Short Context Agent
Keep agentic work inside a short context window using the NVIDIA "You don't need long context" recipe (arXiv:2611.05306) — forward-looking handoff briefs instead of generic summaries, End-of-Conversation / Single-Action compaction, and selective reading of large inputs with grep + bounded windows instead of pasting whole files into the prompt. Use this skill whenever a task is long-horizon (multi-step debugging, migrations, multi-file refactors, long investigations or research), whenever the user mentions context limits, compaction, /compact, handoff, checkpointing, "resume where we left off",
Install
/plugin marketplace add che-shr-cat/short-context-agentThe repo is a Claude Code plugin marketplace. Run /plugin afterwards to pick the plugin that includes this skill.
npx skills add che-shr-cat/short-context-agent --skill short-context-agentgit clone --depth 1 https://github.com/che-shr-cat/short-context-agent.git /tmp/short-context-agent
mkdir -p ~/.claude/skills
cp -r /tmp/short-context-agent/skills/short-context-agent ~/.claude/skills/short-context-agentPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# short-context-agent — compaction and selective reading protocol This skill turns the paper's harness-level recipe into something you can run today. The paper's claim: an agent in a sequential loop never needs its whole past conversation, nor a whole document, in working memory. It needs (1) a compact, forward-looking description of where it stands and (2) a tool to fetch raw data on demand. Done right, accuracy holds flat or improves (stale exploration traces stop distracting the model), and peak context drops by an order of magnitude. Three pieces, in order of how often you'll use them: | Piece | Paper concept | What you do here | |---|---|---| | A. Selective reading | LCW2: long inputs → filesystem + `grep` + bounded `doc-read` windows | Never load a big file whole. Search, then read a window. | | B. Handoff brief | EoCC / SAC compaction prompt (six-part execution brief) | Write `.claude/handoff.md` at milestones and before compaction. | | C. Resume | Fresh conversation = system prompt + task + brief | Continue from the brief without re-deriving or re-trying dead ends. | Context is a cache. State lives in text. If something matters past this turn, it has to be in a file or in the brief. ## A. Selective reading — never paste the document Applies to any input that would cost more than ~2K tokens to load whole: documents, logs, transcripts, large source files, datasets.
Details
- Skill name
short-context-agent- Repository
- che-shr-cat/short-context-agent
- Path
skills/short-context-agent/SKILL.md- License
- MIT
- Repo updated
- 5 November 2026
FAQ
What does the Short Context Agent skill do?
Keep agentic work inside a short context window using the NVIDIA "You don't need long context" recipe (arXiv:2611.05306) — forward-looking handoff briefs instead of generic summaries, End-of-Conversation / Single-Action compaction, and selective reading of large inputs with grep + bounded windows instead of pasting whole files into the prompt. Use this skill whenever a task is long-horizon (multi-step debugging, migrations, multi-file refactors, long investigations or research), whenever the user mentions context limits, compaction, /compact, handoff, checkpointing, "resume where we left off",
How do I install the Short Context Agent skill?
Claude Code plugin: /plugin marketplace add che-shr-cat/short-context-agent. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Short Context Agent popular?
It lives in che-shr-cat/short-context-agent, which has 5 GitHub stars and 4 forks.
Related skills
De AI Writing
9processmission/agent-skills
Detect, mark, and remove AI writing tics from Chinese and English drafts, so the text reads normally: denser, clearer, free of phrasing that obstructs reading. Use to de-AI a text, strip filler or clichés, or clear editing notes out of the body. Applies language- and model-specific rule tables and keeps the editing process out of the article. It rewrites style; personality and authorship stay out of scope.
Prompt Lab
10charlie947/skills-for-everyone
Turn an idea for a prompt into a finished, tested prompt in one turn. You type the idea, and the prompt comes back in a code block, graded out of 100 against Anthropic's current published prompting guidance (fetched live every run) and a plain bar any reader can run. The pass mark is 95, and Claude runs the prompt itself before handing it over whenever it can. Use when the user wants a prompt written, improved or checked. Triggers on "prompt lab", "write me a prompt", "make this prompt better", "grade this prompt", "score this prompt", "is this prompt any good", "make this prompt expert level"
Claude Club Motion Sites
6Samin12/claude-club-motion-library-skill
Use when designing or building animated websites, landing pages, mobile apps, page sections, motion backgrounds, gradients, or design systems. Searches and applies the Claude Club library of 547 responsive motion recipes, including all 341 owner-authorized MotionSites prompts, matching HTML previews, imagery references, and reusable implementation patterns.
Improve System
10charlie947/skills-for-everyone
Mine your own Claude Code history for what keeps going wrong, then fix it at the source. Finds corrections you had to give more than once (a rule should have stopped them), jobs you keep typing by hand (a missing skill), context you keep pasting in (a missing fact), and the kinds of work eating the most hours. Use when the user says "improve the system", "improve my setup", "what am I correcting", "what keeps going wrong", "mine my history", "what skills should I build", "why do I keep repeating myself", or once or twice a week as a routine. Also use with a topic to audit one kind of work that
Connection Doctor
9duvoai/skills
Diagnose and health-check the Connections and credentials a Duvo Agent relies on. Use when the user asks whether a Connection is set up or available, why an Agent is blocked on authentication, what a Connection can do, or wants a proactive check across an Agent or a folder of Agents. In Ask Duvo, for an Automation, starts from getAutomationReadiness. Reads each Build's Connections and their live state via the Duvo public API, names the problem, and proposes one concrete fix — handing off to run-debugger when the evidence lives in a single failing Run.
Anti AI Slop Design
9prangishviliAbe/agent-skills
Design, refine, implement, or critique websites, landing pages, dashboards, apps, and components so nothing looks AI-generated or templated. Use for "looks AI", "too generic", "make it premium", "make it look better", art direction, redesigns, visual polish, a page or component designed from scratch, or a review of how a UI looks, including Georgian and multilingual layouts. Covers every layer where slop appears: page structure, typography, color, spacing, surfaces, icons, imagery, copy, motion, and the code that produces them. Replaces default choices with decisions derived from the subject,
Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.