Observe
Run /skill-doctor in a Claude Code terminal to inspect visible skill context cost and never-invoked skills.
Navigation Explore Code Guide
15 tools mapped Stars read Sep 30, 2026 RTK · lean-ctx · Headroom · LLMLingua · Edgee · Portkey
Context engineering: Tools & ecosystem
The context window is not storage. It's a budget of attention. This page maps the tools that help you spend it well.
TL;DR
I think context engineering will be absorbed by harness engineering, as one part of how a coding agent is built and run. I also expect "context engineer" to exist as a real job title within two years, to carry that work.
The Shift
From prompt engineering to context engineering
Prompt engineering optimizes one request. Context engineering optimizes the entire information architecture: what the model knows before any request begins.
When Claude generates generic output or ignores a convention, inspect the context it actually received before blaming the model. Missing, stale or noisy context is one possible cause; model limits and failed checks are others. The useful question is which cause the evidence supports.
The tools on this page attack that problem from different angles: compressing what enters the context, filtering what shouldn't, routing intelligently, and measuring the results.
| Dimension | Prompt Engineering | Context Engineering |
|---|---|---|
| Question | How to phrase this? | What should the model know? |
| Scope | Single request | Full system |
| Data source | Static user input | RAG, memory, tool outputs |
| Goal | One good response | Reliable system at scale |
| Approach | Heuristic, artisanal | Algorithmic, systematic |
Lean applied to context
Pull the context the decision needs
Just-in-Time is an analogy here, not a claim that a smaller prompt makes a software team Lean. Start with one recurring task and the decision it must make.
Trace the source
Record the authoritative file, revision or retrieval time, and the relevant behavior. A repository path alone does not show that the agent saw current information.
Trace the use
Separate information present, generated, loaded and used. Check the resulting decision against a known failure and a nearby valid case.
Trace the effect
Compare reviewer corrections, rework and the downstream result for similar tasks. Count retrieval and maintenance cost before turning a local fix into a permanent rule.
Test the corresponding harness control and follow the full request to the user.
Three concepts that change how you think
Before picking tools, these mental models determine whether you're solving the right problem.
Minimum Viable Context
MVCProvide exactly what the model needs, nothing more. Over-context degrades adherence as fast as under-context produces hallucinations.
In practice 20 precise rules beat 200 generic ones.
Context Rot
As context length grows, models ignore information in the middle. Instructions at line 400 of a CLAUDE.md are followed ~60% as often as instructions at line 10.
In practice Run /compact at 70%, not 90%.
Semantic Priming
Ultra-compressed context works not by verbatim recall, but by activating the model's pre-trained knowledge. 10 precise keywords can outperform 100 tokens of prose.
In practice "OpenAPI 3.1, strict, no nullable" > two paragraphs.
Seven tool categories
Each category intercepts at a different point in the pipeline, from CLI output to observability layer to operational loop closure.
Output Compression
Filter CLI and tool output before it reaches the model
- RTK CLI proxy that filters shell output ★ 82K
- lean-ctx MCP server: file reads + session memory ★ 3.8K
- Headroom Compression for structured tool output
Prompt Compression
Reduce input tokens before sending to the LLM
- LLMLingua 20x compression, ~1.5% perf loss
- LLMLingua-2 Distillation-based, faster
AI Gateways
Routing, guardrails, and compression at the API layer
- Edgee Compression + routing, 15-20% in prod
- Portkey 250+ LLMs, semantic caching
RAG Optimization
Improve retrieval quality and reduce retrieval noise
- Contextual Retrieval Anthropic, 67% fewer failures
- RAG Triad Evaluation: context / answer / groundedness
Memory Systems
Persist context across sessions without flooding the window
- ICM Infinite Context Memory, dual-layer
- /compact Native Claude Code, use at 70%
LLMOps
Trace, measure, and improve your AI pipelines
- Langfuse Open-source tracing, self-hostable
- Arize Phoenix RAG Triad evaluation specialist
ACE Operations
Close the feedback loop: detect friction, propose config changes as PRs, and flag retirement candidates
- Curator (PR-based) Generates config-change PRs from typed friction signals + canary A/B
- Ejection probes Flags low-use or negative-ROI candidates; humans decide
Skill Governance 1 2 3 4
Low use is a review signal, not a deletion rule
A quiet skill may be obsolete, rarely needed, successfully preventive, or simply missing telemetry. The usage signal starts an investigation; it does not settle the decision.
Bound the evidence
The report requires v2.1.252 or later, excludes bundled and enterprise skills, depends on feature-flag fetching, and is unavailable through Remote Control.
Compare
Disable the candidate, run representative sessions with and without it, and combine those results with task frequency and failure severity.
Retire safely
Keep the retired version in Git history or outside discovered skill directories. An archive under .claude/skills/ can remain visible through nested discovery.
Human decision: account for ownership, assumptions, risk, and whether the skill encodes a temporary model gap or a durable team preference.
Skill retirement checklistBenchmarks worth knowing
Each figure names the quantity it measures. A saving on one layer, such as shell output or cached input, is not the same number as the cost of a whole task.
+6.8%
billed cost, with 38.4% fewer tool-output tokens
arXiv 2607.12161, largest compression setup against unmodified Claude Code. Fewer tokens on one layer did not mean a smaller bill
20x
prompt compression
LLMLingua, 1.5% performance loss on GSM8K & BBH
67%
fewer RAG failures
Anthropic Contextual Retrieval + BM25 + reranking
90-97.5%
lower price on cached input tokens
Anthropic prompt caching: a cache read costs 10% of the base input price on most current models, 5% on Opus 5.5 and 2.5% on Fable 5.1
<4%
GPU memory waste
vLLM PagedAttention (vs 60–80% with naive allocation)
2.5x
TTFT acceleration
SlimInfer dynamic token pruning on LLaMA 3.1
Eight public benchmarks measured token-saving tools end to end, and most results sit far below vendor claims. The site author is a core contributor to RTK. Compare claims and measurements →
Tool selection by use case
Which tools matter depends on where you sit in the stack.
Claude Code developer
| Commands flooding context | RTK |
| Large files read repeatedly in session | lean-ctx (signatures mode) |
| Context growing too long | /compact at 70% |
| Rules ignored in large CLAUDE.md | Path-scoping |
| Session memory lost between runs | lean-ctx CCP |
| Cross-session memory | ICM |
AI application builder
| Tool output JSON too verbose | Headroom |
| Prompts too long | LLMLingua |
| Multi-provider routing | Portkey |
| RAG chunks losing context | Contextual Retrieval |
| Compression + routing for coding agents | Edgee |
| RAG quality measurement | Arize Phoenix |
Self-hosted LLM deployer
| GPU memory fragmentation | vLLM (PagedAttention) |
| Shared-prefix cache reuse | SGLang (RadixAttention) |
| Repeated query caching | Redis semantic cache |
"Context engineering is the art of filling the context window with the right information at the right time."Andrej Karpathy
Three words carry the weight: right information (not all information), right time (not always-on for everything). The tools on this page are the mechanics behind making that "right" a system property rather than a manual judgment call.
Full reference in the guide
The guide covers every tool with install instructions, measured benchmarks, and when-to-choose-what guidance, including LLMOps tools (Langfuse, LangSmith, Arize Phoenix), KV cache infrastructure (vLLM, SGLang), and the experimental catalog (SlimInfer, TopV, AttnComp).