Skiller: Teaching Your AI Agent to Remember
An open-source meta-skill that gives coding agents persistent memory
TL;DR
The repo is at github.com/markdav-is/Skiller. MIT licensed. The install takes about thirty seconds. If your agent has ever forgotten something you wish it hadn’t, give it a look.
Today’s AI coding agents’ biggest drawback isn’t hallucinations or cost, but their lack of memory between sessions. Every developer faces this: you spend ages debugging, your AI figures out the fix, and life moves on. Weeks later, the same issue returns—yet the agent starts fresh, oblivious to past solutions.
This results in wasted effort, with valuable discoveries lost once a session ends. Team knowledge is scattered in heads, Slack threads, or cryptic code comments. Until AI agents gain persistent memory, maintain a shared troubleshooting log or update documentation to capture and reuse insights for future problems.
I built Skiller to fix this.
What Skiller Does
Skiller is an open-source tool that gives AI coding agents persistent memory by capturing significant discoveries—like debugging breakthroughs or complex fixes—during coding sessions. It stores these insights as structured markdown “skills,” which agents automatically load in similar situations, ensuring solutions are reused efficiently. Skiller requires minimal setup: just add it to your repository alongside agents like GitHub Copilot or Cursor.
By recording valuable knowledge directly from developer actions, Skiller replaces outdated wikis or runbooks, making information instantly accessible when relevant. For example, it documents tricky dependency resolutions so future developers receive immediate, step-by-step guidance, preventing repeated mistakes.
Skiller leverages modern AI agents’ native skills systems, supporting progressive loading. As a “meta-skill,” it observes your workflow, extracts non-trivial solutions into skill files, and enables teams and projects to benefit from prior experience, promoting actionable and timely knowledge sharing.
The Quality Problem
Auto-generated documentation often leads to overloaded knowledge bases, filled with low-value entries that make it hard to find useful information. For instance, generic advice like “Restart the server if it crashes” adds clutter, while specific guidance such as “If service X fails with error Y, apply patch Z and restart only the affected module” is genuinely helpful.
Skiller combats this by evaluating every skill entry for reusability, specificity, and effectiveness before it’s saved. Only high-quality, actionable solutions are kept, ensuring the library remains practical and up-to-date.
The description field is optimized for semantic matching, allowing the system to understand user intent and surface relevant skills quickly. This focus ensures teams receive precise, actionable guidance when they need it most, rather than sifting through irrelevant content.
How It Actually Works
Skiller uses the Agent Skills standard. Skills live in a .github/skills/ directory in your repo. Each skill is a markdown file with YAML frontmatter describing what it does and when it applies. The agent reads these descriptions at startup and loads the full content when the situation matches.
The install is three curl commands. No dependencies, no build step, no npm install, no Docker container, no twelve-step onboarding ritual. Just markdown files dropped into your repo.
Special Note to Claude users
Check the Claude Code specific readme
macOS / Linux:
mkdir -p .github/skills/skiller .github/agents .github/prompts
curl -sL https://raw.githubusercontent.com/markdav-is/Skiller/main/SKILL.md \
-o .github/skills/skiller/SKILL.md
curl -sL https://raw.githubusercontent.com/markdav-is/Skiller/main/.github/agents/skiller.agent.md \
-o .github/agents/skiller.agent.md
curl -sL https://raw.githubusercontent.com/markdav-is/Skiller/main/.github/prompts/skiller.prompt.md \
-o .github/prompts/skiller.prompt.mdWindows (PowerShell):
New-Item -ItemType Directory -Force -Path .github\skills\skiller, .github\agents, .github\prompts
Invoke-WebRequest https://raw.githubusercontent.com/markdav-is/Skiller/main/SKILL.md `
-OutFile .github\skills\skiller\SKILL.md
Invoke-WebRequest https://raw.githubusercontent.com/markdav-is/Skiller/main/.github/agents/skiller.agent.md `
-OutFile .github\agents\skiller.agent.md
Invoke-WebRequest https://raw.githubusercontent.com/markdav-is/Skiller/main/.github/prompts/skiller.prompt.md `
-OutFile .github\prompts\skiller.prompt.mdIt works with GitHub Copilot in both VS Code and Visual Studio, Cursor, and any other agent that supports the .github/skills/ convention. Drop it in, start working. The skills accumulate over time.
Standing on Shoulders
Skiller originated as a fork of Claudeception, a project for Claude Code addressing similar challenges. To broaden its impact, I rebuilt Skiller for use with multiple AI agents, leveraging proven concepts like skill libraries. Research—including Voyager, CASCADE’s meta-skills, and Reflexion—shows that agents benefit from frameworks enabling learning and self-reflection. Skiller adapts these academic advances for practical use in software teams, turning theory into productivity.
One standout feature is Skiller’s recursive approach: agents continually build new skills from past experiences. As your team solves problems, Skiller captures valuable solutions as reusable skills, expanding the library over time. This process ensures agents become more effective with each task, streamlining future work and deployments.
Who This Is For
Skiller helps solo developers and teams avoid repeating past solutions by using AI-powered agents to record and share fixes for common problems. On teams, unique expertise is automatically distributed so everyone benefits from collective knowledge. For small civic organizations, Skiller ensures vital information isn’t lost when people leave, making onboarding faster and preserving operational know-how.
I’ve seen many government offices and nonprofits lose key processes when staff turn over. With Skiller, critical steps—like troubleshooting old payroll systems or handling quirks in reporting—are saved for new hires. While it can’t capture everything, Skiller makes retaining essential operational knowledge easier, helping small organizations remain resilient and efficient.
The repo is at github.com/markdav-is/Skiller. MIT licensed. The install takes about thirty seconds. If your agent has ever forgotten something you wish it hadn’t, give it a look.
The future of AI-assisted development is not agents that know everything. It is agents that remember what matters.