Coding & Development

Stops false completion ("hallucinated success"). Before any claim that work is done, fixed, working, passing, deployed or ready, the claim is turned into an acceptance contract written before the code, each criterion is proven by a command or a file whose receipt (exit code, output, working-tree fingerprint) is written by a script, an independent checker re-runs the proofs, and only the computed verdict (DONE, NOT DONE, BLOCKED) may be reported. Use when starting a feature, fix or migration that must really work, before saying done, before a commit, PR, deploy or handover, when a subagent reports success, or when the user doubts a result. Optional Stop hook refuses to end a session on an unproven done. French triggers (preuve avant de dire termine, verifie que c est vraiment fini, ne me dis pas que c est fait sans preuve, criteres d acceptation, fausse completion, il dit que c est fini mais ca ne marche pas, verification avant livraison, prouve-le). Local routing (proves one task's acceptance; for launch read

Use in a monorepo when a change crosses components — designing or adding an HTTP route, a client call, a registry entry (artifact, tool, event) or a contract version, checking whether something that does the job already exists, following or documenting a flow from the app through the BFF to the services, and before finishing such work or opening a PR. The deterministic `atlas` CLI generates the facts and flow drafts; you write the intent (titles, roles, decisions) and interpret component boundaries, and atlas checks what you wrote.

Compile A-ISL / ISL specs to bin-<stack>/ using a coding agent (no paid isl-generator / Gemini / OpenAI). Use ONLY when the user explicitly asks to build, compile, or regenerate implementation from .isl.md after the Spec phase. Prefer preflight, queue --next, and light compile packets; agent writes bin-<stack> files. Never compile unprompted after Spec or UX chat.

Read-only readiness gate for one tracker issue. Determines whether an AI coding agent can safely start implementing the issue right now against the current repository state, and reports exactly one of PREFLIGHT READY, READY WITH ASSUMPTIONS, or BLOCKED with concrete evidence. Use only for explicit requests such as `/preflight ABC-17`, `$preflight <issue URL>`, "preflight this ticket," or "check whether this ticket is ready to implement." Never edits the issue, never edits code, never implements anything. Do not trigger for rewriting a ticket (use `prepare`), implementing one (use `do-ticket`), reviewing a pull request (use `pr-review`), or general architecture questions.

Sets up and runs the agent-driven hardware-in-the-loop engineering method for a hardware project - generator-first KiCad boards with Freerouting and live part verification, parametric CadQuery CAD with plug-in parts and mass/CG tables, host-tested firmware against a digital twin, a PowerShell CI matrix with no-shortcut and portability audits, dev/main branching with merge locks, machine-safety slots, detached jobs, usage-limit handoffs and overnight sessions, in either single-session or swarm mode. Use when the user says "hardware in the loop", "set up a hardware project", "PCB/CAD/firmware swarm", asks to scaffold a hardware repo for agents, or wants agents to design boards, enclosures or firmware with gates.

Which free MCP server or framework to reach for, and its traps — finance/gov data (SEC EDGAR, Yahoo, SAM.gov, USAspending), web research (Exa, Firecrawl, Crawl4AI, Jina reader), design/art/brand (Canva, Inkscape, Penpot, Figma, Orshot, Hugging Face image gen, vectorizing, charts), data/spatial (DuckDB spatial, live ArcGIS layers, OpenStreetMap/Overpass, Census, MapLibre), and ops (PostHog, iMessage, context7 docs). Use when making a social post, OG card, poster, logo/vector, chart, map, or brand asset; when querying parquet/geojson/duckdb or a city's open-data layer; when an MCP "needs authentication" or fails; or when asked "what tools do we have", "is there an MCP for X", "should we add X".

Build a data application on magmalake as one Mojo project, with pipelines run by orca, a lake of Iceberg tables over S3, and a long-running flare web app reading them, then deploy it to AWS with `orca aws`. Use when starting a new magmalake project or laying one out (pixi, mojoshelf, `bin/`, `pipelines/`, `src/`); when a pipeline step writes Iceberg tables that an app reads; when writing a flare web app, console or dashboard over the lake; when running the project locally with Postgres and a local warehouse; or when deploying to AWS (`orca aws up`, `push`, `register`, `app`, `status`, `down`), setting up its hostnames, secrets, public bucket or costs. Pipelines in depth (partitions, keyed, sensors, dista tasks) are the `orca` skill's.

Run a feature from idea to merge in two linked stages - first grill the user relentlessly until every design decision is settled (Matt Pocock's grill-me then to-spec), then carry that shared understanding through spec-driven development (DeepLearning.AI/JetBrains SDD course) - constitution check, branch, dated spec directory with requirements.md, plan.md and validation.md, implement task groups, validate, changelog, merge and replan. Use when the user says "grill me then spec it", "grill to spec", "spec-driven", "SDD", "feature spec", "next phase", or wants a non-trivial change planned and built against a written spec.

Decision-first experiment coach for product bets. Use when a PM or builder wants to know whether to build, invest, price, or keep pursuing something; what to test first; whether a test is even worth running; or what a finished test actually proved. Triggers on "should we build this," "how do I validate," "what's the riskiest assumption," "customers asked for it," "here are my results," or any belief that needs evidence before money or time is committed. Also use it for A/B tests, fake doors, pilots, betas, and pricing tests, and when someone asks how many users or how long a test needs.

Write and run distributed Mojo programs with dista, magmalake's Ray-style task framework, where one compiled binary is the driver, the head and every worker. Use when writing or reviewing Mojo code that imports dista or mentions distributed Mojo, a remote task, `Task`/`Task2`/`TaskMany`/`TaskWith`, `ObjectRef`, `rt.remote`, `rt.get`, `Seq`, an actor with methods or checkpoints, a cluster, a head, workers, `DISTA_ROLE`/`DISTA_HEAD`, the `dista` command (status, drain, scale, timeline, stop), draining, autoscaling, or a dista runtime inside an orca pipeline step; and when a dista job hangs, is refused with a registry mismatch, or fails with a lost result. Pairs with Modular's mojo-syntax skill and magmalake's mojo-data-performance skill.

Load test any project's API or web service with Python, locust and uv. Scans the code base for every endpoint, sorts them into core, support and peripheral features, asks the user which test level to run for each (smoke, expected load, stress) and which user-load scenario to simulate, with options derived from the project's own traffic, limits and timeouts. Then it builds a locust harness, runs it safely, collects server metrics and CPU profiles, and writes an HTML report on capacity, the knee and how the service fails. Use it whenever the user asks for a load test, stress test, soak or spike test, a performance test of an API, capacity or throughput planning, latency under load, or "how many users can it handle", or mentions locust, even if they name only one endpoint. Do not use it for micro-benchmarks of a single function or for frontend page-speed (Lighthouse) audits.

Synthesizes, audits, benchmarks, and refines AI Agent Skills adhering to the agentskills.io open standard. Performs competitor teardowns, ingests domain literature and practitioner transcripts into mental models, auto-discovers MCP tools, enforces 6-tier compaction-resilient layout, writes companion deterministic validation scripts in scripts/, audits against the UC Irvine 26-Smell Taxonomy, and runs verified evals. Trigger on: "create skill", "build a skill", "write skill", "skill creator", "make a skill from notes", "turn this into a skill", "audit skill", "audit an existing skill", "audit existing skill", "build commercial skill". Do NOT use for one-off shell scripting or general codebase refactoring without skills.

Records the user's actions — browser (Playwright event stream + trace) or native desktop (event/window/screenshot stream with accessibility targets and password masking) — on macOS, Windows and Linux, asks the user what the recording cannot show, then turns the demonstration into a reusable agent skill, optionally exported as a ZIP for claude.ai. Use when the user asks you to watch them perform a task, record a workflow, or create/refine a skill from their demonstration, even if they never say "skill". Agent-agnostic — works in Claude Code, Codex, opencode, or any agent with shell access.

小黑式拆解配图工作室。像 Ian 小黑那样先把中文文章读透、拆成“认知锚点 → 配图拆解方案(shot list)”,再按需生成原创 IP 小灯出演的怪诞手绘正文配图;融合小互 IP Studio 的逐节枚举、深层提炼、三轨分流(情绪/解释/四格)、隐喻逐一对照、比例规则、生图脚本和可选换角。用于“分析怎么配图”“拆解配图方案”“配图 shot list”“小灯配图”“正文配图”“文章插图”“隐喻配图”“帮我生成配图”“去标题/改图”等任务。

Own a repo's `.wiki/`, the memory that agents read and maintain themselves. At the end of a session (or during it) it looks at what happened and routes: a daily note, a fix to a topic page, a new map of content with its facet pages, or "nothing worth recording". Then it fixes the dates, regenerates the day index, lints, commits only its own files and syncs. Also looks things up in the wiki cheaply (rg first). Use when the user says "update the wiki", "wiki update", "note this in the wiki", "crea

从绘画构建完整可游览的 Three.js 三维世界，先生成左、右、背后三张观看图，建立四向物件数量清单，再逐件建模、组合并用实际截图验收。用于名画沉浸空间、绘画场景重建、系列世界制作及四周空间和主体细节修正；不用于仅生成一张三维风格图片。

Run an Android app on the owner's phone or the project's emulator — start and stop the emulator, pick the device, build, install, open the app and take screenshots. Use whenever a task needs the Android app running, on a device or emulator, or a screenshot of it, or when the owner asks to start, stop or check the emulator.

Build UX/UI suggestions as tappable mocks in the app's UI lab, record each round in docs/ux-history, and promote the chosen option into the app once the owner approves a plan. Use whenever the owner asks to suggest, mock up, compare or try UX/UI designs for a screen, or to bring back an earlier suggestion.

Works a local issue board shared by a team of coding agents through the earmark CLI. Joins the board as a worker, reviewer or pm profile, claims the next issue in scope, reads only what changed, hands work off to review, files work only another specialist can do (images, audio, copy) as blocking issues, approves or rejects handoffs, and triages and assigns the backlog. Use when the project has a .earmark/ board, when the user mentions issues, tickets, a backlog, the board, a review queue or earm

Compile ONNX models to run on Qualcomm NPUs (Hexagon HTP, via the QNN / QAIRT SDK) for Android apps, and run them with sherpa-onnx or ONNX Runtime. Use when the user wants a model (speech, audio, vision, small transformer) on a Snapdragon phone's NPU for speed or battery, or asks how to compile/convert/quantize a model for QNN, HTP, Hexagon, Snapdragon or "the NPU".

Builds and maintains PHP APIs with Kipchak, the API development kit on Slim 4 and FrankenPHP worker mode. Covers routes, controllers, DTOs with OpenAPI attributes, models, tests, configuration, drivers and middleware, and writing new drivers and middleware. Use when a project's composer.json requires kipchak/core, when the user mentions Kipchak, or when asked to build a lean, explicit PHP API with Kipchak.

Open a local page with the context tree of this project, for each mode of the main session and each agent, showing what is loaded at start (Claude Code's own files, the hecpack hook, the agent body) and what it may read by link. It also lists the files that cost most, the rules by path, the docs nobody reads and the repeated text, with the change since the last run. A script builds it with no model call. Use when the user runs /hecpack:context or asks how much context an agent or mode starts wit

Portable outer-operator procedure for durable Argus missions across five hosts.

Coordinate coding agents on a shared local project with task ownership, file claims, Git checkpoints, context handoffs, independent review and recovery. Use when the user requests collaboration between agents such as Codex, Claude Code or Antigravity.