claude-skills

Conservative cross-repo worktree + branch pruning audit. Classifies every branch/worktree as safe-to-delete (proven dead via merged PR or zero unique commits) or keep-with-reason, then deletes what's proven dead. Use for 'clean up stale branches', 'prune dead worktrees', 'repo hygiene sweep', or any multi-repo branch-pruning ask.

Best practices for AI-authored writing — consolidated rules for when Claude (or any LLM) drafts prose a human will put their name on: emails, posts, newsletters, docs, scripts, bios, marketing copy. Use whenever AI-drafted output must not read as generic AI sludge; triggers on '/ai-writing', 'de-slop this draft', 'make this not sound like AI'. Pulls the AI-relevant rules from the writing and communication skills into one pre-ship gate. NOT a voice profile (use writing-voice for Summer's voice, sme-voice for someone else's) and NOT a conflict/timing layer (use communication-safeguards for heated messages).

Document a dead end so future sessions avoid re-exploring. Writes an abandon record to .claude/abandoned/<date>-<topic>.md capturing what was tried, why it failed, and what was learned. Invoked as `/session-handoff --abandon`, or when the user says "abandon this", "this didn't work", "scrap this approach", or "dead end". Also used by claude-flow Phase 1 (Explore/Archive path) and the research skill. Normal state-export handoffs are written by the `next` skill to docs/plans/<date>-session-handoff.md — not by this skill.

Captures another person's writing voice from samples and applies it when ghostwriting or editing in their voice — primarily for SME (subject-matter expert) scripts at DeepLearning.AI. Phase A builds a saved voice profile from samples; Phase B applies it when editing or extending the SME's scripts. Profiles live in ~/claude_code/agent-vault/sme-voices/ to sync across machines. Trigger on 'capture [name]'s voice', 'edit this in [name]'s voice', 'ghostwrite as [name]', or '/sme-voice'. NOT for writing as Summer herself (use writing-voice), ToneGuard analysis, or summarizing/analyzing documents.

Typography rules — quote marks, dashes, spacing, hierarchy, HTML entities. Auto-apply when generating HTML/CSS/React/Jinja with visible text; audit on "fix the typography", "typography audit", "make this look professional".

Loads Summer Rae's personal voice profile for any writing she'll put her name on — fiction, personal essays, blog posts, newsletters, social posts, course marketing hooks, L0 scripts, business emails and documents. Applies her creative voice (sensory, deadpan dark humor, show-don't-tell) and her business-writing preferences (minimal, direct, lead with the ask, no hollow openers). Default whenever Summer asks you to write or draft anything ('write this for me', 'in my voice'). NOT for ghostwriting or preserving an SME's voice (use sme-voice), ToneGuard analysis, or summarizing/analyzing documents.

Hand off to the next session — write a self-contained continuation prompt/handoff doc, then ship the current work (commit → PR → review → merge) and run cleanup. Use when the user says "/next", "hand this off", "wrap up and write a continuation prompt", or wants to end one session and prime the next.

End-to-end ship pipeline — commit, push, PR, 10-step review, fix, CI, merge. Use on "ship it", "done", "merge it", "push it", "deliver this".

Use when a founder wants to close their first customers, build a repeatable sales process, or decide when to hire a salesperson. Triggers on: founder-led sales, first customers, cold outreach to customers, early sales calls, sales discovery, when to hire sales, B2B sales, close a deal, sales process. Applies frameworks from 16 product leaders on outreach, discovery, closing, and repeatability.

Use when a founder is preparing to raise capital, deciding whether to raise at all, meeting with investors, or asking about fundraising strategy. Triggers on: should I raise, fundraising strategy, venture capital, pitch to investors, raise a round, investor meetings, fundraising process. Applies strategic frameworks from product leaders on narrative, process, and psychology.

Search GitHub code, repositories, issues, and PRs via gh CLI

Use when a founder or PM wants to assess whether they have product-market fit, measure user retention and engagement, run the Sean Ellis survey, or decide whether to scale or keep iterating. Triggers on: product-market fit, PMF, do I have PMF, retention curve, Sean Ellis survey, scale vs iterate, when to scale, measuring PMF. Applies frameworks from 46 product leaders.

Use when a founder or PM wants to design or optimize pricing strategy using willingness-to-pay research and product-led frameworks. Triggers on: pricing strategy, willingness to pay, WTP research, freemium vs paid, self-serve ceiling, pricing model, monetization, price increase, value metric. Distinct from the 'pricing' skill (SaaS tiers/packaging); this skill covers WTP-research-deep and iteration frameworks from 46 product leaders.

Dedicated bug fix orchestrator — Reproduce → Diagnose → Fix (TDD) → Verify. Composes existing debugging skills (investigator, systematic-debugging, TDD) into a streamlined pipeline. Use when fixing bugs, regressions, or unexpected behavior instead of claude-flow.

Use when creating new features, implementing complex changes, building features, adding features, executing implementation plans, or running TDD workflows. Multi-phase agentic build pipeline (Phase 0 context → 1 discovery → 2 exploration → 3 requirements → 4 architecture/plan → 4c verification → 5 TDD implementation → 5.5 reflection → 6 review+ship). Workflow paths: fast, clone, lite, plan, explore, audit, full. Executor/Advisor strategy — Sonnet executes, Opus advises at key decision points. Project-generic router; project-specific skill menu is injected via Phase 0/5 (see references/project-skill-menu.md).

Use when the user needs open-ended idea generation or selection help — "I want to make/build/write/start something", "I'm stuck", "inspire me", "make this weirder", "help me pick between these", "I need to invent X", naming a product/campaign, or unsticking a stale creative approach.

Run health checks on project memory files — broken links, orphan memories, stale entries, contradictions, oversized index lines

Use when you need to reduce LLM API spend, control token usage, route between models by cost/quality, implement prompt caching, or build cost observability for AI features. Triggers: 'my AI costs are too high', 'optimize token usage', 'which model should I use', 'LLM spend is out of control', 'implement prompt caching'. NOT for RAG pipeline design (use rag-architect).

Pick the optimal Claude model (Haiku / Sonnet / Opus) for the current task — two-layer router: mode detection (longContext, background, think, subagent-fleet, webSearch, default) plus weighted complexity scoring. Outputs a one-line recommendation card with model, mode, complexity, reasoning, cost anchor, and the `/model` or `claude --model` command. Use at session start, before dispatching parallel subagents (model choice multiplies cost across N agents), when paying Opus prices on mechanical work, or on 'which model should I use'. NOT for production LLM API spend (use llm-cost-optimizer), PR-reviewer mode (use pr-reviewer-mode), or picking skills (use skill-discovery).

Evaluate and optimize prompts using proven prompting techniques and frameworks. Activate whenever the user asks to "optimize," "improve," "evaluate," or "fix" a prompt, asks if a prompt is "good" or "clear," wants "better versions" of a prompt, asks why a prompt is producing poor results, or wants prompt engineering guidance. Also trigger when the user shares a prompt and seems unsatisfied with Claude's outputs — even if they don't explicitly ask for optimization. Do not use when the user wants the task executed directly ("just do it").

Use when the user says "/articles", "check my articles", "process Claude articles", "triage my reading queue", or asks whether anything in their Mem reading queue is worth pulling into their projects or skills. Pulls unread notes from Mem's "Claude articles" collection ONLY, runs each through `useful-for` against the user's claude-skills repo and active projects (CourierFlow, BetterBurgh, DeepLearning.ai work), and moves processed notes to "Claude articles — reviewed". Triggers on any phrasing about checking, processing, triaging, or reviewing the reading inbox in Mem.

Tailor an existing resume to one or several job descriptions with visible confidence scoring, structured reframing, and positioning help. Triggers on "tailor my resume", "update my resume for this job", "resume to job", "match resume to JD", "resume multiplier", "three resume versions", "multiple job descriptions", "resume keywords", "ATS alignment", "position my experience", "reframe my resume", or pasting JDs alongside a resume.

Use when the user wants to find off-market house-buying candidates — homes that aren't currently listed but where the owner may be motivated to sell (distressed, inherited, long-vacant, sheriff sale, tax delinquent). Two stages: `/off-market discover <county>` ranks candidates from free public data; `/off-market outreach <candidates.csv>` drafts personalized letters in the user's voice. v1 supports Allegheny County (Pittsburgh) via WPRDC; additional counties via plug-in adapters. Free data only — no PropStream / ATTOM / BatchLeads dependency.

Token-efficient Python web scraping. Use whenever the user wants to scrape, crawl, or extract data from a website, HTML page, online article, or GitHub repo — even when they don't say 'token-efficient'; triggers on 'scrape this site' or a pasted URL with an extraction goal. Keeps raw HTML out of Claude's context: fetch and parse in a Python subprocess, return only the structured fields asked for. Default: pick a bundled script, run it, write JSON to disk, show a head; multi-step exploration is the exception.

Real-time state, timing, and intent layer for any Slack message, email, or workplace communication Summer is drafting or answering. Runs alongside ToneGuard — checks internal activation (1–10), HALT signals, and intent before composition; provides the Two-Sentence Response template for conflict replies and the 2-Minute Reset for rumination loops. Trigger on 'how should I respond to', 'I'm fuming about', or any communication carrying conflict heat. Hands the final draft to ToneGuard for passive-aggressive / escalatory scanning.

Unified review skill with auto-tiering — absorbs PlanCraft and adversarial-thinking. Tier 0: single-conversation devil's advocate or steelman, no API calls, for pressure-testing ideas, plans, decisions, strategies, or assumptions ('attack this idea', 'find flaws', 'steelman', 'argue against'). Tier 1: DeepSeek scope check. Tier 2: DeepSeek + second critic. Tier 3: full debate. In Codex, use Claude/Anthropic as the second external critic unless the user explicitly requests GPT-5. Conditional Haiku Style critic for frontend.

Use when designing or running evals for an LLM feature — offline (golden datasets, LLM-as-judge and code-based evaluators, prompt/model/architecture experiments, regression gates) or online (production sampling, blocking guardrails for PII/jailbreak/toxicity, drift detection, shadow/canary rollout, eval-based alerting). Triggers on 'design evals', 'golden dataset', 'LLM-as-judge', 'eval harness', 'production evals'. NOT for prompt-registry mechanics or governance policy (use prompt-governance), variant promotion at scale (use prompt-optimization), tuning one judge prompt's wording (use prompt-optimizer — judge prompts are prompts), or RAG retrieval design (use rag-architect).

Use when authoring an implementation plan, PRD, design doc, or brainstorm artifact that beats its Markdown equivalent by inlining visual mockups, interactive elements, or component examples — specs where a flat text outline loses engagement. Triggers on 'HTML spec', 'PRD with mockups', 'interactive plan', 'living design system', 'micro-UI to edit my plan'. Composes with [`writing-plans`](../writing-plans/SKILL.md) and [`prd`](../prd/SKILL.md) (HTML output variants), [`claude-flow`](../claude-flow/SKILL.md) Phase 4 (visual design-doc mode), and [`excalidraw-canvas`](../excalidraw-canvas/SKILL.md) (whiteboard-style diagrams; html-spec for full single-page artifacts).

Use when the user wants to design, redesign, critique, audit, polish, or otherwise improve a frontend interface — websites, landing pages, dashboards, product UI, app shells, components, forms, settings, onboarding, empty states. Handles UX review, visual hierarchy, information architecture, cognitive load, accessibility, performance, responsive behavior, theming, typography, spacing, layout, color, motion, micro-interactions, UX copy, error states, i18n, and design systems or tokens; also bland-to-bold or loud-to-quiet rebalancing, live browser iteration, and ambitious visual effects. Not for backend-only or non-UI tasks.

Use when the user is considering granting an AI agent or automation access to a system, tool, or account and wants a one-page risk/permission brief before flipping the switch. Triggers on 'permission brief', 'agent risk review', 'scope this agent's access', or a pasted integration/MCP/connector setup with a 'should I let it…' question. Produces allowed/ask-first/never actions, spending and data limits, required logs, failure + rollback, and a low-risk first-week test plan. Pairs with skill-security-auditor (audits skill code) — this scopes the runtime privileges around the agent.

Use when authoring a Product Requirements Document — a durable, sharable spec — before handing a feature to claude-flow or another AI coding agent. Triggers on 'write a PRD', 'spec out [feature]', 'product spec', 'requirements doc'. Output maps onto claude-flow's $requirements contract (user stories, EARS acceptance criteria, scope in/out, edge cases, risk class, non-functional) so Phase 3 ingests it instead of re-asking. Distinct from claude-flow's inline Phase 3: use when the PM / founder / requester wants a spec that survives across sessions.

Router for all prompt-engineering work — dispatches to structured-prompt-builder (single-turn authoring), agent-prompt-architecture (tool-using agents), prompt-optimizer (improving), prompt-optimization (variant analysis at scale), or prompt-governance (production management); centralizes the shared Anthropic prompting principles each sub-skill enforces. Use when a prompt-related ask doesn't cleanly map to one sub-skill or needs the write → improve → govern pipeline; triggers on 'prompt engineering', 'how should I prompt Claude'. If the request clearly fits one sub-skill, invoke that directly.

Write, review, or improve DeepLearning.AI course scripts — Lesson Zero (L0) intros, marketing teasers, Course Map/Overview slides, diagram walkthroughs, Talking Heads (THs). Covers Andrew Ng voice, DL.AI guidelines, teleprompter-aware phrasing, iteration protocol, and an annotated reference-script library. Use for any course intro / L0 / marketing / teaser video script or script-quality review for launch materials. Supersedes the bundled anthropic-skills:sc-marketing-scripts — this version adds curated references and a progressive-disclosure loader.

Security audit and vulnerability scanner for AI agent skills before installation. Use when evaluating a skill from an untrusted source, auditing a skill directory or git repo URL for malicious code, gating installs of Claude Code plugins / OpenClaw / Codex skills, scanning Python for dangerous patterns (os.system, eval, subprocess, network exfiltration), detecting prompt injection in SKILL.md files, or checking dependency supply-chain risk and file-system boundary violations. Triggers on 'audit this skill', 'is this skill safe'.

Heavy, explicit-invocation startup planning workshop spanning Kawasaki Conception (genesis, sweet spot, mantra, business model, cofounder, MATT) and execution planning (Business Model Canvas, TAM/SAM/SOM, MVP/MoSCoW, Go/NoGo, Marp pitch deck), plus optional PRO/CON debate validation, AI stakeholder-clone rehearsal, and customer segmentation matrix. ONLY invoke on explicit '/startup-planner' or a named ask for the Business Model Canvas / Kawasaki workshop / Marp deck. Do NOT auto-fire on vague 'plan a startup' or 'business idea' — use startup-analysis for CLEARFRAME validation instead.

Build a production-ready XML-structured prompt through an interactive critique loop — drafts with Anthropic's Prompting 101 component framework, audits against a checklist, and iterates until clean and efficient. Use whenever the user asks to draft or design a prompt, system prompt, or LLM instruction set — classification, extraction, document analysis, multi-step reasoning, or any output-format-sensitive task — or describes a production LLM workflow without asking for a prompt. NOT for managing prompts in production at scale (use prompt-governance), analyzing performance across prompt versions (use prompt-optimization), or optimizing session context for an active agent (use context-engineering).

Tool-call-level discipline for minimizing session token burn without sacrificing correctness. Use at the start of a cost-constrained session, on heavy mid-session burn (file-slurping, serial searches, verbose narration), before /compact or handoff, or when briefing a subagent on efficient tool use. Teaches combine-discover-and-read, batched edits, parallelized searches, entry-point-first exploration, targeted line ranges, cheap-subagent delegation, DB introspection over schema-file reads, dense-prose compression of always-loaded context files, and preserve-vs-drop compaction categories. NOT for production LLM API spend (use llm-cost-optimizer), choosing which files the agent sees (use context-engineering), or claude-flow Phase 2 exploration (use smart-exploration).

Use for QC / cut-list / pacing analysis of a screencast, tutorial, or narrated recording — generates a timestamped incident log of stumbles, repeated phrases, filler words, dead air, inaudible audio, and visual issues (frozen view, stuck cursor, wrong tab, black frames, audio-visual mismatch) without modifying the source video. Triggers on 'review this video for pacing', 'generate a cut list', 'find dead air'. Companion to automated-screencast-director — that skill records screencasts, this one QCs them.

Use before any "applied", "shipped", "fixed", "merged", "deployed", "done", "complete", or "ready" claim — verify behavior at runtime, not just code state. Check the actual output (endpoint response, UI render, deployed PR state), cite concrete evidence (file_path:line link, git diff --stat, gh pr view, fenced command output), and surface gaps. Required when the claim involves runtime behavior; especially after Edit/Write tool successes (which do not render in the user transcript and are not evidence by themselves).

Market positioning strategy using the April Dunford 5+1 framework, enriched with JTBD discovery, Moore positioning statement, and Neumeier Onliness Test. Produces a positioning document, positioning statements, competitive alternatives map, and market category analysis. Triggers for "positioning", "market position", "differentiate from competitors", "positioning statement", "competitive positioning", "category strategy", "where do we fit", "how are we different", or any request to define, sharpen, or rethink product positioning. Works standalone — no prior session needed, but leverages startup-design or startup-competitors output when available.