juliuswiener
User🍪 Personal Claude Code marketplace — multi-agent skills (review/cleanup/plan/exec), a canonical task-routing hook, a custom statusline (nord-hud), plus an EE/hardware + dev toolkit.
Categories
Indexed Skills (42)
author-skills
Patterns for authoring/revising skills beyond basics: structure (XML vs markdown), tool-preference phrasing, composition, compliance monitoring, choosing skill vs hook vs CLAUDE.md vs command/subagent. Use when writing or revising a skill.
deepinit
Generates hierarchical AGENTS.md files across a codebase — one per significant directory, each linking to its parent, describing what lives there and how it fits together. Use when onboarding a repository that has no agent-facing documentation, or when the user says 'deepinit', 'generate AGENTS.md', 'set this repo up for Claude', 'dokumentier das Repo für Agenten'. To orient yourself without writing files, use `orient` instead.
diagnose
Find WHY something behaves as it does by running competing explanations against each other — evidence for and against each, a rebuttal round, then the one probe that would settle what is left, executed if it can be. Use when a test fails for reasons that are not obvious, when a measurement or benchmark result needs explaining, when two explanations both fit the facts, or on 'warum passiert das', 'wieso ist die Zahl anders', 'was ist da los', 'trace this'. When you can already reproduce and localise the failure, the `debugger` agent is enough. It ends at the decisive probe — applying the fix is `implement`.
find-deals
Best-deal sourcing + negotiation leverage. Triggers: 'find me the best deal', 'cheapest near me', 'is this a good price', 'buy now or wait', 'compare deals', 'negotiate', sourcing/validating vehicle deals.
interactive
Run a command that needs a real TTY or human input — sudo password, ssh passphrase, installer questions, TUI — somewhere the user can actually type. Agent subprocesses have no TTY, so such commands fail ('sudo: a terminal is required to read the password'). For a plain sudo command it uses `sudo -A` with a GUI password dialog and needs no terminal at all; for anything else it opens a real pane, preferring the tmux the user is already in and falling back to spawning a terminal window. Triggers: sudo needs password, interactive prompt, terminal required error, ssh passphrase, installer asks y/n, 'braucht ein Terminal', 'sudo geht nicht', 'da muss ich mein Passwort eingeben'.
manage-skills
Inspect and edit the skills already on this machine — list, add, remove, edit, search, copy between scopes, or set up the directories. Use on 'manage skills', 'local skills', 'add skill', 'remove skill', 'edit skill', 'welche Skills habe ich', 'Skill anlegen'. This is local CRUD; to find a skill in a marketplace use `search-skills` or `install-skill`, to write a good one use `author-skills`.
search-skills
Search the open agent-skills ecosystem (skills.sh / `npx skills`) for a skill that already exists, check whether it is trustworthy, and install it on request. Use on 'find a skill for X', 'is there a skill that…', 'gibt es dafür einen Skill', or when a capability is missing and someone has probably already built it. To write one instead, use `author-skills`; for the prompts.chat registry, `install-skill`.
dart-expert
Production Dart (3.x): null safety, async/streams, sealed classes + exhaustive pattern matching, records, Result-style errors, isolates, testing. Use when writing/reviewing Dart domain logic, packages, CLIs, serialization, or concurrency (non-UI).
pdf-extract
Convert a PDF / Office / scanned document to LLM-ready markdown locally via MinerU (GPU-accelerated on ROCm). Triggers on: 'read this PDF', 'extract from <file>.pdf', 'parse this document', 'get the tables out of', 'OCR this', after read-router picks a text+tables document. Output is TEXT (cheap downstream) — the default document reader, preferred over pixel-read whenever the document is mostly text and tables. Runs on-device; documents never leave the machine.
read-router
Pick the right web-data READ paradigm per URL or file before fetching — never default to one tool for everything. Triggers on: 'scrape this', 'read this page/PDF', 'extract from <url>', 'get the data from', 'pull this document', 'what does this site say', or any time content must be pulled from a web URL or a document file. Routes to web-scrape (HTML→markdown, local), doc-extract (PDF/Office→markdown, local GPU), pixel-read (visual/layout→images), or search (`research --web` / exa) when the URL is unknown. Local-first and data-sovereign by default.
visual-read
Render a web page or document to images (tiled JPEGs) so Claude can SEE layout that text extraction loses — charts, infographics, scanned forms, complex tables, visual diffs. Triggers on: 'look at this chart/figure', 'what does this dashboard show', 'read this scanned form', 'the layout matters', or when read-router / web-scrape / doc-extract flag that the meaning is visual. Local (pixelrag pixelshot), no backend. EXPENSIVE in tokens — use only when layout IS the information, not as a default reader.
web-scrape
Fetch a web page to clean markdown locally via Crawl4AI — no external API, no per-request cost, content never leaves the device. Triggers on: 'scrape <url>', 'get the text/content of this page', 'read this website', 'extract the article', after read-router picks a normal or anti-bot web page. Default for ~90% of pages. Use --stealth for anti-bot/login/heavy-JS sites. Firecrawl (external) is only a last-resort fallback for the rare page local can't crack.
install-skill
Find + install Agent Skills from the prompts.chat registry via MCP (search_skills/get_skill). Use to discover/install an existing skill before hand-rolling one. Requires the prompts.chat MCP connected.
implement
Build something behind a deterministic gate — a test, a linter or a compiler exit code is the only verdict, never a self-assessment and never a judge model. `--from task` for one story (the only mode a worker may drive), `--from goal` to decompose a goal and drive it until everything is green, `--from plan` to take the split from a finished plan, `+ --parallel` for disjoint stories at once. Use when the user says implement, build, fix, 'make the tests pass', 'bau das', 'mach das fertig', 'implementier das', 'setz den Plan um', or hands over a plan to execute. Cheap $0 workers do the volume and this thread takes over after three consecutive reds; needs Claude Code launched through claude_bridge.
orient
Find out where you are in an unfamiliar repository without reading files — returns the whole-repo symbol skeleton ranked by PageRank over tree-sitter tags, so the answer is 'here is the important code' rather than 'here is all the code'. No network, no build step, no reasoning pass; about a second in any tree. Use before planning a change in a repo you do not know, when asked to map or explain a codebase, or on 'wo bin ich hier', 'was ist das für ein Repo', 'zeig mir die wichtigsten Stellen'. For ONE file use smart_outline; for what depends on what use the graphify skill.
plan
Decide what to build and how, and stop at whatever maturity you actually need. `--stage ideas` for an idea board, `--stage shortlist` for a ranked set with the weak ones killed, `--stage spec` for pinned requirements via one question per round; with no stage it runs the planning tournament and returns a plan. Add `--deep` for post-tournament Planner→Architect→Critic validation on vague or expensive work. Use for planning a change, exploring options, deciding between ideas, pinning down what is actually being asked, or on 'plan das', 'wie gehen wir das an', 'welche Optionen haben wir', 'was genau soll das können'.
prompt-engineering
Writes and repairs prompts for language models — the section order a prompt needs, which technique fits which signal, per-model formatting, and how to diagnose one that gives inconsistent results. Use when writing a prompt for a model, optimizing or rewriting one, designing an agent or skill instruction, when a prompt produces varying output and the cause is unclear, or on 'schreib einen Prompt', 'Prompt verbessern', 'warum antwortet das Modell so unterschiedlich'. For a Claude Code skill, subagent or CLAUDE.md the artifact rules differ — see the artifact section below, and `author-skills` for the full set.
review
Judge finished work and return severity-rated findings, each anchored to a file:line that was actually read. Three axes: `--scope diff|repo|plan` (what is under review — a change, a whole codebase, or a plan before anyone builds it), `--lens security|a11y|claims` (narrow to one question), `--deep` (parallel specialists with every finding adversarially verified, instead of one pass). Defaults to a single pass over the current diff. Use for reviewing a diff or PR, auditing a codebase, checking whether a plan is sound, asking whether the documentation is true, or on 'schau dir das mal an', 'review das', 'ist der Code gut', 'stimmt das so'.
flutter-ui
Flutter UI/UX: components, screens, animations, responsive layouts, custom widgets/themes, performance + accessibility. Triggers: 'create/build Flutter UI/screen', 'Flutter animations', 'responsive Flutter layout', 'custom Flutter widgets', 'Flutter theme'.
bom-manager
BOM management for electronics; orchestrates DigiKey/Mouser/LCSC/element14/JLCPCB/PCBWay/KiCad. Triggers: component sourcing, pricing, ordering, BOM export, fab prep, 'what parts do I need', 'cost estimate', 'order components', 'compare pricing'.
circuit-design
Circuit Weaver entry point: new design (wizard + IC selection + passives + schematic gen) or open existing for review. Triggers: 'design a circuit', 'new design', 'circuit-weaver', a design directory path.
datasheet-extract
Extract structured specs from component datasheet PDFs: pinouts, electrical characteristics, peripherals, topology. Triggers: 'extract datasheet', specs for MPN X, verify extraction, check pin functions.
design-wizard
Interactive guided circuit-design wizard: requirements to IC selection to BOM to schematic to PCB, quote-ready KiCad. Triggers: 'new design', 'start a project', 'design wizard', 'help me design', 'I want to build', 'new board/circuit'.
digikey-search
Search DigiKey for components and download datasheets via API. Triggers: component/part search, datasheet, pricing, stock, MPN, footprint, 'sync datasheets', distributor lookup. Default prototype distributor + datasheet source.
ee-reference
Electrical/electronic engineering reference: circuit analysis, component selection, power supply, signal integrity, RF, thermal, EMC, test & measurement. Triggers: EE design questions, calculations, component vetting.
element14-search
Search Newark/Farnell/element14 for components: pricing, stock, datasheets (free API). Triggers: Newark, Farnell, element14, non-US/regional distributor, cross-region price compare.
emc-precheck
EMC pre-compliance risk analysis for KiCad PCBs (FCC Part 15, CISPR 32/25, MIL-STD-461G). Triggers: EMC, EMI, radiated/conducted emissions, ground plane, decoupling, clock routing, 'will this pass FCC', EMC test plan.
jlcpcb-fab
JLCPCB PCB fab + assembly: BOM/CPL, basic vs extended parts, design rules, ordering. Triggers: JLCPCB, order PCBs/assembled boards, JLCPCB design rules, PCB cost/turnaround.
kicad-analyze
Analyze KiCad projects and PDF schematics: review designs, extract BOM, trace nets, DRC/ERC/DFM, power trees. Triggers: KiCad, .kicad_sch/.kicad_pcb/.kicad_pro, schematic/PCB review, Gerber, netlist, 'check my board', 'review before fab'.
kicad-docs
Generate engineering docs from KiCad: HDD, CE technical file, ICD, design review + manufacturing packages; render schematic/PCB SVGs and block diagrams to PDF. Triggers: 'generate documentation/report', HDD, CE file, ICD, render schematic/layout, block diagram, generate PDF.
lcsc-search
Search LCSC (JLCPCB parts library): stock, pricing, datasheets by Cxxxxx or MPN (free API). Triggers: LCSC, JLCPCB assembly parts, production sourcing, Cxxxxx numbers, BOM for JLCPCB assembly.
mouser-search
Search Mouser for components: pricing, stock, datasheets. Triggers: Mouser, DigiKey out of stock, cross-distributor price compare, parts DigiKey lacks.
pcbway-fab
PCBWay PCB fab + turnkey/consigned assembly: design rules, ordering. Triggers: PCBWay, turnkey assembly, non-LCSC parts, globally-sourced assembly, PCBWay vs JLCPCB.
spice-sim
Auto SPICE simulation of subcircuits from KiCad (ngspice/LTspice/Xyce): filters, dividers, opamp gain, LC resonance, crystal load. Triggers: simulate, run spice, verify with simulation, check filter cutoff, divider voltage, opamp bandwidth.
vivado-fpga
Xilinx/AMD Vivado FPGA build: batch TCL, ADI HDL reference designs, Zynq-7000/MPSoC block design, synth/impl/bitstream. Triggers: Vivado, Zynq, UltraScale+, ADI transceiver/HDL, FPGA build.
abort
Release an active persistence loop started by `implement --from goal|plan`, so the gate-persist Stop-hook stops refusing to let the session end. Use on 'cancel', 'stop the loop', 'abort the implement run', 'kill the persistence', 'abbrechen', 'stopp die Schleife', or when a session will not finish because stories are still red.
cleanup
Make a repository tidier, in one of two ways that are never mixed. `--remove` runs parallel detectors for dead code, duplication, AI slop, over-abstraction and unused dependencies, verifies each candidate is safe to delete, and returns a deletion-first plan. `--move files|modules` relocates instead: docs, scripts and tests out of the root into conventional folders, or modules regrouped vertically by feature. There is no default — you have to say which. Use on 'clean this up', 'remove dead code', 'reorganise the repo', 'räum das auf', 'aufräumen', 'toten Code raus', 'sortier das Repo'.
commit-message
Write one commit message — Conventional Commits, subject ≤50 characters, a body only where the 'why' is not obvious. It outputs the message and nothing else: no staging, no commit, no amend, no push, which is what the `-message` in the name is for. Use on 'write a commit', 'commit message', 'generate commit', 'schreib einen Commit', 'Commit-Nachricht', or when changes are staged and the text is missing.
research
Answer a question that needs searching rather than deciding. By default it investigates THIS repository: a goal is split into 3-7 staged hypotheses, each stage routed by complexity, up to 16 agents concurrently, cross-validated for contradictions and gaps — local files, grep and git only, never the web. `--web` instead decomposes the question into 2-5 parallel lookups against documentation, APIs, specs and standards, for what this repository cannot answer. Use on 'find out how X works', 'research this', 'wie funktioniert das hier', 'finde raus', 'recherchier das'.
verify
Prove a change works before anyone claims it does — run the narrowest decisive command and report only what it actually showed. Use before saying done, fixed or passing, and when the user asks 'does this work', 'is it done', 'verify my fix', 'confirm it passes', 'check if tests pass', 'läuft das', 'ist das fertig', 'hast du das geprüft'. Runs in this session and spawns nothing.
organize-modules
Organize modules vertically by feature domain, not horizontally by language construct. Use when creating new modules, reorganizing code, or deciding where to put types and functions.
circuit-sim
Circuit simulation dispatcher: SPICE subcircuits (-> spice skill, ngspice), RF chains (scikit-rf), power/clock transient + AC. Triggers: simulate, run simulation, check stability, ripple analysis, S-parameters, RF chain.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.