Observed Signal · May 15, 2026 · Technical Guidance · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
CLAUDE.md Wastes Tokens — Use Skills Instead
A developer post argues that large, static CLAUDE.md (or agent.md) files harm agent performance and increase API cost because their full contents are injected into the model context on every conversation turn. A typical 1,000-line CLAUDE.md can be 7,000–10,000 tokens, which multiplies across multi-turn sessions. The author recommends using skill files with only name and description loaded into context (progressive disclosure), teaching workflows through successful conversational runs, then codifying skills from those runs, and iterating. The post also warns about context-window degradation as it fills and gives guidance for scaling sub-agents only after skills are battle-tested.
Practical developer guidance that reduces token costs and improves reliability of LLM-driven agents; relevant to teams building agentic workflows but not industry-shifting.
Track LinkedIn Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- A full CLAUDE.md is injected into the model context at the start of every conversation turn, not just once.
- A 1,000-line CLAUDE.md is roughly 7,000–10,000 tokens; repeated across 20–30 turns consumes significant context and cost.
- Skill files inject only the skill name and description into context; the detailed body stays on disk until the agent deems it relevant (progressive disclosure).
- Recommended workflow: run the workflow conversationally until a successful run exists, then ask the agent to generate a skill from that run and iterate.
- Filling the context window degrades model quality; keeping context lean preserves performance during long sessions.
Connected Companies & Entities
2 Entities mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
3 Hidden Token Sinks in Claude Code
A Dev.to technical follow-up by Prakash Ponali describes three additional sources of token waste in Claude Code after applying a Skill Vault pattern. Starting from a ~51K token per-session baseline, the author identified and fixed a bloated root CLAUDE.md (saving ~1.5K tokens), disabled claude-mem's SessionStart timeline auto-injection (~2K tokens), and re-vaulted 27 unused skills (~1.4K tokens). Combined, these changes shave roughly 4.9K tokens per session without losing capability. The post details file locations, audit scripts, and configuration edits (including a caution about plugin upgrades), and advocates auditing auto-loaded context and moving episodic content to on-demand access to improve LLM attention and efficiency.
How to write effective Claude Code SKILL.md
This technical guide explains the SKILL.md format for Claude Code skills, focusing on how to make skills reliably trigger. It describes the two-stage loading model (description always visible; body loads only on use), recommended frontmatter fields (name, description, disable-model-invocation, allowed-tools), argument substitution syntax, and the recommended 500-line limit with supporting files for large references or scripts. The post also covers common triggering problems and fixes, subagent contexts (context: fork) for noisy procedures, and notes that SKILL.md follows the Agent Skills open standard used by multiple coding agents. Publication date: 2026-07-22.
What to Put in Your CLAUDE.md
This tutorial explains how to write an effective CLAUDE.md for Claude Code agents. It recommends including only lines that materially change Claude's behaviour—start with a one- or two-line project description and explicit stack versions, a top-level directory map, build and test commands, non-enforceable conventions, and explicit 'do not touch' notes. It warns against including personality instructions and rules already enforced by tools. The author introduces a pragmatic "one-line test": remove a line and keep it only if its absence would cause Claude to make a mistake. The post argues brevity improves runtime reliability because CLAUDE.md is loaded into Claude's context each session. The article links to a free CLAUDE.md cheat sheet and a paid, deeper guide on configuration stacks and agent tooling.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.