Development · Phase 04 of 05

Teach the model,
then let the agent work.

The middle of the workshop deck shifts from prompting into agent behavior. It distinguishes passive knowledge from active autonomy, shows how Ask/Edit/Agent/Plan modes change the risk profile of a task, and outlines how Claude Code sessions should be structured before developers hand over execution.

Ask Mode Edit Mode Agent Mode Plan Mode

Skills are reusable knowledge. Agents are autonomous behavior.

The deck spends real time on this because teams routinely blur the two. A skill file gives the model domain expertise. An agent reasons, calls tools, inspects results, retries, and decides what to do next.

Skill = passive context

Static instructions that teach the codebase

Form
Markdown documents such as CLAUDE.md, SKILL.md, or project instruction files.
Purpose
Capture conventions, file locations, architectural patterns, and domain constraints once.
Behavior
No tools, no loops, no decisions. Skills inform the model but do not act.
Cost profile
Loaded once into context, which makes them cheap compared with iterative agent loops.
Agent = active execution

Reasoning loop with tools and feedback

Form
Claude Code, Copilot Agent, subagents, and other runtimes that can read, edit, search, and run commands.
Purpose
Inspect the repo, choose files, call tools, handle errors, and continue until the task is done.
Behavior
Dynamic and stateful. Tool results land back in context and influence the next action.
Cost profile
More expensive because every step consumes turns, tokens, and sometimes multiple cooperating workers.
Step 1User prompt
sent
Step 2Model reasons →
tool call JSON
Step 3Runtime
executes tool
Step 4Result injected
into context
Step 5Model continues
or loops

Match the mode to the task before you grant autonomy.

Both Copilot and Claude Code support multiple ways of working. The deck’s recommendation is to start with lower-autonomy modes when the task is simple and switch up only when scope, uncertainty, or execution needs justify it.

Mode 01

Ask

Q&A, explanation, architecture understanding, and lightweight exploration. No files change. This is the cheapest place to clarify intent before you risk code churn.

Mode 02

Edit

You choose the file set and the AI proposes diffs inside that boundary. This is the workshop’s controlled middle ground for refactors, renames, and contained changes.

Mode 03

Agent

The model decides what to inspect and modify, runs tools, and loops until it believes the job is complete. Powerful, but only safe when the task and constraints are already clear.

Mode 04

Plan

Think first, then execute. The deck repeatedly treats planning mode as the default starting point for multi-file work because it exposes file impact, edge cases, and risk before edits begin.

Copilot Ask Copilot Edit Copilot Agent Claude --plan
Tips & Tricks

If the change crosses files, tests, or migrations, start in plan mode even when the implementation feels obvious. The extra minute up front is usually cheaper than unwinding a confident wrong diff.

Project guidance belongs in layers, not one bloated system prompt.

One of the strongest pages in the deck is the skill-file ecosystem diagram. It separates always-on instructions from demand-loaded expertise so the model gets the right knowledge at the right time without carrying unnecessary weight every turn.

Always On

CLAUDE.md

Project-wide Claude Code instructions. The deck’s guidance is to keep it tight because it loads every turn and therefore costs tokens every turn.

On Demand

/skills/*/SKILL.md

Domain-specific expertise such as auth, billing, or release patterns. Skills are where detailed conventions belong once they stop being universal.

Agents

.claude/agents/*.md

Custom subagents with specific tools or roles. The workshop positions these as a way to standardize specialized workers rather than keep reteaching the same behavior.

Shared

AGENTS.md

Cross-agent coordination instructions that can keep Claude, Copilot, and other agent runtimes aligned on team standards.

Copilot

.github/copilot-instructions.md

Copilot-specific project guidance, with optional custom agents under .github/agents when the team wants more structured behavior inside GitHub tooling.

Instruction rule

The deck’s practical guidance is simple: system prompts should stay compact, skills should hold depth. That split keeps always-loaded context lean while preserving detailed domain knowledge when an agent actually needs it.

The terminal workflow is agentic, but it still needs checkpoints.

The workshop’s Claude Code section is intentionally operational. It walks through how to start a session, describe a task, force a plan, execute, then review and compact before moving on.

1

Launch

Open Claude Code at the project root so it can read CLAUDE.md, discover available skills, and start with the right repository context.

2

Describe

Use plain English, but keep it specific. Files, errors, constraints, and desired outcomes matter more than polished prose.

3

Plan

Use --plan or explicitly ask for a plan first. Review touched files, edge cases, and what the agent says it will not change.

4

Execute

Only after the plan is approved should the agent read, write, test, and call tools in a loop. The deck treats this as a governed handoff, not blind trust.

5

Review

Check the diff, inspect test results, and compact or reset before the next task so stale assumptions do not leak across sessions.

Parallelism

Subagents are the default scale-up pattern

They work in a hub-and-spoke model: focused workers with separate context windows that report back to a parent session. The deck recommends them for exploration, test runs, and code review.

High-End Scale

Agent teams are for coordinated, peer-to-peer work

The deck positions agent teams as a more expensive option for full-stack features and complex parallel reviews where workers need to coordinate directly instead of reporting only to a parent.

Page takeaway
Plan mode is the safety boundary. Skills sharpen the agent, but planning is what keeps autonomy reviewable and compatible with normal engineering controls.