8 Best OpenCode Skills for More Reliable Coding Agents
TL;DR
The best OpenCode Skills improve process reliability, not just code generation. Start with planning, debugging, verification, and review Skills before adding framework-specific packs.
Obra Superpowers is the strongest general-purpose starting point.** It packages repeatable workflows for brainstorming, test-driven development, systematic debugging, review, and completion checks.
Vercel React Best Practices is a better choice than a generic frontend prompt for React and Next.js repositories.** Its rules are narrower, easier to test, and maintained by the team closest to that ecosystem.
A Skill should be evaluated by trigger accuracy, evidence quality, maintenance, permission scope, and measurable output—not repository stars alone.
Coding agents still need current information outside the repository. A bounded web research Skill can use Nstproxy Crawl to retrieve selected documentation and public pages while keeping source validation inside the workflow.
Best OpenCode Skills at a Glance
The best OpenCode Skills are Obra Superpowers, Vercel React Best Practices, frontend-design, agent-browser, a repository-specific security review Skill, a database migration review Skill, a release documentation Skill, and a bounded web research Skill. This ranking prioritizes whether a Skill changes the quality of a real development decision, not how many instructions it contains.
OpenCode discovers Agent Skills from project and user directories, reads their SKILL.md metadata, and exposes matching Skills through its skill tool. The official OpenCode Agent Skills documentation also supports permission patterns, so teams can allow trusted Skills, deny internal ones, or require approval for experimental packs. When a Skill needs current public documentation rather than only repository context, Nstproxy Crawl can provide the bounded retrieval layer.
Connected planning, TDD, debugging, review, and verification workflows
Opinionated process can feel heavy for small changes
2
Vercel React Best Practices
React and Next.js work
Specific, prioritized performance guidance
Poor fit outside the React ecosystem
3
frontend-design
UI implementation
Forces visual intent and component-level review
Quality varies by the pack and project design system
4
agent-browser
Browser-based validation
Gives the agent a repeatable inspection and interaction contract
Requires careful permission and authentication boundaries
5
Security review
Threat-focused change review
Applies repository-specific security checks before merge
Generic security prompts create false confidence
6
Database migration review
Schema and data changes
Checks reversibility, locks, backfills, and rollout order
Must be customized to the actual database and deployment model
7
Release documentation
Changelogs and release readiness
Converts diffs and tests into a consistent release artifact
Cannot decide product significance without team context
8
Web research and Crawl
Current technical research
Retrieves source-linked public pages beyond the codebase
Still requires source selection and claim validation
How the Ranking Was Evaluated
The ranking uses five fields that materially affect whether an OpenCode Skill belongs in a working repository.
Trigger precision: The description should activate for a recognizable job and remain quiet for unrelated work.
Evidence contract: The Skill should say what files, tests, documentation, or live surfaces must support its conclusion.
Failure behavior: A useful Skill stops or qualifies the result when prerequisites, credentials, tests, or evidence are missing.
Permission scope: The Skill should request only the tools and external access needed for its outcome.
Maintenance ownership: The instructions need an identifiable source and a practical update path when frameworks or team policies change.
The Awesome OpenCode directory is useful for discovery, but inclusion in a collection is not proof that a Skill is safe, current, or suitable for a repository. Read the SKILL.md, supporting scripts, license, and recent changes before installation.
1. Obra Superpowers: Best Overall OpenCode Skill Pack
Obra Superpowers is the best starting pack for teams that want a coding agent to follow an engineering process rather than jump directly to edits. The official Superpowers repository includes Skills for brainstorming, planning, test-driven development, systematic debugging, code review, worktrees, and verification before completion.
Choose it when: Agents frequently make plausible changes before establishing requirements, reproducing bugs, or checking results.
Why it ranks first: The Skills reinforce one another. Planning produces an executable path; implementation invokes tests; debugging requires root-cause work; completion requires evidence.
Trade-off: Superpowers is intentionally prescriptive. A team should decide which workflows are mandatory and which are excessive for trivial documentation or configuration changes.
A sensible pilot is one defect and one feature. Measure whether the pack reduces reverted edits, unverified completion claims, and review churn rather than judging it by how polished the agent's prose sounds.
2. Vercel React Best Practices: Best for React and Next.js
Vercel React Best Practices is the best specialized OpenCode Skill for repositories built around React or Next.js. The official Vercel Agent Skills collection organizes performance guidance into prioritized rules covering areas such as request waterfalls, bundle size, rendering, and client behavior.
Choose it when: The repository uses React or Next.js and performance regressions are a recurring review concern.
Why it ranks highly: The rules are maintained close to the framework ecosystem and are more actionable than a broad instruction such as “write performant React.”
Trade-off: Framework-specific advice can conflict with a repository's architecture, older runtime, or deliberate product trade-offs. The Skill should inspect the installed framework version and project conventions before proposing changes.
Do not install it as a universal frontend authority. Pair it with repository tests and performance measurements, and require the agent to distinguish verified regressions from theoretical optimization.
3. frontend-design: Best for Deliberate UI Work
A focused frontend-design Skill is valuable when agents produce functional but visually generic interfaces. The best versions require the agent to inspect the existing design language, state a visual direction, reuse tokens and components, and review the rendered result at relevant breakpoints.
Choose it when: The failure is design inconsistency, not missing React syntax.
Why it matters: UI quality cannot be verified from source code alone. A design Skill connects intent, implementation, and rendered inspection.
Trade-off: Community packs vary widely, and some replace one set of clichés with another. Prefer a Skill adapted to the team's design system over a long universal style manifesto.
The Skill should require screenshots or a live browser check, keyboard navigation, visible focus, error and empty states, and a comparison against the existing product. For related browser-centered workflows, agent tools and their safety boundaries provide a useful design reference.
Give OpenCode Skills Current Web Evidence
Use Nstproxy Crawl to retrieve selected public documentation and research pages as structured, source-linked inputs.
agent-browser is a strong choice when OpenCode must verify a local interface, inspect a rendered page, or reproduce a browser-only defect. Its Skill defines interaction commands, snapshots, sessions, authentication handling, and trust boundaries rather than leaving the agent to improvise clicks.
Choose it when: A change cannot be validated through unit tests or static analysis alone.
Why it matters: Browser automation becomes safer when navigation, credential handling, and page content are treated as explicit boundaries.
Trade-off: Browser access expands the operational surface. Project instructions should limit target domains and require approval before messages, uploads, purchases, permission changes, or other external side effects.
This Skill complements testing; it does not replace it. Use unit and integration tests for deterministic logic, then browser checks for the narrow rendered behavior that matters.
5. Security Review: Best as a Repository-Specific Skill
A security review Skill works best when it encodes the application's actual trust boundaries, sensitive assets, supported authentication flows, and prohibited changes. A generic “find vulnerabilities” prompt is too broad to establish coverage.
Choose it when: Pull requests touch authentication, authorization, secrets, file handling, network boundaries, deserialization, or dependency execution.
What the Skill should inspect: Changed entry points, attacker-controlled inputs, authorization decisions, logs, credential exposure, unsafe defaults, and tests for failure paths.
Trade-off: A Skill is not a penetration test or a guarantee. It can improve review consistency while still missing architectural issues or framework-specific exploits.
Treat every finding as a hypothesis with file evidence and a realistic attack path. The Skill should distinguish a proven vulnerability, a defense-in-depth improvement, and an unresolved question.
6. Database Migration Review: Best for High-Blast-Radius Changes
A database migration review Skill is useful because schema changes fail differently from ordinary application code. The Skill should check forward and rollback paths, long locks, compatibility between old and new application versions, backfill behavior, indexes, nullability, and deployment order.
Choose it when: A change modifies production schema, moves data, or changes an identifier used across services.
Why it matters: A syntactically valid migration can still block writes, exhaust resources, or break a rolling deployment.
Trade-off: The Skill must be database- and deployment-specific. PostgreSQL, MySQL, serverless databases, and managed migration systems expose different operational risks.
Require the output to separate pre-deploy, deploy, backfill, validation, and cleanup phases. A single “run migration” step is rarely enough for a material production change.
7. Release Documentation: Best for Consistent Handoffs
A release documentation Skill turns merged changes and verification evidence into a consistent changelog, upgrade note, or release checklist. It should identify user-visible changes, breaking behavior, migrations, configuration additions, and known limitations.
Choose it when: Release notes are often incomplete or assembled manually from memory.
Why it matters: The code diff shows what changed, but the Skill can enforce the questions needed to explain operational impact.
Trade-off: The agent cannot infer product importance reliably from a diff alone. The workflow needs issue context, labels, or a human decision for ambiguous changes.
For OpenCode itself, keep installation and compatibility statements tied to the current documentation rather than copying old setup posts. The distinction between coding-agent products is also covered in OpenCode versus Claude Code.
8. Web Research and Crawl: Best for Current External Context
A web research and Crawl Skill is useful when a coding task depends on current vendor documentation, release notes, standards, public issues, or comparison evidence. The Skill should define which sources are acceptable, how many pages may be retrieved, what metadata must be preserved, and when conflicting evidence blocks a conclusion.
Nstproxy Crawl can act as the retrieval component for that Skill. It accepts selected public URLs or bounded site jobs and returns structured page artifacts for downstream analysis. JavaScript rendering helps when documentation or product pages are not present in the initial HTML. Page and depth limits keep collection tied to the research question. Markdown, HTML, JSON, Links, and PDF outputs support different evidence workflows. The Skill still owns source quality, deduplication, claim checking, and citation; Crawl does not make evidence true.
Bounded site discovery
Use maximum-page, depth, and include/exclude controls so “research this library” does not become an unbounded crawl.
Rendered page retrieval
Use browser rendering only for pages that need it, and record the final URL, retrieval time, page title, and status alongside the extracted content.
Evidence-ready outputs
Store source-linked Markdown or JSON as evidence inputs, then require the agent to cite the page supporting each change-sensitive claim.
Install a Skill only after reading its contents and testing its trigger on a disposable branch. OpenCode recognizes project-local paths such as .opencode/skills/<name>/SKILL.md and compatible Agent Skills locations documented on the current product page.
A safe rollout has four steps:
Pin the source. Record the repository and revision or vendor release used.
Inspect supporting files. Scripts and templates can have a larger effect than the frontmatter description suggests.
Start with ask-level permissions. Require approval for experimental or external-access Skills until their behavior is understood.
Run positive and negative trigger tests. Confirm the Skill activates for its intended task and stays inactive for neighboring requests.
The portable Agent Skills specification explains the common SKILL.md structure, but OpenCode's current documentation remains authoritative for discovery paths and permission behavior.
How Many OpenCode Skills Should You Install?
Install the smallest set that covers recurring failures in the repository. Five well-owned Skills are usually more useful than a directory of overlapping packs because descriptions compete for activation and instructions can conflict.
Review the catalog whenever the framework, deployment process, or team policy changes. Remove a Skill when no one can explain its trigger, evidence requirements, or owner. For adjacent extension choices, compare Skills with MCP servers for developers: Skills define repeatable procedures, while external tools provide capabilities and live data.
Conclusion
The best OpenCode Skills make agent work more testable and easier to review. Start with Obra Superpowers for process discipline, add one stack-specific Skill such as Vercel React Best Practices, and build repository-owned review Skills only for recurring high-impact decisions.
Next, install two candidates on a disposable branch and score them on trigger accuracy, permission scope, evidence, and review time. If current external documentation is a regular dependency, add a bounded web research Skill and use Nstproxy Crawl as the public-page retrieval layer; test it on a small allowlist before expanding coverage. For teams operating additional proxy-backed collectors, Nstproxy Proxy Manager is the related routing and monitoring capability to evaluate.
OpenCode Skills are reusable Agent Skills whose SKILL.md files describe when a workflow applies and how the agent should execute it. OpenCode discovers available Skills and lets the agent load a matching one through the skill tool.
Q: Where do OpenCode Skills go?
OpenCode supports project and user Skill directories documented on its current Agent Skills page, including .opencode/skills/<name>/SKILL.md for repository-local workflows. Use the current official paths because older community posts may describe different layouts.
Q: What is the best OpenCode Skill for beginners?
Obra Superpowers is the best general starting point when a beginner wants structured planning, debugging, testing, and verification. A smaller repository may prefer to enable only the workflows it will actually enforce.
Q: Are Claude or Codex Skills compatible with OpenCode?
Some Skills follow the portable Agent Skills format and can be discovered from compatible locations, but compatibility should be tested rather than assumed. Tool names, permissions, supporting scripts, and product-specific instructions can still require adaptation.
Q: Can an OpenCode Skill research the web?
An OpenCode Skill can research the web when the runtime exposes an appropriate search or crawl tool. The Skill should set source and evidence rules, while a bounded service such as Nstproxy Crawl retrieves authorized public pages.
Q: Are community OpenCode Skills safe to install?
Community OpenCode Skills should be treated as code and operational instructions, not harmless prompt text. Inspect every file, pin the source, restrict permissions, and test behavior on a disposable branch before repository-wide use.
Ivy Lin
Aug. 25th 2026
110M+ real IPs with 99.9% access success
Blazing-fast average response ~0.5s for high-concurrency tasks
From only $0.1/GB
Get immediate access to premium residential, datacenter, IPv6 and ISP proxy pools.
