cost ~80x a read — keep Claude Code sessions warmtip

   A cold 1-hour cache write on current Claude pricing runs roughly 80x the cost of a warm cache read, so returning after a break can be surprisingly expensive.

   Why it matters: Teams paying per-token for Claude Code should know that idle gaps between prompts (not just total tokens) directly drive cost via repeated cache misses.

   How to apply: Batch related work into continuous sessions rather than sporadic check-ins, or explicitly ping the session periodically to keep the cache warm before asking a real question.

   claude-codecostcaching

   Read more: Keep Claude Code’s cache warm. On Fable 5.1, a 1-hour cache write costs 80x a read.

5. #5 Curated Claude Skills add double-digit accuracy lift, per multiple benchmarkstechnique

   Cited results (Microsoft SkillOpt, NVIDIA SkillEvaluator, SkillsBench) show well-written skills lifting task accuracy by ~16-60 points depending on domain.

   Why it matters: Skill-authoring quality can matter as much as model choice — a spreadsheet skill reportedly took Claude from 22 to 82 on SpreadsheetBench.

   How to apply: Audit existing custom skills against these benchmarks' patterns (explicit steps, verified sub-tasks) rather than assuming a skill is helping just because it exists; consider A/B testing task accuracy with a skill on vs. off.

   claudeskillsagents

   Read more: Why your skills suck(and how to improve them)

6. #6 Cache-to-Cache: direct semantic communication between LLMspaper

   A 2025 arXiv paper (resurfacing on HN) proposes letting LLMs exchange KV-cache representations directly instead of round-tripping through text, for faster multi-model pipelines.

   Why it matters: Teams building multi-agent or multi-model pipelines (e.g., a small local model handing off to Claude) could cut latency and token overhead if cache-level exchange is applicable to their stack.

   How to apply: Read the paper before investing in custom inter-model protocols; it's most relevant to teams running open-weight models locally where KV-cache access is possible (less applicable to closed API-only setups).

   paperllmmulti-agent

   Read more: Cache-to-Cache: Direct Semantic Communication Between LLMs (2025)

7. #7 FreeComputerUse: a cheaper local hybrid browser agenttool

   An open-source alternative to Claude/GPT computer-use agents pairs a small local model (DeepSeek Flash) for planning with Playwright executing and verifying actions locally.

   FreeComputerUse · hybrid browser agent

   A small local model plans; Playwright executes and verifies on the machine

   1

   Plan

   DeepSeek Flash, local

   2

   Execute

   Playwright acts

   3

   Verify

   check page state

   back to the model only on a new decision

   Open-source take on computer-use: the loop stays local, so repetitive browser work stops burning tokens.

   Why it matters: Full computer-use agents burn tokens fast on repetitive browser tasks; this pattern only calls the model when a new decision is needed, cutting cost for routine automation.

   How to apply: For repetitive browser workflows (form-filling, scraping, monitoring), consider this plan-with-small-model/execute-locally pattern instead of a full vision-based computer-use loop.

   agentslocal-llmbrowser-automation

   Read more: Computer use was too expensive for my daily tasks, so I built a hybrid browser agent

8. #8 Local Whisper + Claude Code pipeline for turning recordings into structured outputtechnique

   An engineer ran Whisper large-v3 locally on an RTX 5090 to transcribe 14 unlabeled trade-show recordings, then had Claude Code build the folder structure and CRM entries from the transcripts.

   LOCAL-MODEL + AGENT PATTERN

   Audio never leaves the box; only text reaches the agent

   1

   Raw audio

   14 unlabeled trade-show recordings

   2

   Whisper

   large-v3, local on RTX 5090

   Costly, private step runs on-device

   3

   Transcripts

   plain text, on disk

   4

   Claude Code

   classify + label

   5

   CRM entries

   folder tree + records

   Local inference for the expensive private step; the agent handles structuring.

   Why it matters: Demonstrates a practical local-model-plus-agent pattern: use local inference for the expensive/private step (audio transcription) and Claude for the structuring/reasoning step, avoiding cloud audio-upload costs and privacy concerns.

   How to apply: For similar unstructured-recording backlogs, run Whisper (or whisper.cpp/GGUF variants) locally first, then feed transcripts to Claude Code for classification, labeling, or CRM-style structuring.

   local-llmwhisperclaude-code

   Read more: How I Used Claude Code to Turn 14 Trade Show Recordings Into My Follow-Up

9. #9 AI CFO Skills: open-source finance skill library for Claude/ChatGPTrepo

   An open-source skills library gives Claude or ChatGPT the ability to answer runway, margin, and profitability questions directly from financial data.

   Open-source repo

   A drop-in finance skill set for Claude and ChatGPT

   AI CFO Skills

   tool Open source

   Skills library · Claude / ChatGPTSep 2026

   Claude ProjectClaude Code skills folderChatGPT

   runPoint it at exported P&L / ledger data → runway, margin and profitability answers

   Inspectable prompting logic for financial analysis, instead of rebuilding it per team.

   Why it matters: Ready-made, inspectable skills for a common but fiddly domain (financial analysis) save teams from building this prompting logic from scratch.

   How to apply: Pull the skill library into a Claude Project or Claude Code skills folder and point it at exported P&L/ledger data to get board-pack-style summaries and anomaly checks.

   open-sourceclaudeskills

   Read more: Turn Claude into your own CFO · Turn ChatGPT or Claude into your own CFO

10. #10 SPARSEUP: 149M-parameter open-source sparse embedding modeltool

    Linkup Research open-sourced SPARSEUP, a SPLADE-style sparse retriever scoring 56.4 on BEIR-13 with human-readable token weights.

    Why it matters: Sparse embeddings that expose interpretable token weights are easier to debug than opaque dense vectors, useful for teams building or troubleshooting RAG/search systems.

    How to apply: Evaluate SPARSEUP as a lightweight, self-hostable retrieval component where explainability of match scores matters, rather than defaulting to a black-box dense embedder.

    open-sourceembeddingsretrieval

    Read more: Linkup Research Releases SPARSEUP: A 149M-Parameter Open-Source Sparse Embedding Model

11. #11 SparsePR: training-free sparse attention for video generationrepo

    A new open-source method adds training-free sparse attention to video-generation and world-model backbones (HunyuanVideo, Wan2.2-I2V, Cosmos-Predict2.5), released with code on GitHub.

    Why it matters: Training-free speedups are immediately usable against existing checkpoints without a costly fine-tuning pass, lowering the barrier to optimizing video-gen inference.

    How to apply: Teams running any of the supported video backbones can drop in SparsePR from the GitHub repo to test inference speedups before investing in custom attention-kernel work.

    open-sourcevideo-generationattention

    Read more: SparsePR: Training-Free Sparse Attention for Video Generation and World Models