All Skills > SDK Skill Creator
Create a Sentry SDK Skill Bundle
Produce a complete, research-backed SDK skill bundle — a main wizard SKILL.md plus deep-dive reference files for every feature pillar the SDK supports.
Invoke This Skill When
- Asked to "create a Sentry SDK skill" for a new platform
- Asked to "add support for [language/framework]" to sentry-agent-skills
- Building a new
sentry-<platform>-sdkskill bundle - Porting the SDK skill pattern to a new Sentry SDK
Read
${SKILL_ROOT}/references/philosophy.mdfirst — it defines the bundle architecture, wizard flow, and design principles this skill implements.
Phase 1: Identify the SDK
Determine what you're building a skill for:
# What SDK? What's the package name?
# Examples: sentry-go, @sentry/sveltekit, sentry-python, sentry-ruby, sentry-cocoa
Establish the feature matrix — which Sentry pillars does this SDK support?
| Pillar | Check docs | Notes |
|---|---|---|
| Error Monitoring | Always available | Non-negotiable baseline |
| Tracing/Performance | Usually available | Check for span API |
| Profiling | Varies | May be removed or experimental |
| Logging | Newer feature | Check minimum version |
| Metrics | Newer feature | May be beta/experimental |
| Crons | Backend only | Not available for frontend SDKs |
| Session Replay | Frontend only | Not available for backend SDKs |
| AI Monitoring | Some SDKs | Usually JS + Python only |
Reference existing SDK skills to understand the target quality level:
ls skills/sentry-*-sdk/ 2>/dev/null
# Read 1-2 existing SDK skills for pattern reference
Phase 2: Research
This is the most critical phase. Skill quality depends entirely on accurate, current API knowledge. Do NOT write skills from memory — research every feature against official docs.
Research Strategy
Spin off parallel research tasks (using the claude tool with outputFile) — one per feature area. Each task should:
- Visit the official Sentry docs pages for that feature
- Visit the SDK's GitHub repo for source-level API verification
- Write thorough findings to a dedicated research file
Read ${SKILL_ROOT}/references/research-playbook.md for the detailed research execution plan, including prompt templates and file naming conventions.
Research the Sentry Wizard
Before diving into feature research, check whether the Sentry wizard CLI supports this framework:
# Check the SDK's docs landing page for wizard instructions
# Visit: https://docs.sentry.io/platforms/<platform>/
# Look for: "npx @sentry/wizard@latest -i <integration>"
If a wizard integration exists:
- Document the exact wizard command and
-iflag - Document what the wizard creates/modifies (files, config, build plugins)
- Note that the wizard handles authentication interactively — login, org/project selection, and auth token creation/download all happen automatically
- Note whether the wizard sets up source map upload — this is critical for frontend SDKs
- This will become "Option 1: Wizard (Recommended)" in Phase 3 of the generated skill
Why this matters: The wizard handles the entire auth flow (login, org/project selection, auth token) and source map upload configuration automatically. Without source maps, production stack traces show minified code — making Sentry nearly useless for frontend debugging. And without the auth token, source maps can't be uploaded at all. The wizard is the most reliable way to get both right in a single step.
Research Batching
Batch research tasks by topic area. Run them in parallel where possible:
| Batch | Topics | Output file |
|---|---|---|
| 1 | Setup, configuration, all init options, framework detection | research/<sdk>-setup-config.md |
| 2 | Error monitoring, panic/exception capture, scopes, enrichment | research/<sdk>-error-monitoring.md |
| 3 | Tracing, profiling (if supported) | research/<sdk>-tracing-profiling.md |
| 4 | Logging, metrics, crons (if supported) | research/<sdk>-logging-metrics-crons.md |
| 5 | Session replay (frontend only), AI monitoring (if supported) | research/<sdk>-replay-ai.md |
Important: Tell each research task to write its output to a file (outputFile parameter). Do NOT consume research results inline — they're large (500–1200 lines each). Workers will read them from disk later.
Research Quality Gate
Before proceeding, verify each research file:
- Has actual content (not just Claude's process notes)
- Contains code examples with real API names
- Includes minimum SDK versions
- Covers framework-specific variations
# Quick verification
for f in research/<sdk>-*.md; do
echo "=== $(basename $f) ==="
wc -l "$f"
grep -c "^#" "$f" # should have multiple headings
done
Re-run any research task that produced fewer than 100 lines — it likely failed silently.
Phase 3: Create the Main SKILL.md
The main SKILL.md implements the four-phase wizard from the philosophy doc. Keep it focused — the main file should cover the wizard flow, quick start config, framework tables, and reference dispatch. Deep-dive details for individual features belong in references/ files, not here. Be thorough but not redundant.
Gather Context First
Before writing, run a scout or read existing skills to understand conventions:
- Frontmatter pattern (name, description, license)
- "Invoke This Skill When" trigger phrases
- Table formatting and code example style
- Troubleshooting table conventions
SKILL.md Structure
---
name: sentry-<platform>-sdk
description: Full Sentry SDK setup for <Platform>. Use when asked to "add Sentry
to <platform>", "install <package>", or configure error monitoring, tracing,
[features] for <Platform> applications. Supports [frameworks].
license: Apache-2.0
---
# Sentry <Platform> SDK
## Invoke This Skill When
[trigger phrases]
## Phase 1: Detect
[bash commands to scan project — package manager, framework, existing Sentry, frontend/backend]
## Phase 2: Recommend
[opinionated feature matrix with "always / when detected / optional" logic]
## Phase 3: Guide
### Option 1: Wizard (Recommended) ← if wizard exists for this framework
[blockquote telling the user to run the wizard themselves — it requires interactive browser login. Include the command in a copy-pasteable code block inside the blockquote. Tell them to come back when done. Add a line after the blockquote: "If the user skips the wizard, proceed with Option 2 (Manual Setup) below."]
### Option 2: Manual Setup ← always include
### Install
### Quick Start — Recommended Init
### Source Maps Setup ← required for frontend/mobile SDKs
### Framework Middleware (if applicable)
### For Each Agreed Feature
[reference dispatch table: feature → ${SKILL_ROOT}/references/<feature>.md]
## Configuration Reference
[key init options table, environment variables]
## Verification
[test snippet]
## Phase 4: Cross-Link
[detect companion frontend/backend, suggest matching SDK skills]
## Troubleshooting
[common issues table]
Key Principles for the Main SKILL.md
- Keep it lean — deep details go in references, not here
- Wizard-first for framework SDKs — if the Sentry wizard supports this framework, present it as "Option 1: Wizard (Recommended)" before any manual setup. The wizard requires interactive browser login and cannot be run by the agent — present it in a blockquote telling the user to copy-paste the command into their own terminal, and come back when done. If the user skips the wizard, the agent proceeds with full manual setup. See
${SKILL_ROOT}/references/philosophy.mdfor the full pattern. - Source maps are non-negotiable for frontend/mobile — the manual setup path must include source map upload configuration (build tool plugin + env vars). Without source maps, production stack traces are unreadable minified code.
- Detection commands must be real — test them against actual projects
- Recommendation logic must be opinionated — "always", "when X detected", not "maybe consider"
- Quick Start config should enable the most features with sensible defaults
- Framework middleware table — exact import paths, middleware calls, and quirks
- Cross-link aggressively — if Go backend, suggest frontend. If Svelte frontend, suggest backend.
Phase 4: Create Reference Files
One reference file per feature pillar the SDK supports. These are deep dives — they can be longer than the main SKILL.md.
Reference File Structure
# <Feature> — Sentry <Platform> SDK
> Minimum SDK: `<package>` vX.Y.Z+
## Configuration
## Code Examples
### Basic usage
### Advanced patterns
### Framework-specific notes (if applicable)
## Best Practices
## Troubleshooting
| Issue | Solution |
|-------|----------|
What Makes a Good Reference
Read ${SKILL_ROOT}/references/quality-checklist.md for the full quality rubric.
Key points:
- Working code examples — not pseudo-code, not truncated snippets
- Tables for config options — type, default, minimum version
- One complete example per pattern — don't show 5 variations of the same thing
- Framework-specific notes — call out when behavior differs between frameworks
- Minimum SDK version at the top — always
- Honest about limitations — if a feature was removed (like Go profiling), say so
Feature-Specific Guidance
| Feature | Key things to cover |
|---|---|
| Error Monitoring | Capture APIs, panic/exception recovery, scopes, enrichment (tags/user/breadcrumbs), error chains, BeforeSend, fingerprinting |
| Tracing | Sample rates, custom spans, distributed tracing, framework middleware, operation types |
| Profiling | Sample rate config, how it attaches to traces, or honest "removed/not available" |
| Logging | Enable flag, logger API, integration with popular logging libraries, filtering |
| Metrics | Counter/gauge/distribution APIs, units, attributes, best practices for cardinality |
| Crons | Check-in API, monitor config, schedule types, heartbeat patterns |
| Session Replay | Replay integration, sample rates, privacy masking, canvas/network recording |
Note for frontend/mobile SDKs: Source map upload configuration belongs in the main SKILL.md (Phase 3: Guide), not in a reference file. It's part of the core setup flow — every frontend production deployment needs it. Cover the build tool plugin, the required env vars (
SENTRY_AUTH_TOKEN,SENTRY_ORG,SENTRY_PROJECT), and add.envto.gitignore.
Phase 5: Verify Everything
Do NOT skip this phase. SDK APIs change frequently. Research can hallucinate. Workers can fabricate config keys.
API Verification
Run a dedicated verification pass against the SDK's actual source code:
Research prompt: "Verify these specific API names and signatures against
the <SDK> GitHub repo source code: [list every API from the skill files]"
Things that commonly go wrong:
- Config option names with wrong casing (
SendDefaultPiivsSendDefaultPII) - Fabricated config keys that don't exist (
experimental.tracing— verify it's real) - Deprecated APIs used instead of modern replacements (
configureScope→getIsolationScope) - Features listed as available when they've been removed (profiling in Go SDK)
- Wrong minimum version numbers
Review Pass
Run a reviewer on the complete skill bundle:
- Technical accuracy of code examples
- Consistency between main SKILL.md and reference files
- Consistency with existing SDK skills in the repo
- Agent Skills spec compliance (frontmatter, naming)
Fix Review Findings
Triage by priority:
- P0: Misleading claims (advertising removed features) — fix immediately
- P1: Incorrect APIs, deprecated methods — fix before merge
- P2: Style inconsistencies, version nitpicks — fix if quick
- P3: Skip
Phase 6: Register and Update Docs
After the skill passes review:
- Update README.md — add to the SDK Skills table
- Update AGENTS.md — if the philosophy doc or skill categories section needs it
- Add usage examples — trigger phrases in the Usage section
- Document the bundle pattern — if this is a new SDK, note the references/ structure
Commit Strategy
Each major piece gets its own commit:
feat(<platform>-sdk): add sentry-<platform>-sdk main SKILL.md wizardfeat(<platform>-sdk): add reference deep-dives for all feature pillarsdocs(readme): add sentry-<platform>-sdk to available skillsfix(skills): address review findings(if any)
Checklist
Before declaring the skill complete:
- Philosophy doc read and followed
- All feature pillars researched from official docs (not from memory)
- Research files verified (real content, correct APIs, >100 lines each)
- Main SKILL.md is focused — wizard flow + quick start + reference dispatch; deep dives in references
- Main SKILL.md implements all 4 wizard phases
- Wizard CLI checked — if supported, presented as "Option 1: Wizard (Recommended)" with auth flow + source map benefits described
- Source map / debug symbol upload covered in manual setup path (frontend/mobile SDKs)
- Reference file for each supported feature pillar
- APIs verified against SDK source code
- Review pass completed, findings addressed
- Profiling/removed features honestly documented (not advertised)
- Cross-links to companion frontend/backend skills
- README.md updated
- All commits polished with descriptive messages