Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/sontek/sontek-skills itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
Skills do not use an MCP config. Follow the steps in the README below.
README
synced Oct 23, 2026Sontek Skills
Agent skills for Codex and Claude Code, following the Agent Skills open format. The repository also ships Claude Code-specific subagents.
Installation
Codex from local clone
# Clone the repository
git clone git@github.com:sontek/sontek-skills.git ~/sontek-skills
# Register this repository's marketplace and install the plugin
codex plugin marketplace add ~/sontek-skills
codex plugin add sontek-skills@sontek-skills-local
Start a new Codex thread after installation so the skills are discovered.
Codex loads the skills under plugins/sontek-skills/skills/. The persona files
under plugins/sontek-skills/agents/ are Claude Code-specific; skills that ask
for parallel or delegated reviewers can still use Codex's native sub-agent
support when it is available.
Claude Code from local clone
# Clone the repository
git clone git@github.com:sontek/sontek-skills.git ~/sontek-skills
Then inside Claude Code, run these slash commands:
/plugin marketplace add ~/sontek-skills
/plugin install sontek-skills@sontek-skills-local
After installation, restart Claude Code. Skills and agents are automatically invoked when relevant to your task.
Quick Test (No Installation)
# Run Claude Code with the plugin directory directly
claude --plugin-dir ~/sontek-skills/plugins/sontek-skills
Updating
Inside Claude Code, run:
/plugin update sontek-skills@sontek-skills-local
Or use /plugin to open the interactive plugin manager.
What's in this plugin
This plugin ships two kinds of capabilities:
- Skills — portable atomic workflows loaded by Codex or Claude Code.
verb-objectnaming (create-pr,review-code,write-skill). - Agents — Claude Code-specific persona subagents that run in isolated
context and compose skills with judgment.
role-nounnaming (senior-engineer,security-auditor).
See Skills vs agents for the decision guide.
Available Skills
Git workflow
| Skill | Description |
|---|---|
| create-branch | Create git branches following <type>/<description> conventions |
| commit | Conventional commits with strict AI attribution rules (only Co-Authored-By) |
| create-pr | Create PRs; respects repo PULL_REQUEST_TEMPLATE.md when present |
| iterate-pr | Iterate on a PR until CI passes and review feedback is addressed |
| update-changelog | Add notable user-facing changes since the last tag to CHANGELOG.md's ## Unreleased section |
Code review & audit
| Skill | Description |
|---|---|
| review-code | Prioritized code review (P0-P3) with fail-fast error-handling rubric; branch mode (diff vs. main) or paths mode (explicit file list). Fans out to specialist reviewers, then a fresh-context finding-verifier pass refutes weak findings |
| review-pr | Multi-reviewer PR review consolidated into human-toned comments for your approval (never posts unprompted). Two modes: incoming (someone else's PR) and self (self-review your own PR for completeness and merge-readiness) |
| auto-review-code | Loop review-code and simplify-code until no safe fixes remain; batches risky changes for a single end-of-run approval pass |
| simplify-code | Route changed code through the code-simplifier agent's AI-slop rubric and apply the fixes (the built-in /simplify lacks the rubric) |
| review-security | OWASP-aligned security review with confidence-based reporting |
| review-django-access | Django/DRF IDOR and access-control review |
| review-django-perf | Django performance review (N+1, unbounded queries, missing indexes) |
| review-perf | Application-tier performance review for non-Django stacks (Flask, FastAPI, Go, Node) — blocking I/O on async, algo blow-ups, per-item network loops, missing app cache |
| optimize-perf | Apply perf fixes from review-perf with before/after benchmarks; reverts if the measured delta doesn't meet a threshold |
| review-gha-security | GitHub Actions security review (pwn requests, expression injection, credential theft) |
Implementation
| Skill | Description |
|---|---|
| fix-issue | Implement a fix for a bug/issue from scratch, reproduce-first (failing check → root-cause fix → verify → edge cases → hand off to commit/create-pr) |
| debug | Diagnose a bug by the scientific method (stabilize → hypothesize → experiment → fix root cause → regression test); pattern-reuse gate, anti-patterns, time limits. Composes with sentry and iterate-pr |
Planning & design
| Skill | Description |
|---|---|
| clarify | Interview the user one question at a time to resolve ambiguity or stress-test a plan — classifies the gap (intention/premise/parameter/expression) and asks the highest-information question, with a confirmatory mode for "just do it". Formerly grill-me |
| plan-implementation | Structured plan for NEW features (writes IMPLEMENTATION_PLAN_*.md) |
| plan-refactor | Tiny-commit refactor plan (writes REFACTOR_PLAN_*.md) |
| review-plan | Route an existing plan (file or plan-mode draft) through senior-engineer and product-manager for a judgment pass |
| auto-review-plan | Loop review-plan and auto-apply small blocking edits until the plan is clean; flags structural recommendations for end-of-run approval |
| improve-architecture | Find architectural friction and design module-deepening refactors |
| document-architecture | Generate a human-readable architecture document for a module/service using 4 parallel Explore sub-agents; embeds validated Mermaid diagrams via draw-mermaid-diagram / draw-infra-diagram |
| frontend-design | Design distinctive frontend UIs with strong aesthetic direction (greenfield/redesign work) |
Diagrams
| Skill | Description |
|---|---|
| draw-mermaid-diagram | Create and validate Mermaid diagrams of any type (flowchart, sequence, class, ER, state, gantt). Houses tools/validate.sh |
| draw-infra-diagram | AWS / cloud architecture specialist on top of draw-mermaid-diagram — adds shape vocabulary, color palette, region tinting, debugging-oriented composition |
Session management
| Skill | Description |
|---|---|
| handoff | Write a structured HANDOFF_<slug>.md so a fresh session can continue the current work without re-reading the transcript. Use before /clear or when starting a new window mid-task |
Tools & integrations
| Skill | Description |
|---|---|
| librarian | Cache remote git repos at ~/.cache/checkouts/<host>/<org>/<repo> for stable, throttled-refresh local access |
| sentry | Fetch Sentry issues / events / transactions / logs via the Sentry API for debugging |
| uv | Use uv for Python in repos that have adopted it — uv run, uv add, inline script metadata, uv_build |
Sales & GTM
| Skill | Description |
|---|---|
| sales-email-draft | Draft or polish a sales/GTM email in John Anderson's voice as Founder of Drape, applying Drape's brand voice and a Win Without Pitching posture |
| sales-email-prospect | Draft a prospecting/outreach email from the context you provide, aimed at earning a Probative Conversation |
| sales-email-followup | Draft a post-meeting follow-up grounded in the actual meeting notes pulled from Google Docs (Google Drive) |
| sales-email-reply | Draft a reply to a pasted email thread; applies the Win Without Pitching posture for sales-adjacent threads |
Meta
| Skill | Description |
|---|---|
| write-skill | Create new skills and agents with proper structure and triggers |
| review-skill | Audit existing skills against the write-skill rubric (frontmatter, trigger coverage, structure, references) |
| write-agents-md | Write concise, high-signal AGENTS.md / CLAUDE.md files |
| review-tone | Make any prose (email, PR description, commit, doc) read human, not robot: strips em-dashes, cuts AI-tell filler, signals weight in words. Shared final pass for the sales-email-* family |
Available Agents
Agents are persona subagents that run autonomously in isolated context. They use the skills above but bring experience and opinion.
| Agent | Description |
|---|---|
| code-simplifier | Autonomous refactoring specialist — refines recently modified code for clarity (opus) |
| senior-engineer | 15+ year SaaS engineer — opinionated architecture, scalability, maintainability (opus) |
| security-auditor | Offensive + defensive security engineer — HIGH-confidence findings with PoC (opus) |
| product-manager | 15+ year PM with UX sensibility — user journeys, scope discipline, prioritization (opus) |
| researcher | Web research specialist — source-cited findings with explicit confidence and conflicting-source flags (opus) |
See agents/README.md for invocation triggers and when to use each one.
Skills vs agents
| Skills | Agents | |
|---|---|---|
| What | Atomic capabilities loaded inline | Personas/roles running in isolated context |
| Naming | verb-object (create-pr, review-code) | role-noun (senior-engineer, code-simplifier) |
| Location | plugins/sontek-skills/skills/<name>/SKILL.md | plugins/sontek-skills/agents/<name>.md |
| Frontmatter | allowed-tools: | model: (opus/sonnet/haiku) |
| Context | Runs in main conversation — user sees every step | Runs in isolated context — returns a final report |
| Best for | Conversational workflows | Fire-and-forget audits, persona judgment |
Use the write-skill skill to create either one.
Creating new skills
Skills follow the Agent Skills specification. Each skill requires a SKILL.md file with YAML frontmatter.
Skill template
Create a new directory under plugins/sontek-skills/skills/:
plugins/sontek-skills/skills/my-skill/
└── SKILL.md
SKILL.md format:
---
name: my-skill
description: A clear description of what this skill does. Use when [specific triggers]. Include keywords that help the agent identify when this skill is relevant.
---
# My Skill Name
## Quick start
[Minimal working example]
## Workflow
[Step-by-step guidance]
## Examples
Concrete examples showing expected input/output.
Naming conventions
- name: 1-64 characters, lowercase alphanumeric with hyphens only.
verb-objectfor skills (e.g.create-pr),role-nounfor agents (e.g.code-simplifier). - description: Up to 1024 characters, include trigger keywords and "Use when..." clause.
- Keep
SKILL.mdunder ~100 lines; split longer content intoreferences/*.md.
Wrap prose at 80 chars if desired:
npx prettier --write --prose-wrap always --print-width 80 your-file.md
Optional frontmatter fields
| Field | Description |
|---|---|
license | License name or path to license file |
compatibility | Environment requirements (max 500 chars) |
model | Override model for this skill (e.g., sonnet, opus, haiku) |
allowed-tools | Space-delimited list of tools the skill can use |
metadata | Arbitrary key-value pairs for additional properties |
---
name: my-skill
description: What this skill does
license: Apache-2.0
model: sonnet
allowed-tools: Read Grep Glob
---
References
- Agent Skills Specification
- getsentry/sentry-skills — upstream we fork from
- mattpocock/agent-skills — source of
clarify(formerlygrill-me),improve-architecture,plan-refactor - ryanthedev/code-foundations and systems-design.skill (Code Complete / APOSD / Clean Architecture / Legacy Code) — techniques folded into
clarify,debug,code-reviewer,finding-verifier,improve-architecture,plan-refactor,plan-implementation,review-plan - mitsuhiko/agent-stuff — source of
librarian,update-changelog,sentry,uv,frontend-design, and the Mermaid validation tool used bydraw-mermaid-diagram/draw-infra-diagram
License
MPL-2.0