ckorhonen

Create production-ready GitHub Actions workflows for automated testing, building, and deploying applications. Use when setting up CI/CD with GitHub Actions, automating development workflows, or creating reusable workflow templates.

Implement comprehensive evaluation strategies for LLM applications using automated metrics, LLM-as-judge, human feedback, and benchmarking. Use when testing LLM performance, measuring AI application quality, comparing prompts/models, or establishing evaluation frameworks. Covers RAGAS for RAG pipelines, evals-as-code CI/CD integration, and modern 2025/2026 practices including structured output evaluation and agentic task success measurement.

Create or expand an Idea.md / IDEA.md file from a rough description, existing repo, conversation history, notes, or other early-stage product inputs. Use when the user asks to "write an Idea.md", "turn this into an idea file", "capture this product idea", "expand this concept", or wants a repo-grounded concept brief before validation, PRD, or implementation work.

Run a rigorous autonomous experiment loop for any optimization target using explicit hypotheses, repeated trials, structured experiment logs, and local HTML reports. Use when asked to "run autoresearch", "optimize X in a loop", "start experiments", or "improve this with benchmark-driven iteration".

Capture reusable knowledge and create new skills automatically during coding sessions, problem solving, and retrospectives. Use when finishing a complex task, running a retrospective, discovering a reusable pattern, or when user says "save this as a skill", "capture this solution", "run a retrospective", or "what did we learn from this".

Build professional native macOS apps in Swift with SwiftUI and AppKit. Full lifecycle - build, debug, test, optimize, ship. CLI-only, no Xcode. Use when asked to: create macOS apps, build Swift apps, develop SwiftUI interfaces, fix macOS app issues, add macOS app features, or when user says 'build a Mac app', 'create a Swift project', 'develop for macOS'.

Use when the user asks to generate or edit images via the OpenAI Image API (for example: generate image, edit/inpaint/mask, background removal or replacement, transparent background, product shots, concept art, covers, or batch variants); run the bundled CLI (`scripts/image_gen.py`) and require `OPENAI_API_KEY` for live calls.

Perform language and framework specific security best-practice reviews and suggest improvements. Trigger when the user requests: security review, security report, threat assessment, vulnerability scan, security audit, secure-by-default coding, hardening guidance, or security best practices guidance. Supported languages: Python, JavaScript/TypeScript, Go. Covers OWASP Top 10, supply chain security, secrets management, dependency auditing, and modern cloud deployment hardening.

Babysits an open pull request end-to-end by polling every minute, triaging new comments and reviews, handling CI failures, iterating on reviewer scores, and stopping only when the PR is ready to merge. Use when a user says `babysit this PR`, `watch this PR until it's merge-ready`, `handle review feedback on my PR`, or `keep iterating on this PR until reviewers are happy`. Do NOT use for a one-shot PR risk review, merged PR production monitoring, or repo-wide CI debugging with no scoped PR.

Initializes an optional repo-local agent collaboration preference file at `.ai/swe.json` by running a short interview or a zero-question quick mode. Use when a user says `initialize agent settings for this repo`, `set up my local agent prefs here`, `run quick init for this project`, or `create .ai/swe.json for how I like to work`. Do NOT use for `npm init`, project scaffolding, dependency installation, or environment bootstrap.

Hunts for concrete performance bottlenecks in a scoped repository surface using profiler output, benchmarks, query plans, traces, bundle analysis, or repo evidence, then returns the smallest high-value follow-up experiments or fixes. Use when a user says `find performance bottlenecks`, `why is this slow`, `profile this flow`, `hunt hot paths`, or asks for a recurring performance review. Do NOT use for live incident response, generic observability audits, speculative micro-optimization, or broad architecture rewrites with no bottleneck evidence.

Plans the narrowest trustworthy validation path for a scoped code change or diff. Use when a user asks what to run before merging, how to validate a specific change, whether the current checks are enough, or wants a bounded command order from narrow to broad. Do NOT use for writing tests, fixing the code change itself, broad QA sweeps, or generic debugging that needs root cause analysis.

Creates or revises `swe:` skills for this repository using a repeatable authoring workflow: define concrete use cases, tune trigger boundaries, apply progressive disclosure, add matching eval assets, and validate the package. Use when adding a new skill, tightening an existing skill that over- or under-triggers, or distilling an external workflow into a repo-ready skill. Do NOT use for generic documentation edits or for executing the workflow the skill would describe.

Audits a repository for human-facing or operational documentation that drifted from code, config, interfaces, workflows, or repo structure changes. Use when a user says `check docs drift`, `docs are stale`, `update the runbook after this change`, or `what documentation is missing after this change`. Do NOT use for agent-guidance updates, generic documentation rewrites, or writing docs without evidence that they are stale.

Audits post-incident engineering follow-through after a sev or incident to verify whether the durable follow-up happened: regression tests, monitors, docs, runbooks, ownership updates, tickets, rollback learnings, and remaining backlog. Use when a user asks whether incident follow-up is complete, what still needs to be done after a postmortem, or how to close the engineering loop. Do NOT use for live incident response, root-cause analysis, or a generic bug hunt unrelated to an incident.

Inspects services, jobs, and code paths for missing or weak logs, metrics, traces, alerts, dashboards, or deployment-linked telemetry, then returns a tightly scoped backlog of observability gaps. Use when a user says `find observability gaps`, `audit telemetry coverage`, `what logs or metrics are missing`, `check alerting coverage`, or asks for a recurring telemetry review. Do NOT use for live incident response, root-cause analysis, generic performance tuning, or a broad code review.

Finds manual processes that could be automated: manual deploy steps in READMEs, setup instructions that could be scripts, repetitive git workflows, manual testing procedures, and undocumented tribal knowledge. Also checks for artifacts not in source control. Use monthly to surface automation ROI. Do NOT use for code quality or style audits.

Detects entropy signals in a codebase: stale TODOs, disabled tests, lint suppressions, commented-out code, dead imports, empty catch blocks, and deprecated API usage. Designed for daily runs to catch quality erosion early. Do NOT use for feature work, refactoring planning, or security audits.

Finds duplicated code patterns across the codebase including exact copies and structural duplication (same logic with different variable names). Ranks by frequency and suggests extraction candidates. Designed for weekly runs. Do NOT use for method length, naming, or style concerns.

Finds code that obscures its intent: unclear variable names, nested ternaries, boolean parameters without names, overly clever one-liners, and comments that restate code instead of explaining why. Focuses on recently changed files. Do NOT use for method length, duplication, or test naming concerns.

Surfaces magic numbers, unexplained string literals, hardcoded URLs, configuration values buried in code, and other "magical" values that should be named constants or config entries. Use when you want a weekly sweep for unnamed literals. Do NOT use for style or formatting audits.

Scans a codebase for methods and functions that exceed configurable length thresholds. Reports a ranked list of the longest methods with file:line, language, and surrounding complexity signals. Designed for scheduled weekly runs. Do NOT use for style formatting, indentation, or naming concerns.

Flags classes, modules, and functions with vague generic names like Manager, Handler, Processor, Helper, Utils, Service, Base, Common, and Misc. Suggests more specific alternatives based on what the code actually does. Use weekly to prevent naming drift. Do NOT use for variable-level naming or style enforcement.

Checks that test method and function names read as complete sentences describing behavior. Flags cryptic names like test1, testFoo, or abbreviated names that do not describe what is being tested. Designed to run on every PR. Do NOT use for test coverage, test structure, or non-test code.

Review a work session, transcript, or tool-call log and compound the learnings into durable improvements: agent rules (AGENTS.md/CLAUDE.md), new or updated skills, slash commands, deterministic scripts, sub-agent delegation, and context-bloat reduction. Use when the user says "reflect", at the end of a work session, on a schedule, or when asked how to make the agent faster, cheaper, or more reliable in this project. Do not use for ordinary code review unless the requested output is agent/process improvement.

Scan local agent session logs (Claude Code, Codex) across many sessions, cluster the common failure and friction patterns, then manage the skill portfolio to address them: find and install existing ecosystem skills, create or modify local skills, fix under- or over-triggering descriptions, and remove skills that misfire. Use when the user asks to scan session logs for failure patterns, audit or tune their skills, or asks why a skill never fires. For single-session extraction use reflect instead.

Audit whether past reflect improvements are actually working. Reads the .reflect/applied.md ledger and checks each applied rule, skill, command, or script for evidence it fired and helped; grades entries verified, unused, or regressed, and prunes dead weight. Use on a schedule (weekly or every ~10 sessions), when the user asks "are these improvements working?", or before a reflect-memory consolidation pass. Requires prior reflect runs — if no .reflect/ ledger exists, run reflect first.

Lifecycle maintenance for accumulated agent memory: dedupe, merge, expire, and resolve contradictions across AGENTS.md/CLAUDE.md agent rules, .reflect/ state, and other agent instruction files reflect maintains. Use monthly, when instruction files feel bloated or contradictory, when the agent cites stale rules, or after many reflect runs have accumulated. Not for extracting new learnings — that is the reflect skill.