Skills — prompt chains, task prioritization, and cost optimization with Andy and Alex
Thu, 09 Jul 26
What Skills Are and Why They Matter
- A skill turns a good prompt into a reliable workflow
- Best for tasks done repeatedly where quality matters every time
- Not ideal for one-off tasks or processes that change constantly
- Full skill system includes: skill.md, sub-agents, assets, references, and evals
- Evals act as a scoring framework: they check output against the voice guide, the structure guide, etc.
Prompt to Prompt Chain to Skill
- Start with a single prompt; escalate only when it’s insufficient
- Prompt chain: multiple prompts linked in sequence
- Each step can go deeper
- A human review loop can be inserted between steps
- Once the chain is dialed in, convert to a skill (skill.md in markdown)
- After building skills, bundle related ones into plugins to stay organized
Task Prioritization Demo (Andy’s Hot Seat)
- Framework recommended: Navy SEALs’ CEE model
- Critical: urgent, must act now (flat tire)
- Essential: needed soon (low gas)
- Enhancing: nice to have (new paint job)
- Workflow: dump task list into AI, apply CEE framework, output in preferred format
- Demo ran live using a “C Task Prioritizer” skill in Claude’s Cowork mode
- Cowork runs sub-agents and performs better than standard chat for multi-step skills
- Output exported to Excel/TSV for potential import into task management tools
- Trello is better suited for teams; personal task management works well in Notion or Obsidian
Notion as an AI-Connected Knowledge Base
- Notion connects natively to Claude Desktop via built-in connector
- Claude can build and update Notion databases via plain-English commands
- Demo: created a personal to-do database, added tasks by voice, and queried bills by name
- Session recaps saved directly to Notion databases
- Daily task queries: “show only tasks due today”
- Status updates by voice: “Mark that task as done”
- Notion subscription: ~$12/month; recommended for daily AI-connected use
- Downside: requires files marked offline for airplane/offline access
Security: Skills, Keys, and Connectors
- Do not import skills from unknown sources: 17-20% of community skills are estimated to contain malicious code
- Avoid skills that take over browser sessions or store passwords in Markdown files
- Never put API keys or passwords in skill files, dot-env files, or config.json
- Claude’s config.json is a known target; anyone who accesses it gets all MCP keys
- Recommended approach: use a password manager (e.g., 1Password) to reference keys at runtime
- Keys are never stored in plaintext on the machine
- For connectors and integrations: use marketplace-vetted connectors only, not custom browser-takeover scripts
Claude Subscription, Costs, and Scheduling
- $20/month: fine for occasional use, limited to Sonnet-class models
- $100/month: recommended for daily heavy use; can run all day on single sessions
- Cost strategy: use higher-tier models for strategy, lower-tier for execution
- OpenRouter now offers intelligent routing: auto-selects a model by task complexity
- Claude Desktop scheduled tasks: runs skills on a timer (daily, weekly, etc.)
- Computer must be on and app open for scheduled tasks to fire
- Six built-in templates: inbox triage, meeting prep, content ideas, etc.
- Recommended resource: “The AI Daily Brief” podcast/YouTube by Nathan (covers news, trends, job market impact)
Next Steps
- Post Alex's skills collaboration idea to the School community
- Explore building a shared skill that routes tasks to the lowest-cost model capable of handling them.
- Think about September session topics and post suggestions to the Skool
- Agents are likely not the focus; the group is looking for an intermediate topic between skills and full agent workflows.