2389-research

Use when user says "simmer this", "refine this", "hone this", "iterate on this", or asks to improve a specific artifact over multiple rounds. Runs an iterative refinement loop with investigation-first judges that read the code, understand the problem, and propose evidence-based improvements. Auto-selects single judge or multi-judge board based on complexity. Works on any artifact type: documents, prompts, specs, emails, creative writing, API designs, pipelines, codebases. Supports multi-file workspace targets, runnable evaluators, and open-ended optimization (model selection, pipeline topology, prompt tuning).

Judge board subskill for simmer. Dispatches a panel of judges with different lenses, runs one deliberation round where they challenge each other's scores, then synthesizes consensus scores + single ASI. Drop-in replacement for simmer-judge that produces identical output format. Do not invoke directly — dispatched by the simmer orchestrator when JUDGE_MODE is board.

Reflect subskill for simmer. Records iteration results in trajectory table, tracks best candidate, handles regression rollback, and passes ASI forward to the next round. Supports both single-file and workspace modes. Do not invoke directly — called by simmer orchestrator after each judge round.

Setup subskill for simmer. Inspects the artifact or workspace, infers evaluation contracts and search space, proposes a complete assessment to the user, and produces a setup brief after confirmation. Conversational, not form-based — the agent does the work of understanding the problem, then presents what it found. Do not invoke directly — called by simmer orchestrator.

Generator subskill for simmer. Produces an improved version of the artifact based on the judge's ASI feedback. Handles both single-file and workspace targets. Do not invoke directly — dispatched as a subagent by the simmer orchestrator.

Judge subskill for simmer. Scores a candidate artifact against user-defined criteria on a 1-10 scale and produces ASI (highest-leverage direction) for the next generator round. Supports judge-only, runnable evaluator, and hybrid evaluation modes. Do not invoke directly — dispatched as a subagent by the simmer orchestrator.

This skill should be used when analyzing binaries, executables, or bytecode to understand what they do or how they work. Triggers on "binary", "executable", "ELF", "what does this do", "reverse engineer", "disassemble", "decompile", "pyc file", "python bytecode", "analyze binary", "figure out", "marshal". Routes to sub-skills for triage, static analysis, dynamic analysis, synthesis, or tool setup.

Use when you need to run a binary, trace execution, or observe runtime behavior. Runtime analysis via QEMU emulation, GDB debugging, and Frida hooking - syscall tracing (strace), breakpoints, memory inspection, function interception. Keywords - "run binary", "execute", "debug", "trace syscalls", "set breakpoint", "qemu", "gdb", "frida", "strace", "watch memory"

Use when analyzing binary structure, disassembling code, or decompiling functions. Deep static analysis via radare2 (r2) and Ghidra headless - function enumeration, cross-references (xrefs), decompilation, control flow graphs. Keywords - "disassemble", "decompile", "what does this function do", "find functions", "analyze code", "r2", "ghidra", "pdg", "afl"

Use when ready to document findings, generate a report, or summarize binary analysis results. Compiles analysis findings into structured reports - correlates facts from triage/static/dynamic phases, validates hypotheses, generates documentation with evidence chains. Keywords - "summarize findings", "generate report", "document analysis", "what did we find", "write up results", "export findings"

Use when reverse engineering tools are missing, not working, or need configuration. Installation guides for radare2 (r2), Ghidra, GDB, QEMU, Frida, binutils, and cross-compilation toolchains. Keywords - "install radare2", "setup ghidra", "r2 not found", "qemu missing", "tool not installed", "configure gdb", "cross-compiler"

Use when first encountering an unknown binary, ELF file, executable, or firmware blob. Fast fingerprinting via rabin2 - architecture detection (ARM, x86, MIPS), ABI identification, dependency mapping, string extraction. Keywords - "what is this binary", "identify architecture", "check file type", "rabin2", "file analysis", "quick scan"

Executor subskill for thrifty. ONE cached agent that builds an ordered list of units from their briefs in a single session — honoring the shared contract, running the gate, and self-fixing — then reports per-unit results. (Also handles the degenerate single-unit case.) Do not invoke directly — dispatched as a Haiku subagent by the thrifty orchestrator.

Use when the user says "thrifty", "delegate this", "tiered build", or asks to execute a spec cheaply. thrifty takes a spec you already have (written by hand or by any model — thrifty doesn't care which) and EXECUTES it for less. Orchestrates a tiered-delegation run: the architect (Sonnet) refines the spec into a contract that pins every cross-cutting decision, ONE cached Haiku executor subagent builds the whole ordered unit list in a single session (gate in-loop, self-fixing — not one subagent per unit), and Sonnet verifies/fixes only where a test can't judge. Works for code and non-code tasks alike. PREFER the lean dispatch flow (thrifty-dispatch) by default — it is the benchmarked, cheapest path; THIS subagent substrate is the richer, pricier fallback.

Lean tiered-build orchestrator (JSONL-dispatch architecture). Use for "thrifty fast / lean / dispatch", or a multi-sprint build where cost matters. The architect (Sonnet) writes a contract + a sprints.jsonl and calls a dispatch SCRIPT that loops cheap bare model calls (claude -p --bare, subscription, no API key) to write each sprint to disk — NO subagents. The orchestrator never ingests sprint outputs, only a manifest, so its context stays tiny. Then it runs the gate and surgical-fixes.

Brief-writer subskill for thrifty's split planning mode. Expands ONE terse unit spec (from the architect/director) plus the shared contract into a full, self-contained BRIEF with right-sized approach and concrete acceptance criteria. Has authority WITHIN its unit only; never re-decides anything the contract already pins. Do not invoke directly — dispatched as a Sonnet subagent by the thrifty orchestrator.

Checker subskill for thrifty. Verifies one completed unit against its acceptance criteria, attempts a surgical fix for small localized defects, and returns a structured verdict that diagnoses the failure so the orchestrator can route it. Do not invoke directly — dispatched as a Sonnet subagent by the thrifty orchestrator.

Planning discipline for thrifty, run by the architect (Sonnet) itself. Decomposes a task into units, writes CONTRACT.md (cross-unit decisions + dependency graph), one self-contained BRIEF per unit with concrete acceptance criteria, and initializes the LEDGER. Invoked by the thrifty orchestrator during the Plan step.

This skill should be used when adding features to existing Firebase projects. Triggers on "add function", "create endpoint", "new tool", "add api", "new collection", "implement", "build feature". Guides TDD workflow with test-first development, security rules, and emulator verification.

This skill should be used when troubleshooting Firebase emulator issues, rules violations, function errors, auth problems, or deployment failures. Triggers on "error", "not working", "debug", "troubleshoot", "failing", "broken", "permission denied", "emulator issue".

This skill should be used when initializing a new Firebase project with proven architecture. Triggers on "new firebase project", "initialize firebase", "firebase init", "set up firebase", "create firebase app", "start firebase project". Guides through CLI setup, architecture choices, and emulator configuration.

This skill should be used when working with Firebase projects, including initializing projects, adding Cloud Functions or Firestore collections, debugging emulator issues, or reviewing Firebase code. Triggers on "firebase", "firestore", "cloud functions", "emulator", "firebase auth", "deploy to firebase", "firestore rules".

This skill should be used when reviewing Firebase code against security model and best practices. Triggers on "review firebase", "check firebase", "validate", "audit firebase", "security review", "look at firebase code". Validates configuration, rules, architecture, and security.

Scoring framework for speed-run showdown and any-percent. Invoked at Phase 4 to evaluate implementations using 5-criteria scoring. Do not invoke directly - called by showdown/any-percent.

Same design, multiple parallel runners compete using hosted LLM for code generation. Each runner creates own plan, generates code via Cerebras, pick the best. Part of speed-run pipeline.

Token-efficient code generation pipeline using hosted LLM. Triggers on "speed-run", "fast build", "turbo build", "use hosted LLM", "use cerebras". Routes to turbo (direct codegen), showdown (parallel competition), or any% (parallel exploration).

Explore different architectural approaches in parallel using hosted LLM for code generation. No restrictions on approach - fastest path to comparing real implementations. Part of speed-run pipeline.

Direct code generation via hosted LLM (Cerebras). Write a contract prompt, generate code, fix surgically. Part of speed-run pipeline.

Use when a project needs multi-perspective review — pre-launch audit, post-refactor check, inherited codebase assessment, or periodic health check. Dispatches parallel expert reviewer agents with persona framing.

Use when you want to see a site through fresh eyes — how different types of people experience it for the first time. Dispatches agents with varying tech sophistication who land on the site cold and report confusion, friction, and first impressions.

Use when you want to verify a site works by having agents act as real users completing common tasks — browsing, subscribing, purchasing, searching. Each agent has a goal and clicks through the real flow using browser MCP tools.

Use when you want pedantic, nitpicky, opinionated feedback on a site or project — the kind of feedback you'd get from Hacker News commenters, typography snobs, grammar pedants, and standards purists. Finds the things that professional reviewers skip because they're "too minor."

This skill should be used when implementing features with parallel exploration or competition. Triggers on "build", "create", "implement", "try both approaches", "compare implementations". Routes to omakase-off (entry gate for design exploration) or cookoff (exit gate for parallel implementation).

This skill should be used when moving from design to implementation. Triggers on "let's build", "implement this", "looks good let's code", "ready to implement". Presents options for parallel agent competition (cookoff), single subagent, or local implementation. Each agent creates own plan from shared design for genuine variation.

This skill should be used as the entry gate for build/create/implement requests. Triggers on "build X", "create Y", "implement Z", "add feature", "try both approaches", "not sure which approach". Offers brainstorm-together or omakase (chef's choice parallel exploration) options. Detects indecision during brainstorming to offer parallel exploration.

Scoring framework for test-kitchen cookoff and omakase-off. Invoked at Phase 4 to evaluate implementations using 5-criteria scoring. Do not invoke directly - called by cookoff/omakase-off.

Use when a question needs parallel deep analysis from multiple perspectives - code reviews, architecture decisions, research synthesis - spawn specialized agents as a clearness committee rather than doing everything in one context

Use when facing questions with ethical weight, multiple valid approaches, significant trade-offs, or potential for harm - before answering, convene internal voices to discern rather than conclude

Use when user has a stake or perspective in a decision and wants to participate in discernment rather than receive advice - facilitates user alongside agent voices with participatory discipline teaching

Use when facing decisions requiring careful consideration - ethical questions, architecture choices, trade-offs, or when user has a stake and wants to participate in finding clarity. Routes to discernment (internal), clearness (multi-agent), or gathered (participatory).

This skill should be used when creating new styled components or adding new CSS classes. Triggers on "create component", "new button", "new card", "add styles", "style component", "build UI element". Guides semantic naming, Tailwind composition, dark mode support, and test coverage.

This skill should be used when refactoring existing CSS from inline styles or utility classes to semantic patterns. Triggers on "refactor CSS", "extract styles", "consolidate CSS", "convert inline", "clean up styles", "migrate to semantic". Transforms to semantic classes with dark mode and tests.

This skill should be used when working with CSS, creating components, styling elements, refactoring styles, or reviewing CSS code. Triggers on "CSS", "styles", "Tailwind", "dark mode", "component styling", "semantic class", "@apply", "stylesheet". Routes to specialized sub-skills for creation, validation, or refactoring.

This skill should be used when reviewing or auditing existing CSS code for consistency with established patterns. Triggers on "review CSS", "audit styles", "check CSS", "validate stylesheet", "CSS review". Checks semantic naming, dark mode coverage, Tailwind usage, and test coverage.

Use when preparing a codebase for first-time public/open-source release. Full lifecycle from audit through documentation, hardening, and final review.

Use when auditing a codebase for openness, checking for secrets, missing docs, or gaps before public release. Re-runnable at any point.

Use when preparing a codebase for public/open-source release, reviewing a repo for openness, or auditing for secrets and missing documentation before making code public. Triggers on "open source", "public release", "prepare repo", "openness", "repo review", "make public", "open-source prep".

CI/workflow monitoring and fixing. Triggers on "CI failing", "workflow failed", "checks red", "build broken", "tests failing", "rerun workflows".

PR health assistant - monitors CI status, triages review comments, fixes issues with systematic prevention. Triggers on "check my PR", "PR health", "make PR green", "fix PR", "ready to merge", "handle PR feedback".

Review comment triage and handling. Triggers on "review comments", "address feedback", "reviewer asked", "changes requested", "handle comments", "triage reviews".

Use when working with contacts, companies, deals, or relationships in the Pagen CRM. This skill ensures proper contact management, duplicate prevention, and interaction logging.

Use when handling email tasks (checking inbox, drafting replies, managing threads, adding events to calendar). Prevents common mistakes like broken threading and ensures proper MCP tool usage.