Development

Use when the user wants an idea taken from conversation to implemented, pushed branches in one pass — "/session-build <ideia>", "builda isso", "brainstorma e implementa", "aprimore ao máximo", "faça as melhorias", "vamos mapear e escrever estas specs", "pega essa ideia e leva até a branch pronta". ALSO invoke mid-conversation without being asked: if this conversation has already produced a spec, a plan, or a code edit toward a multi-part build, stop and invoke this skill before continuing — y...

A planning and unblocking copilot for AI-assisted coding. It interviews the user, scans the real repository with deterministic scripts instead of guessing, and writes the documents a coding assistant needs — prd.md, architecture.md, tasks.md, database.md — plus focused recovery plans when you are stuck. Use when the user wants to start vibe coding, plan an app, write a PRD / spec / architecture / data model / build plan, turn an idea into requirements, audit an existing codebase, debug a myst...

> 适用范围：本技能仅用于**交互式对话中响应真人用户的开发请求**。如果你收到的是"当前阶段: X / 执行指令"这类被指派的阶段任务（你是工作流子代理），**不要使用本技能**，直接执行你的阶段工作。 你在一个 Forge 工作区里，你是【编排者】：自己一行代码都不读、不写、不改。所有真正读/写/分析本工作区仓库代码的活，都必须派【真·子代理】去做（独立进程、cd 到项目目录、加载该项目 skills/rules、显示在检查器）。**绝对禁止**用你 CLI 内置的 Task/subagent 自己把活干了——那不是 Forge 的子代理，违背核心设计。你有两条派活的路径，按"是否跨阶段"判断走哪条：

Installs a timestamp hook into a target project's .claude/settings.local.json so every prompt in that project arrives stamped with the current date, weekday, and time. Use when asked to "add the time hook", "make this project time-aware", "install timestamps in this project", "put the timestamp thing in another folder", or when Claude keeps not knowing the date or how much time passed between messages. Creates or merges the settings file, adds gitignore coverage, verifies the command, and nev...

Add Arcjet security protection to HTTP routes, AI agent tool calls, MCP servers, background jobs, and queue workers. Covers rate limiting, bot detection, email validation, prompt injection, sensitive information blocking (including Rampart NER), content moderation, capture/flush, remote Guard policies, typed inputs, and abuse prevention. Works in JavaScript/TypeScript, Python, and Go. HTTP frameworks share this skill. JS Guard adapters that ship on npm are per-adapter reference files loaded f...

Use when creating a new spec-driven workflow — the /specode:spec entry that walks requirements → design → tasks → 执行方式 → execute → acceptance, autonomously calling superpowers skills for the heavy lifting with a first-class specode-native fallback, and filing 4 fixed artifacts (requirements.md / design.md / tasks.md / implementation-log.md) under the user document directory. Trigger /specode:spec <request>, or 进入 spec 模式 with a new requirement.

Use when running specode's requirements phase — project analysis + code location + clarification, then writing requirements.md. Produces <specsRoot>/<slug>/requirements.md (with the project_root frontmatter contract) plus the location pointers handed to the design phase. Invoked by name (specode:intake) via the Skill tool by the specode orchestration SKILL; not a user-triggered command.

Use when building or refreshing the RagKit index for a project knowledge-base/ (e.g. after distill added or updated knowledge points). Trigger /ragkit:embed. Querying instead uses ragkit:query; index health uses ragkit:status.

Extracts durable, reusable engineering knowledge from completed work, debugging sessions, migrations, incidents, and difficult implementation tasks. Converts verified experience into compact, evidence-backed Solutions, Lessons, Decisions, Constraints, and Workflows while rejecting task noise, duplicates, and unsupported conclusions. Read-only: produces knowledge proposals; memory-edit applies them.

Evidence-based project knowledge classification skill. Classifies verified repository findings - including claims extracted from existing multi-origin knowledge sources (AGENTS.md, CLAUDE.md, .cursor/rules/, .claude/, docs/) - into current facts, architecture, decisions, solutions, lessons, constraints, workflows, reference, historical, or obsolete knowledge; determines current-state status, durability, and knowledge value; resolves cross-source conflicts with evidence; detects semantic dupli...

Use when the developer asks you to hand work to someone else - spin up an implementer, get a reviewer on it, put a researcher on a question - or when you need to know who else is working on this project.

Use when a decision needs the developer, when work is finished and they need to understand what it means, when a spec needs approving before you build, or when you are stuck - writes the report page and decision they read to make the call.

Use when answering a question from the developer in a Bench chat turn - writes the answer as a rendered page rather than a wall of prose.

Edit a published or draft post in content/blog/ (copy, frontmatter, images), keeping the two dates and the URL intact, then prove it with the content lint and the build. Use when asked to update, revise, fix, reword, retitle or re-image a post, or to change its SEO fields; a rewrite that needs an outline and a critique is write-post in revise mode.

Review a content file against content/VOICE.md for what the lint cannot judge - claims, numbers, regulation naming, rhythm, structure - after quoting the lint's own findings on it, and report HIGH / MEDIUM / LOW with a verdict. Use when asked to brand-review, voice-check, lint, audit or read a draft or a page for voice, claims, AI-tic prose or readiness, or "is this on-brand" / "is this ready to publish".

Take a post or a page off the site without breaking an indexed URL - a post becomes a draft (its URL stays indexed), a page is removed only once a redirect is in place - and prove it with the lint and the build. Use when asked to retire, unpublish, hide, take down, archive or delete a post or a page.

Add a blog post to content/blog/ as a draft, from a brief, a draft file or the workshop's drafts, with its images, frontmatter and the load-bearing body conventions, then prove it with the content lint and the build. Use when asked to add, create or import a post, bring a draft into the site, or publish a draft; to write one from scratch, write-post runs first and hands off here.

Add a site page as content/pages/<slug>.yaml made of existing section types, with its SEO block, structured data and OG image, then prove it with the lint, the build and a visual capture. Use when asked to add, create or scaffold a page, a landing page or a route on the site.

Critical editorial review of one post draft - a prose-quality critique as the reader content/VOICE.md describes (dead transitions, paragraphs that do not earn their place, ungrounded generalities, AI-tic sentences the writer rationalised, cuts the writer would not make) and a two-pass fact-check of every externally verifiable claim against primary sources, with a verdict, a confidence and an action per claim. Read-only on the draft. Use before a post ships, from write-post at its critique sta...

Report the state of the site's content from the files, read-only, with nothing invented - posts by date with drafts and stale drafts, pages and when their copy changed, FAQ sets, reviews, use cases, the editorial calendar and backlog, the workshop's briefs in flight, recent changes, and the content lint's warnings by rule. Use when asked what is live, what is in draft, what is planned, what changed recently, or for a content status, blog status or editorial snapshot.

Design a site built on agentic-cms, from the wireframe to a look of its own or from one look to the next - tokens first, then the chrome, then the sections - one element at a time through candidates the owner picks from, a build on the dev server the owner looks at, and a pixel proof before anything merges. Use when asked to design, restyle, theme, brand, redesign or "make it look like" anything on the site, or to start designing a site from scratch.

Prove a change on the site's pixels with the screenshot harness - a baseline of the commit it starts from, a capture after, a compare read as a report - so a refactor shows it moved nothing and a design change shows it moved only what it meant to. Use before any merge that touches src/, styles or images, when asked to prove, verify, check parity or "make sure nothing else changed", and after a `SIZE`, `CHANGED` or `reflow` line needs reading.

Run a design round the way it works - the owner says what they want or what looks wrong, the agent brainstorms three to five ideas in one message, the owner picks which to build, the agent builds them as real components on a throwaway demo route (`pnpm kit demo new`) in the section's real frame with the page's real copy on the site's own theme, the current version last, the owner looks on the dev server and picks by letter or edits the pick in words, and on the owner's word the winner becomes...

Turn a screenshot claim ("unaligned", "too big", "appears too early", "looks broken", "cut off") into numbers before editing anything - the element's box, computed styles, stacking contexts, reveal state and timeline from `pnpm kit probe`, a crop from `pnpm kit shot` - so the fix follows the measurement. Use whenever the owner sends a screenshot or describes what a page looks like, and before touching any layout, spacing, size, z-index or timing.