9 Essential Claude Skills for AI Engineers Building Production Agents + 1 Bonus Skill

Claude Code & Coding AgentsMaximem Team2 February 202620 min read
9 Essential Claude Skills for AI Engineers Building Production Agents + 1 Bonus Skill
On this page
  1. What a Claude skill is, and what changed on 22 October 2026
  2. The nine skills at a glance
  3. The essential 9 skills
  4. Bonus skill: synap, persistent memory for agents (built by us)
  5. Which skill for which stage of an agent's life
  6. How to install Claude skills in October 2026
  7. How to check a skill before you install it
  8. Writing your own SKILL.md
  9. Claude skills for mechanical, civil and other engineering
  10. What skills still do not do
  11. Frequently asked questions
  12. Resources

Published 2 February 2026. Updated 29 October 2026: every repository below was re-checked against the GitHub API on 29 October 2026, and every star count and push date is as of that morning.

Three of the nine Claude skills we recommended to AI engineers in February no longer exist at the address we gave you. The prompt-injection skill went when its parent repository was rebuilt as LifeOS, and the prompt-engineering and agent-debugging skills now return a 404 because their repositories were deleted. This update replaces those three with working skills that do the same jobs, moves two more to the repositories they migrated to, and corrects the install commands, because Claude Code installs skills through /plugin and a skills folder, not through the claude skills install command the February version showed. The nine below cover the work that decides whether an agent survives production: evaluating it, securing it, structuring its prompts and catching regressions before a merge.

Ownership of the format changed too. On 22 October 2026 Anthropic donated Agent Skills to the Agentic AI Foundation, the Linux Foundation project that already hosts the Model Context Protocol, and reported that more than 70 agent platforms have integrated the standard, that 1 million skills have been published and that nearly 280 million have been installed (Anthropic's own figures). For you that means a skill written for Claude Code is now a vendor-neutral artifact. If you want the twelve skills that shipped after February, such as Trail of Bits, Superpowers and the Langfuse and LangSmith tracing plugins, they have their own list in Best Claude Code Skills in 2026; this page stays with the core toolkit.

What a Claude skill is, and what changed on 22 October 2026

A Claude skill is a folder containing a SKILL.md file, which holds YAML frontmatter (a name and a description) followed by markdown instructions, plus any scripts, reference documents or templates the instructions point to. Claude reads only the name and description of every installed skill at the start of a session, roughly 100 tokens each according to the Agent Skills specification, and loads the full instructions only when a task matches the description. Supporting files load later still, and only if the instructions call for them. That staged loading is why you can install dozens of skills without paying for them on every turn.

Claude Code runs a skill in two ways, as its skills documentation describes: automatically, when your request matches the skill's description, or directly, when you type /skill-name. A skill installed through a plugin is called as /plugin-name:skill-name. The @skill-name syntax in the February version of this page was never how Claude Code invokes a skill.

One folder now works well beyond Claude. Anthropic launched skills in October 2025, published the format as an open standard in December 2025, and on 22 October 2026 handed the name, the logo, the GitHub organisation and the agentskills.io domain to the Agentic AI Foundation, which Anthropic co-founded with Block and OpenAI. Engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals, with David Soria Parra and Mahesh Murag as lead maintainers, and the proposed roadmap adds versioned releases, a formal proposal process and security guidance for skill authors and installers. Nothing in the format changed on the day of the donation, so every skill on this page keeps working.

The nine skills at a glance

#SkillJob in a production agentRepositoryStarsLast pushStatus since February
1promptfoo-evaluationPrompt and model regression tests with Promptfoodaymade/claude-code-skills1,50229 Oct 2026Same skill, still maintained
2ai-securityPrompt-injection, jailbreak and agent tool-abuse assessmentalirezarezvani/claude-skills28,93027 Oct 2026Replaces promptinjection (removed)
3evaluating-llmsMetrics, LLM-as-judge and RAG validationancoleman/ai-design-components52511 Dec 2025Same skill, repository idle since December
4senior-prompt-engineerEval-driven prompt iteration and structured-output contractsalirezarezvani/claude-skills28,93027 Oct 2026Replaces ai-prompt-engineering (repository deleted)
5ai-threat-testingOWASP Top 10 for LLM applications, offensive testingtransilienceai/communitytools59929 Jul 2026Same skill
6llm-evaluationHallucination detection and benchmark creationstefan-jansen/claude-code-toolkit907 Jul 2026Same skill, repository moved
7llm-patternsArchitecture patterns for apps where an LLM runs core logicalinaqi/maggy71013 Oct 2026Same skill, repository renamed
8coding-agentDelegating coding work to Claude Code, Codex or OpenCode as background workersopenclaw/openclaw393,31329 Oct 2026Same skill, runs inside OpenClaw
9agent-regression-evalsStatistical eval gate that blocks a merge when an agent gets worsealirezarezvani/claude-skills28,93027 Oct 2026Replaces agent-debugger (repository deleted)
BonussynapPersistent, per-user memory for agents (built by us)maximem-ai/maximem_synap_sdk7729 Oct 2026Same skill

Stars and push dates are for the whole repository, read from the GitHub API on 29 October 2026, so a large repository's star count says little about the single skill inside it.

The essential 9 skills

1. promptfoo-evaluation

What it does: Configures and runs LLM evaluations with Promptfoo, the open-source CLI for testing and comparing model outputs. The skill writes promptfooconfig.yaml files, Python custom assertions and llm-rubric LLM-as-judge checks, and it manages few-shot examples inside your prompts.

Use it when: you are about to change a prompt or swap a model and need proof the agent did not get worse, or you want an A/B comparison of two prompts with numbers attached.

Why it is on the list: It was our first pick in February and it is the strongest survivor. The repository has grown from 531 stars then to 1,502 now and was pushed on the morning of this update, which makes it the most actively maintained evaluation skill we checked.

Install:

/plugin marketplace add daymade/claude-code-skills
/plugin install promptfoo-evaluation@daymade-skills

Try: Create a promptfoo suite for my support agent with ten edge cases and an llm-rubric check for invented refund policies.

Source: daymade/claude-code-skills (MIT).

2. ai-security

What it does: Assesses AI and LLM systems for prompt injection, jailbreaks, model inversion risk, data poisoning exposure and agent tool abuse, maps findings to MITRE ATLAS techniques, and scores adversarial robustness. Its own instructions separate it from general application security, which lives in a sibling security-pen-testing skill.

Use it when: an agent reads untrusted text (uploaded documents, web pages, tickets, email) and can call tools, which is the combination where one injected instruction becomes an action.

Why it is on the list: It replaces the February pick, Daniel Miessler's promptinjection skill, which disappeared when his Personal AI Infrastructure repository became LifeOS; the LifeOS skill list we pulled on 29 October has no prompt-injection skill. ai-security covers the same ground and sits in one of the largest maintained skill collections, at 28,930 stars and pushed two days before this update.

Install (adds the engineering-team bundle, which also contains skill 4):

/plugin marketplace add alirezarezvani/claude-skills
/plugin install engineering-skills@claude-code-skills

Try: Run an ai-security assessment on my RAG agent. Focus on indirect injection through uploaded PDFs and on which tools an injected instruction could reach.

Source: ai-security SKILL.md (MIT).

3. evaluating-llms

What it does: Evaluates LLM systems with automated metrics, LLM-as-judge patterns and standard benchmarks, and covers prompt quality, RAG pipeline validation, safety checks for hallucination and bias, and model comparison for deployment decisions.

Use it when: you are choosing between models for a specific task, or you need a first measurement of how often your RAG pipeline answers from the wrong document.

Why it is on the list: It is the broadest evaluation skill on the list and pairs well with promptfoo-evaluation: this skill designs the evaluation and promptfoo runs it. One flag we will not hide: the repository has not been pushed since 11 December 2025, so check any library versions it suggests against current releases.

Install:

npx degit ancoleman/ai-design-components/skills/evaluating-llms ~/.claude/skills/evaluating-llms

Try: Compare two models on fifty of my support tickets and report accuracy, latency and cost per ticket.

Source: ancoleman/ai-design-components (MIT).

4. senior-prompt-engineer

What it does: Eval-driven prompt engineering. It designs prompt templates and structured-output contracts, measures RAG retrieval quality on relevance, faithfulness and coverage, validates agent and tool configurations, and budgets tokens and cost, using three Python tools that need nothing beyond the standard library. It is deliberately model-agnostic and never hardcodes model IDs or prices, so you supply your provider's current rates when you want dollar figures.

Use it when: a prompt works in the playground and fails on real inputs, or you need a JSON contract the rest of your system can depend on.

Why it is on the list: It replaces ai-prompt-engineering by vasilyu1983, whose repository returned a 404 on 29 October 2026. The replacement is stronger on the part that matters in production, because it ties every prompt change to an eval set instead of to a pattern catalogue.

Install: included in the engineering-team bundle from skill 2.

Try: Design an extraction prompt for invoices with a strict JSON schema, then build a twenty-case eval set and report faithfulness.

Source: senior-prompt-engineer SKILL.md (MIT).

5. ai-threat-testing

What it does: An offensive AI security testing framework that runs ten specialised agents against the OWASP Top 10 for LLM Applications, including prompt injection, model extraction, data poisoning and supply-chain attacks, and plugs into a wider penetration-testing workflow.

Use it when: you have permission to attack the system and need a structured pre-launch assessment, not a checklist. The skill's own instructions limit it to authorised assessments.

Why it is on the list: ai-security (skill 2) finds weaknesses; this one tries to exploit them. The repository is maintained by Transilience AI, was last pushed on 29 July 2026, and its README includes a script that runs it inside a Kali Linux container, which is the isolated environment it should have.

Install:

npx degit transilienceai/communitytools/skills/ai-threat-testing ~/.claude/skills/ai-threat-testing

Try: Run an OWASP LLM Top 10 assessment against our staging chatbot and rank findings by exploitability.

Source: transilienceai/communitytools (MIT).

6. llm-evaluation

What it does: Evaluation and testing patterns for LLM applications: prompt testing, output validation, hallucination detection, benchmark creation, A/B testing and quality metrics.

Use it when: you need an internal benchmark that tracks one agent over months, or a hallucination suite built from adversarial queries.

Why it is on the list: It is the most compact of the three evaluation skills here and reads well as a reference. Its repository moved from applied-artificial-intelligence to stefan-jansen/claude-code-toolkit (GitHub redirects the old address), so update any pinned paths.

Install:

npx degit stefan-jansen/claude-code-toolkit/skills/llm-evaluation ~/.claude/skills/llm-evaluation

Try: Build a hallucination suite for my RAG pipeline with twenty queries designed to draw answers the documents do not support.

7. llm-patterns

What it does: Patterns for AI-first applications, where an LLM handles core logic such as classification, extraction and generation, together with LLM testing and prompt management.

Use it when: you are deciding how much of a feature to hand to the model and how to test the part you did.

Why it is on the list: It is a pattern reference, not a task runner, which suits the design stage. Two changes since February: the repository was renamed from claude-bootstrap to alinaqi/maggy, and the skill now sets user-invocable: false, so Claude loads it automatically when the topic comes up and it does not appear in your / menu.

Install:

npx degit alinaqi/maggy/skills/llm-patterns ~/.claude/skills/llm-patterns

8. coding-agent

What it does: Delegates coding work to Codex, Claude Code or OpenCode running as background workers, for tasks larger than a simple edit or a read-only lookup.

Use it when: you run agents from scripts or CI and want one agent to farm out coding jobs to others.

Why it is on the list: It lives in OpenClaw, the largest repository on this page at 393,313 stars. Be clear about where it runs: it is an OpenClaw skill, it needs the claude, codex or opencode binary on the machine, and it is switched on through the skills.entries.coding-agent.enabled setting in OpenClaw's configuration; you do not install it into Claude Code. If you do not run OpenClaw, skip it.

Install: ships with OpenClaw; enable it in OpenClaw's configuration.

9. agent-regression-evals

What it does: Proves that an agent, prompt, model or tool change did not make the agent worse. It runs repeated trials, reports pass@k and pass^k, compares a candidate against a baseline with paired statistics, grades tool-call trajectories, finds flaky eval cases, and fails a CI gate when a case drops, a critical case regresses or the agent makes a forbidden tool call.

Use it when: you want an eval result to block a merge instead of producing a report nobody reads.

Why it is on the list: It replaces agent-debugger by AvivK5498, whose repository returned a 404 on 29 October 2026. We chose a regression gate over another debugger because trace-driven debugging has moved to first-party observability plugins, covered in our list of newer skills, and because a single eval run cannot tell a real drop from noise, which is the problem this skill's paired statistics address. For the methodology behind it, see our guide to agent evals.

Install:

/plugin marketplace add alirezarezvani/claude-skills
/plugin install agent-regression-evals@claude-code-skills

Try: Compare baseline.jsonl and candidate.jsonl from my agent's eval run and tell me whether the pass-rate drop is real or noise.

Source: agent-regression-evals SKILL.md (MIT).

Bonus skill: synap, persistent memory for agents (built by us)

Disclosure first: this is our skill, and Maximem Synap is our product.

What it does: Teaches a coding agent (Claude Code, Cursor, Codex and others) to add Maximem Synap to your application. Synap is a memory service for AI agents: your application sends it conversations as they happen, and when the same person returns it asks Synap what it knows and gets back a short, ranked set of facts ready for the prompt. The skill covers SDK setup, scoping memory per user, customer or organisation, ingestion, retrieval and the adapter package for each of the 23 agent frameworks Synap supports, including LangChain, LangGraph, LlamaIndex, CrewAI, LiveKit, Pipecat and the OpenAI and Claude agent SDKs. Its first step checks whether memory fits your project at all and says so when it does not, because for single-turn calls with no cross-session state a memory layer is overkill.

Use it when: your agent needs to remember a user across sessions, or several tenants need memory that never leaks between them.

Install:

npx skills add maximem-ai/maximem_synap_sdk --skill synap -y

Try: Add Synap long-term memory to my LangGraph support agent, scoped per user, with retrieval before each turn.

Source: maximem_synap_sdk/skills (Apache-2.0; 77 stars and pushed on 29 October 2026, a small repository stated plainly) and the setup guide for coding agents.

Which skill for which stage of an agent's life

StageQuestion you are answeringSkill
DesignWhich parts of this feature should the model own?llm-patterns
PromptingWill this prompt hold its output contract on real inputs?senior-prompt-engineer
First measurementHow good is it, and against which model?evaluating-llms, llm-evaluation
Regression testingDid this change make it worse?promptfoo-evaluation, then agent-regression-evals as the CI gate
Security reviewWhat can an attacker make it do?ai-security, then ai-threat-testing in an isolated environment
AutomationCan other agents do the coding work?coding-agent (OpenClaw only)
StateDoes it remember the right things about each user?synap

Order matters more than selection. A team that runs ai-security before it has an eval suite cannot tell whether its fix for an injection broke the happy path, so build the regression suite first and let every later change, including a security fix, pass through it. In Claude Code the chain looks like four requests in one session: design the prompt with senior-prompt-engineer, generate the suite with promptfoo-evaluation, attack it with ai-security, and gate the result with agent-regression-evals.

How to install Claude skills in October 2026

Claude Code, from a plugin marketplace. Add the marketplace once, then install a plugin from it. Anthropic's own directory works the same way, for example /plugin marketplace add anthropics/claude-plugins-official followed by /plugin install skill-creator@claude-plugins-official. Run /reload-plugins after installing, according to the Claude Code skills documentation.

Claude Code, by hand. Put the skill folder at ~/.claude/skills/<skill-name>/SKILL.md to use it in every project, or at .claude/skills/<skill-name>/SKILL.md inside a repository to share it with everyone who clones that repository. npx degit owner/repo/path/to/skill ~/.claude/skills/<skill-name> copies a single skill folder out of a larger repository without its git history. Claude Code watches these folders and picks up edits within the session; if you create a new top-level skills folder mid-session, run /reload-skills.

Claude Code, checking what is installed. /skills lists available skills, and /skill-doctor (Claude Code 2.1.252 and later) reports each skill's context cost and usage.

Claude.ai and the Claude desktop app. Skills are available on the Free, Pro, Max, Team and Enterprise plans and need code execution switched on, per Anthropic's help article. Turn individual skills on under Customize > Skills, and upload your own as a ZIP of the skill folder through the "+" button. Owners of Team and Enterprise organisations can provision a skill for every member, and it then appears in each member's list with a team indicator.

Other agents. Because the format is a shared standard, the same folder works in the agents that read it, which Anthropic says now number more than 70, including ChatGPT, Cursor, GitHub Copilot and VS Code. The npx skills add <owner/repo> --skill <name> command from skills.sh, which the synap setup guide uses, installs a skill straight from a GitHub repository.

How to check a skill before you install it

A skill can carry scripts that run on your machine with your permissions, and its instructions reach the model with the weight of your own request, so a third-party skill deserves the review you would give a dependency with shell access. Our research summary on agent skill security reports that a static scanner flagged 2,414 of 138,133 public skill files (1.75%), a lower bound because it read only the SKILL.md text and not the scripts beside it.

Five checks catch most problems, and the first two would have flagged every change on this list since February:

  • Does the repository still resolve, and under which name? Two of our February picks now live under new names and three are gone. A redirect is fine; a 404 is not.
  • When was it last pushed? evaluating-llms has been idle since December 2025. Idle is not broken, but it means nobody is updating it for new model or library versions.
  • What does every script do? Read the scripts folder before you read the instructions, and look for network calls, downloads piped into a shell and reads of credential files.
  • Which tools does it pre-approve? An allowed-tools line in the frontmatter grants those tools without asking you during the turn the skill runs.
  • Can you pin it? Copy a known commit into your skills folder instead of tracking a moving branch, so the skill you reviewed is the skill you run.

If your agents also keep memory, the same supply-chain logic applies to what they write down; our guide to memory poisoning covers that side.

Writing your own SKILL.md

At minimum you need a folder whose name matches the skill, containing a SKILL.md that starts with this frontmatter:

---
name: release-notes
description: Drafts release notes from merged pull requests since the last tag. Use when the user asks for release notes, a changelog entry or a summary of what shipped.
---

Two fields are mandatory under the Agent Skills specification: name (1 to 64 characters, lowercase letters, digits and single hyphens, matching the folder name) and description (1 to 1,024 characters), and the reference validator checks both with skills-ref validate ./my-skill. Claude Code is looser on the name, which defaults to the folder name, and adds optional fields such as disable-model-invocation, allowed-tools, paths and context: fork, which runs the skill in a subagent.

Three rules decide whether a skill gets used. The description is the only part Claude sees before choosing, so put the use case first and name the words a user would actually type; Claude Code truncates the combined description and when_to_use text at 1,536 characters in its listing. Keep SKILL.md under 500 lines and move reference material into separate files, because the whole body loads on activation and the specification recommends under 5,000 tokens for it. Write the skill around a procedure rather than a set of facts: either it adds a capability Claude lacks, usually with a script, or it encodes how your team does something Claude can already do, and both kinds age well. A capability skill such as ai-threat-testing ships code that does something Claude cannot do from instructions alone, while a preference skill such as the release-notes example above is usually a single SKILL.md describing how your team wants a familiar task done. Facts that change, such as a customer's plan or last week's decision, go stale in a file nobody updates. How to write a SKILL.md that actually gets selected goes further on selection, and Anthropic's skill-creator skill will draft one for you.

Claude skills for mechanical, civil and other engineering

Most of what appears under "engineering" in skill directories is software engineering, and that is also true of this page. For physical engineering the most complete collection we found is Soljourner/claude-engineering-skills, which holds 36 SKILL.md files for mechanical, aerospace and general engineering: structural analysis, fluid dynamics, thermodynamics, pump design and cavitation analysis, unit conversion, Python packages such as SciPy, SymPy and Pint, and integrations for SolidWorks, ANSYS, COMSOL and OpenFOAM. It is MIT-licensed with 95 stars, and it has not been pushed since 7 November 2025, so treat its tool integrations as a starting point and check them against the versions you run. For electronics, the alirezarezvani collection used above includes an embedded-iot-mentor skill.

A skill does not make Claude a licensed engineer. It gives the model a procedure and, where it ships scripts, a calculator; the engineering judgment, the safety factors and the sign-off stay with you.

What skills still do not do

Skills hold procedure well and state badly. A skill file is the same for every user and every session, which is right for "how we run an eval" and wrong for "what this customer told us on Tuesday". Researchers who labelled 528 matched sets of runs from the SKILL.md literature found that a skill was steadying the agent's procedure 65.7% of the time and supplying a fact the model lacked 4.5% of the time, as we summarised in We do not need a memory layer. We have skills. That is a third-party labelling result, not a Maximem measurement, but it matches what the format is built for.

Skills are also not a replacement for MCP. MCP gives an agent access to a system (a database, a ticket queue, an API), while a skill tells it how to use that access; the two now sit in the same foundation and are designed to be used together. And skills do not give your own product's agents memory of their users. Claude Code's own memory files cover a developer's preferences inside a repository, as our comparison of coding-agent memory explains, but a support or sales agent that has to remember thousands of end users across sessions needs a memory layer behind it, which is the job the bonus skill sets up.

Frequently asked questions

What are the best Claude skills for AI engineers?

For an engineer building production agents, start with an evaluation skill (promptfoo-evaluation), a regression gate (agent-regression-evals) and an AI security skill (ai-security), because those three decide whether a change is safe to ship. Add senior-prompt-engineer when prompts are the bottleneck and ai-threat-testing before launch. All nine on this page were checked against GitHub on 29 October 2026.

Which February skills no longer work?

promptinjection (Daniel Miessler), ai-prompt-engineering (vasilyu1983) and agent-debugger (AvivK5498) are gone; this update replaces them with ai-security, senior-prompt-engineer and agent-regression-evals. llm-evaluation and llm-patterns still exist but moved to stefan-jansen/claude-code-toolkit and alinaqi/maggy.

How do I install a Claude skill?

Run /plugin marketplace add <owner/repo> in Claude Code and then /plugin install <plugin>@<marketplace>, or copy the skill folder into ~/.claude/skills/. In Claude.ai, upload a ZIP of the folder under Customize > Skills. There is no claude skills install command.

Do Claude skills work in Cursor, Codex and GitHub Copilot?

They do, for skills that follow the Agent Skills standard. Anthropic reports more than 70 platforms have integrated it, naming ChatGPT, Cursor, GitHub Copilot and VS Code. Fields such as context: fork and hooks are Claude Code extensions the specification does not define, so other agents may ignore them.

What is the Agentic AI Foundation, and why does it matter for skills?

It is the Linux Foundation project, co-founded by Anthropic, Block and OpenAI, that hosts the Model Context Protocol and, since 22 October 2026, Agent Skills. Anyone can use or contribute to the standard without joining, and the format did not change with the move.

Claude skills vs MCP: which do I need?

Usually both. MCP connects an agent to an external system; a skill teaches it the procedure for using that system well.

Are Claude skills free?

Skill files are free and open, under each repository's licence (most on this page are MIT). In Claude.ai, skills are available on every plan including Free, with code execution switched on.

Can several skills be active at once?

Several can. Claude sees every installed skill's description and can load more than one in a task. After Claude Code compacts a long conversation, it re-attaches the first 5,000 tokens of each skill used, within a combined budget of 25,000 tokens.

Are there Claude skills for mechanical engineering?

A small number exist. Soljourner/claude-engineering-skills has 36 skills covering structural analysis, fluid dynamics, thermodynamics and pump design, plus integrations for SolidWorks, ANSYS, COMSOL and OpenFOAM. It was last updated in November 2025.

Resources

Official documentation

Skill collections

Security

Star counts on this page are as of 29 October 2026. We re-check this list when a repository on it moves, dies or is replaced by something better.

From the team at Maximem

Stop rebuilding agent memory from scratch

Maximem Synap is the context management layer we built after hitting every problem in this post ourselves. Persistent recall across sessions, entity resolution and conscious forgetting, in Python, TypeScript and REST.

Looking for more to read? See our directory of the best engineering blogs to follow.