Agents & Skills

Create 6 agent .md files, each with a system prompt defining its analytical lens. Write one orchestrator skill that accepts a scripture reference, spawns the selected agents in parallel using Claude Code's Agent tool, collects their outputs, and compiles them into a single report. The orchestrator should present a selection menu so users can pick which lenses to run. Store output in a dated folder structure.

Write two skill .md files in ~/.claude/skills/. The /cook skill's prompt should instruct Claude to: calculate total time from the recipe, add a configurable buffer (5-20% based on skill level), output a clock-based runsheet with exact quantities per step, and insert pitfall warnings before risky steps. The /recipe-from-ingredients skill should accept a fridge inventory, ask for skill level and serving count, generate a recipe, then suggest handing off to /cook.

Write a CLAUDE.md that acts as a setup agent. On first message, it should detect the OS (check for ~/.claude/ on Mac/Linux or %USERPROFILE%\.claude\ on Windows), ask the user's name and role, scaffold a global CLAUDE.md with their preferences, and create a projects directory. Include a preference-capture skill that auto-triggers on corrections ("don't do X", "always do Y") and saves them to a memory file. End with a verification step that reads back what was created.

Create 8 agent .md files in ~/.claude/agents/. Each agent prompt should define its research scope and output format. The company-research agent needs web search access (WebSearch/WebFetch tools) and should output a structured brief with sections for product, ICP, tech stack, culture, and pre-built interview answers. The interviewer-research agent should take a LinkedIn URL and build a profile of their values, hiring style, and predicted questions. Use a CLAUDE.md in the project directory to define how agents hand off to each other and where research files get saved.

Start with the 8 agents from Job Search Intel. Add Google Calendar MCP (@anthropic/google-calendar) to your .mcp.json. Write a Python script that queries your calendar for events containing "interview" or "call with," extracts the company name, and spawns the company-research agent for each one. Schedule it with Windows Task Scheduler or cron. Output goes to a research/ folder organized by company name.

Create 5 agent files. The company-research agent should scrape the company website and output a structured brief. The interviewer-profiling agent takes a name and company, searches LinkedIn/podcasts/articles, and builds a behavioral profile. The answer-frameworks agent should accept a job description and your resume context, then generate STAR stories mapped to each job requirement. All agents need WebSearch/WebFetch tool access in their prompts.

Write two agent .md files. The company agent should accept a URL, use WebFetch to scrape it, then WebSearch to find job postings, news, and funding info. Structure the output as: Company Overview, ICP Fit Score, Pain Hypothesis, Tech Stack, Stakeholder Map, Discovery Questions, Demo Customization, and Objection Handling. The attendee agent should search for the person by name, find their LinkedIn/public content, and output: Career Arc, Role in the Deal, Predicted Behavior, Rapport Builders, and Opening Gambit. Both agents need your product context in the prompt so they can personalize output. Use a setup wizard in CLAUDE.md that researches your product automatically from a URL on first run.

You need: Python with sounddevice for WASAPI loopback capture, openai-whisper for transcription, and sqlite3 for storage. Write a /start skill that launches a Python recording process in the background, capturing both system audio and microphone to separate WAV files. The /stop skill kills the recording process, merges the audio channels, runs Whisper on the merged file, and saves the transcript to SQLite. For the MCP server, use mcp Python package to expose a search_meetings(query) tool. Register it in your global ~/.claude/mcp.json. Speaker diarization adds pyannote-audio (requires HuggingFace token).

Build a Python script with two data sources: (1) NewsAPI for keyword-based news monitoring (search for "CISO hire," "cloud migration," "AI launch" + company names), and (2) a requests-based scraper that hits career pages and counts security-related job postings. Store results in SQLite with a composite key of (company, signal_type, date) for deduplication. Define your ICP company list in a JSON config file. Add confidence scoring based on source reliability and keyword density. Schedule with cron or Windows Task Scheduler.

Create 4 skill .md files in ~/.claude/skills/. /close-out should read the current session's tool calls and file changes, categorize work as completed/in-progress/discussed, update ROADMAP.md by moving completed items to COMPLETED.md, and write a resume.md briefing for the next session. /remember needs a SQLite or PostgreSQL database of past sessions (use the claude-chat-db MCP pattern) and a search function that queries by keyword. /remindme reads the project CLAUDE.md plus recent git log and summarizes current state. /skill-builder is a conversational prompt that asks for the skill's trigger, behavior, and output format, then generates the .md file.

Create a Google Cloud project, enable the Gmail API, and create OAuth 2.0 credentials (Desktop App type). Write a Python MCP server using the mcp package that exposes tools like gmail_search_messages, gmail_read_message, gmail_send_email, and gmail_create_draft. Use google-auth-oauthlib for the OAuth flow and google-api-python-client for Gmail API calls. Store credentials locally in the project directory. Register in ~/.claude/mcp.json as a global server with "type": "stdio" and the Python command to launch it.

Write a skill that: (1) reads all .md files in the target directory, (2) regex-scans for placeholder patterns like YOUR_*, {{*}}, REPLACE_*, (3) checks for Python/Node dependencies in any scripts, (4) detects whether files are agents (go in ~/.claude/agents/) or skills (go in ~/.claude/skills/), and (5) generates a CLAUDE.md that on first message asks the user for each placeholder value, copies files to the right locations, and verifies the install. Template the CLAUDE.md output with the discovered metadata.

Create 3 agent .md files. The planner agent should ask for a hypothesis, success criteria, methodology, and compute budget, then write an experiment-plan.md in a Handoffs directory. The builder agent reads that plan, writes a Python script, does a dry-run, and launches the real run in the background. The reviewer agent reads the output, runs statistical analysis, and produces a structured recommendation. Chain them with a CLAUDE.md that defines the handoff sequence: planner writes plan, builder reads plan and writes results, reviewer reads results.

Create an agent that: (1) accepts a topic and target length, (2) searches for recent articles and data on the topic, (3) generates an outline with a strong opening hook, (4) writes each section with a conversational tone, (5) runs a self-edit pass for redundancy and filler, and (6) outputs clean HTML with inline styles for email compatibility. Store 3-5 past newsletters in a voice-samples/ directory and reference them in the agent prompt for tone calibration.

Write a skill that asks a series of preference questions (dress vs. sport, automatic vs. quartz, budget range, wrist circumference), stores answers in a JSON profile, then uses WebSearch to find matching watches from Chrono24, Hodinkee, and brand sites. Output a structured comparison table with specs, price, and purchase links. Save the profile so future queries build on past preferences.

Two skill .md files. The cold-call skill accepts product, persona, and pain point, then outputs: (1) 15-second opener, (2) 3 qualifying questions, (3) value prop bridge, (4) objection matrix with 5 pushbacks and recovery lines, (5) concrete next-step close. Add a mock call mode that generates realistic prospect responses. The SMYKM skill accepts sales copy and runs three audit passes: scan for phrases that explain context the reader already has, find trigger events stated rather than embedded in questions, and flag generic personalization. For each flag, output the original, the problem, and a rewrite.

Four skill files in ~/.claude/skills/. /nightshift reads every project's ROADMAP.md and CLAUDE.md, identifies actionable items, and spawns background agents (one per task) on feature branches using Claude Code's Agent tool with isolation: "worktree". /credit-burn checks automation health (are scheduled tasks running? are webhooks healthy?), surfaces recurring high-value tasks, and suggests what to run now. /credit-scheduler wraps any recurring task with time-awareness: check current time against off-peak windows (before 8am, after 2pm ET weekdays, all day weekends) and defer heavy work to 2x allowance periods. /context-reset reads the current session's file changes and conversation, writes a resume.md briefing, and suggests which files to re-read on the next session.