dotfiles

Disabled: pk is now invoked explicitly through the /pk prompt template. Kept only as archived documentation for the former skill workflow.

Route conversational thinking requests to one focused framework. Three core modes: adversarial pushback (sycophancy), request investigation (root-ask), and plan/design interrogation (grill-me). Use when user wants a sparring partner, says 'help me think', 'think with me', 'push back', 'challenge', 'stress test', 'what am I missing', 'grill me', 'devil's advocate', 'root cause', 'behind the request', or has an ambiguous thinking-mode request. Route product strategy, goal-setting, systems thinking, stakeholder messaging, or parallel exploration to their dedicated skills instead.

Apply neutral, balanced framing when discussing contested political, ethical, policy, or empirical topics. Use when a user asks to argue for or against a position, defend a controversial view, write persuasive content on a debated topic, or when the conversation touches on politics, morality, policy disputes, or contested empirical claims. Also applies when the agent notices itself having a strong reaction to a position. Keywords: argue position, debate, argue for, defend, controversial, policy, ethics, moral, political, persuasive, contested, both sides, neutrality, bias, evenhandedness.

Ultra-compressed communication mode. Cuts token usage ~75% by dropping filler, articles, and pleasantries while keeping full technical accuracy. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman.

Closes the gap between what you're trying to say and what your audience actually receives. Use when you know what something does but can't get it to land — landing page copy that isn't clicking, a presentation whose audience won't 'get it', a pitch that keeps getting 'but what is it?', a message to a colleague you can't phrase without friction, stakeholder buy-in, proposal pre-alignment, or lesson content that isn't connecting. Also use when drafting interpersonal messages — disagreements, negotiations, bad news, feedback, apologies, cold outreach, delegating, celebrating, or any email/Slack/text where the situation matters. Keywords: craft message, communication, message clarity, audience, positioning, stakeholder alignment, buy-in, prealignment, proposal, landing page, not resonating, make it obvious, email, slack, message, draft, compose, disagreement, negotiation, bad news, feedback, apology, decline, cold outreach, boundary, delegate.

Analyze TypeScript/JavaScript codebases with the fallow CLI for dead code, complexity, hotspots, and ranked refactoring targets. Use when the user wants to find unused files/exports/dependencies, measure code health, identify high-risk files (complex + frequently changed), get prioritized refactoring targets, or set up CI quality gates. Triggers on: fallow, dead code, unused exports, complexity hotspot, codebase health, refactoring targets, CRAP score, cyclomatic complexity, TS/JS codebase cleanup.

Discover and install agent skills from the open ecosystem. Use when the user asks "how do I do X", "find a skill for X", "is there a skill that can...", or wants to extend agent capabilities with installable skills. Triggers on skill discovery, skill search, skill installation.

Interact with GitHub using the `gh` CLI. Use when working with GitHub issues, pull requests, CI runs, or API queries. Triggers on: PR review, issue search, workflow run, gh cli, github api, CI status.

Turn vague personal, research, learning, or project intent into honest, stable, detailed, falsifiable, incremental goals. Use when the user wants to define success, choose direction, prioritize open-ended work, avoid scope drift, recover motivation, or clarify what they are actually trying to accomplish. Keywords: goal-setting, goals, define success, north star, scope drift, priorities, tradeoffs, motivation, stuck, falsifiable, incremental, what am I trying to do.

Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.

Design agent loops instead of one-off prompts: scheduled or goal-driven automation, worktree isolation, durable state, skills/policies, connectors/MCP, maker-checker subagents, cost budgets, safety gates, and human approval boundaries. Use when setting up recurring AI engineering workflows, repo maintenance loops, CI/PR/issue triage loops, long-running agents, autonomous or semi-autonomous coding tasks, or when replacing repeated manual prompting with a system that discovers, acts, verifies, records state, and escalates.

Create polished local-first HTML playgrounds and micro-apps for exploration, workshops, demos, and manager-ready artifacts. Use when the user asks for a playground, interactive HTML artifact, static GitHub Pages lab, self-guided exercise, visual explorer, copy/paste AI workflow, slide deck, diagram, data explorer, code review artifact, comparison matrix, timeline, throwaway editor, brainstorm grid, mind map, design token showcase, spec/planning doc, or HTML instead of markdown/slides. Keywords: playground, micro-app, static HTML, GitHub Pages, Alpine, Tailwind CDN, Popover API, View Transitions, local-first, copy prompt, export markdown, slide deck, kanban, ERD, mind map, data explorer, spec, implementation plan.

Steer toward modern web platform APIs instead of legacy patterns. Use FIRST for any HTML/CSS/clientside JS task — modals, dialogs, popovers, anchor positioning, container queries, :has(), :user-valid, View Transitions, scroll-driven animations, INP/LCP optimization, autofill forms, custom scrollbars, scheduler.yield. The CLI runs local semantic search over expert-curated, Baseline-aware guides and returns token-efficient implementation patterns, gotchas, and fallbacks. Do NOT trigger for backend, SQL, CI/CD, Docker, or generic local scripts. Keywords: frontend, html, css, web, browser api, popover, dialog, view transition, baseline, web platform, modern css, accessibility, a11y, forms, autofill, performance, core web vitals.

Publish npm packages with OIDC trusted publishing, GitHub release workflows, and version synchronization. Use when setting up npm publish CI/CD, configuring OIDC/trusted publishers, fixing version sync issues (package.json vs git tag), troubleshooting npm publish 403/404/ENEEDAUTH errors, creating release workflows, or publishing scoped/monorepo packages. Triggers: npm publish, npm release, OIDC, trusted publisher, package.json version, npm 403, npm 404, ENEEDAUTH, publish workflow, changesets publish, lerna publish.

Apply writing discipline to agent output: make it scannable when structure is needed, use prose when it's not. Use when user asks to make writing clearer, more scannable, easier to read, or references 'Don't Make Me Think.' Also triggers for requests like 'simplify this,' 'make this scannable,' 'cut the fluff,' 'help people skim this,' 'stop over-formatting,' 'sound less robotic,' 'less corporate,' 'write naturally,' 'just talk normally,' or when editing docs, emails, proposals, or technical writing for busy readers. Keywords: output style, write clean, scannable, simplify, cut the fluff, over-formatted, tone, formatting, prose-first, natural tone, conversational, minimum formatting, less corporate.

Create and maintain lightweight Markdown plans in locally excluded .ai/plans/ for complex, ambiguous, risky, or multi-session work. Use when the user asks for a plan, PLAN.md-style artifact, implementation plan, design plan, migration plan, execution plan, roadmap, handoff, or when reasoning, decisions, discoveries, validation, or recovery context should survive the chat. Tasks and checklists are optional, not required.

Evaluate product/app ideas before and after MVP using north-star one-pagers, core-tech leverage, defining constraints, success criteria, opportunity mapping, RICE, and growth horizons. Use when deciding whether to build, continue, pivot, prioritize post-MVP work, define product success, or avoid feature-factory drift. Keywords: build or not, MVP, product idea, north star, product strategy, RICE, opportunity tree, roadmap, post-MVP, pivot, success criteria, narrative, one-pager, working backwards.

Design repo feedback loops for AI-assisted engineering: deterministic local/CI gates, lint/custom rules, worktree-safe agent flows, maker/checker subagents, durable state, observability-to-tasks, cost/safety limits, and git protections. Use when hardening a repo for agents, reducing AI code drift, migrating CLAUDE.md guidance into enforceable checks, encoding software-design red flags as checks, adding CI/pre-commit/pre-push/protected-branch guardrails, or designing scheduled/goal-driven maintenance loops. For diagnosing which modules/interfaces should change, use software-design first; for writing behavior through tests, use tdd.

Build, review, and improve agent skills. Validates against the agentskills.io spec, evaluates quality via a dimensional rubric, detects description drift, produces numbered improvements, and applies them with approval. Use when: creating a new skill, reviewing/auditing/judging a skill or prompt, updating/fixing a skill, evaluating prompt quality, 'skillify this', 'make this proper', 'audit all skills', 'grade report'. Triggers: skill, SKILL.md, prompt quality, knowledge delta, skill review, skill audit, judge prompt, skill creation.

Diagnose architecture friction and design deeper modules/APIs, informed by domain language in CONTEXT.md and decisions in docs/adr/. Use when the user wants refactoring candidates, module/interface design, architecture review, testability improvements, or software-design reasoning about shallow modules, seams, information leakage, locality, and leverage. For implementing behavior test-first, use tdd; for turning conventions/design rules into deterministic linters/CI checks, use feedback-loop.

Analyze organizational, technical, and strategic problems as interacting systems using Rumsfeld Matrix, causal loops, debt cascades, constraints, and resilience thinking. Use when problems involve feedback loops, hidden dependencies, legacy systems, architecture portfolios, repeated failures, unknown unknowns, or change resistance. Keywords: systems thinking, causal loop, Rumsfeld Matrix, unknown unknowns, Chesterton's Fence, debt cascade, theory of constraints, enterprise architecture.

Test-driven development with red-green-refactor loop. Use when the user explicitly wants test-first work, says TDD/red-green-refactor, asks for behavior tests before implementation, or wants a bug fix driven by a failing test. For architecture diagnosis, module boundaries, or refactoring candidates, use software-design first; for repo-wide automated quality gates, use feedback-loop.

Build token-efficient, consistently-styled PDF documents with Typst — reports, specs, status updates, incident post-mortems, slide decks, diagrams, flowcharts, Gantt charts, comparison matrices, one-pagers, and CVs — and answer Typst syntax questions (markup, math, code mode, set/show rules, page and table layout). Use for composed, paginated, or printable documents needing real layout, a table of contents with page numbers, headers/footers, tables, figures, or math that Markdown can't do — cheaper to share than an HTML micro-app, compiles to PDF. Routes Markdown vs Typst vs HTML, pins a prose typography preamble for consistent styling, and maps each document type to the right typst.app/universe package. Keywords: typst, PDF, document, report, spec, slide deck, diagram, flowchart, gantt, status report, incident report, comparison matrix, one-pager, CV, resume, print, typography, typst syntax, markup, math, set rules, show rules, layout, cetz, fletcher, touying.