Notebook

Appearance

Dashboard AI Week ​

What are Skills, Agents, Prompts, Instructions and how to use them? ​

  • A prompt is a one-time chat message.
  • An instruction is a permanent rule for a project.
  • A skill is a tool the AI loads only when needed.

AGENTS.md ​

  • inspired by CLAUDE.md from Anthropic
  • Instead of having md files for each vendor
  • .github/copilot-instructions.md used for github copilot as your PR code reviewer
  • /init
  • AGENTS.md sections
    • Architecture overview
    • Dos and Donts
    • Comments
  • Better to have AGENTS.md for each project module as well

Custom Agents ​

  • agents folder
  • They are like specialized colleagues, can be given some personality
  • Orchestrator Agent, Planner Agent and Coder Agent
    • cleaner context memory
  • .github/agents, uses tools
  • github/awesome-copilot
  • SubAgents vs Agent Teams
    • Subagent reports back to orchestrator
    • Agent team communicate with other agents

image

Hooks ​

  • hooks folder
  • Hooks run before agents are loaded
  • They have lifecycle methods like sessionStart, sessionEnd, userPromptSubmitted etc
  • Try some hooks from awesome-copilot
  • example : secret scanner and governance-audit
  • Show agent logs in vscode, ask a questions in chat -> setting (gear icon) in chat -> agent logs
    • Loaded instructions -> skills -> hooks -> agents

Instructions ​

  • instructions folder
  • has applyTo liquid attribute
  • mention in main AGENTS.md when to refer which instruction
  • A global rule file that stays active across the whole project. example use java 11 and gradle instead of maven

Prompts ​

  • prompts folder
  • use /explain-code for .github/prompts/explain-code.prompt.md
  • example: brief overview, step-by-step breakdown, key concepts and terms, common use cases

Skills ​

  • skills folder
  • Folder based capability packages with a required SKILL.md that are loaded by an Agent on-demand
  • .github/skills/breakdown-plan/SKILLS.md will create some docs
  • invokable
  • Persistent, modular files loaded only when triggered.

Workflow ​

  • workflow folder
  • Copilot coding agent works anonymously in a Github actions-powered env to complete development tasks and creates pull requests with the result.
  • daily-issues-report

AI-DLC ​

  • AI driven development lifecycle (AI-DLC)
    • AI Creates Plan -> Human Verify Plan -> AI refines Plan -> AI executes plan -> Human verify outcome
  • Approach 1 : AI Managed
    • AI autonomously builds and maintains software with leasat ot zero human involvement
  • Approach 2 : AI Assisted
    • Developers still perform the intellectual heavy lifting and apply AI in narrow tasks
  • Steps
    • Inception (Mob Elaboration)
      • Build Context on existing codes
      • Elaborate intent with user stories.
      • Plan with units of work.
    • Construction (Mob Construction)
      • Domain model (component model)
      • Add architectural components
      • genrate code and test
    • Operation (CICD)
      • Deploy in Prod with IaaC
      • Manage incidents - AI first, Fast feedback to dev
  • Avoid vibe coding
    • Can you debug every line of code?
    • Do you understand the code end-to-end?
  • Start with small simple tasks
    • No complex task that may succeed
    • Start by vreating simple tasks that always succeed
    • Continue breaking into the specific functions, files or flows
  • Clear context, keep context focused
  • Ask AI to mimic existing code examples rather than complex instructions
  • semantics ratio in the tokens
    • good : refactor using builder pattern
    • bad : ability to build complex object blah blah
  • Beware fo Model's stale training dataset
  • Keep number of MCP tools minimal
    • required MCP uses context window or something else
  • Old measuring metrics don't work anymore like lines of code
  • Is re-writing faster than patching
  • Example : Fast API to support QUERY http method
    • .amazonq folder had AI-DLC rules
    • it outputs aidlc-docs folder based on the instructions
      • rule had inception/reverse-engineering folder
  • 🔗 github aidlc core-workflow

Harnesses in AI ​

  • Why? Reliability
  • An AI harness (also called an agent harness) is the software infrastructure, tools, and rules wrapped around an AI model to make it work reliably in the real world.
  • While a raw Artificial Intelligence model is just a "text-in, text-out" brain, it cannot actually perform tasks on its own. The harness is the system that connects that brain to the outside world, giving it tools, memory, and safety guardrails.
  • Agent = Model + Harness
  • Six Core Layers of an AI Harness
    • The Execution Loop: It keeps the AI running in a continuous loop (Observe ➔ Decide ➔ Act ➔ Verify) until the job is fully done.
    • Tool Management: It gives the AI the ability to click buttons, search the web, read files, and run code.
    • Context Engineering: It filters information so the AI only sees what is relevant, preventing it from getting overwhelmed.
    • Persistence (Memory): It saves the AI’s progress to a database so it does not forget what it was doing if the system restarts.
    • Verification: It automatically checks the AI's work—like running a code test—to catch mistakes before a human sees them.
    • Constraints (Guardrails): It blocks the AI from taking dangerous actions, like deleting important company files.
  • Agent Harness
    • Tool Registry
    • Model
    • Context Management
    • Guardrails
    • Agent Loops
    • Verify
  • Example : Use ChatFPT 3.5 and harness it