Development · Phase 04 of 05

Start with the model,
the prompt, and the context.

The attached workshop deck is an AI coding primer, not a validation workflow. Its first move is to give developers a shared operating model: when to use Copilot vs Claude Code, how model choice affects cost and output quality, and why prompt scope plus context hygiene determine whether an agent helps or thrashes.

GitHub Copilot Claude Code Claude Sonnet 4.6 Claude Opus 4.6 GPT-5.4

Copilot and Claude Code solve different shapes of development work.

The deck positions them as complementary, not interchangeable. Copilot starts inside the IDE and shines on in-flow coding. Claude Code starts with the whole repository in scope and is built for terminal-first, multi-file execution.

Tool 01

GitHub Copilot lives where developers already work

VS Code and JetBrains workflows map cleanly to Copilot’s Ask, Edit, Agent, and Plan modes. It is the natural fit for inline edits, scoped file changes, and backlog items that can be assigned and returned as PRs.

Best fit: file-level to multi-file work when you want strong IDE ergonomics and quick iteration.
Tool 02

Claude Code starts from repo awareness, shell access, and planning

The CLI workflow assumes the agent can inspect the codebase, run commands, execute tests, and work across the repository from the start. That makes it stronger for coordinated edits, refactors, and tooling-heavy tasks.

Best fit: multi-file changes, terminal workflows, repo-wide reasoning, and tool-driven execution.
Mode Match

Use the lightest interaction mode that can finish the job

Ask mode is for explanation. Edit mode is for controlled diffs. Agent mode is for autonomous search, edits, and terminal loops. Plan mode exists to make multi-file work reviewable before a single line changes.

The workshop repeatedly pushes one behavior: review the blueprint before approving autonomous edits.
Decision Rule

Tool choice is really scope choice

If the task is local, stay local. If the task needs shell commands, repo search, tests, or orchestration, step up to a true coding agent. The deck frames tool selection as matching workflow shape, not brand preference.

Pick the interface that gives enough autonomy without paying for unnecessary complexity.
GitHub Copilot Claude Code
Lives inVS Code, JetBrains, Eclipse, XcodeTerminal (CLI) + IDE extensions
InteractionAsk, Edit, Agent modes + Chat UIConversational prompts + planning
ScopeFile → multi-file → full repo (agent)Multi-file, full codebase from start
ModelsGPT-5.4, Claude Sonnet 4.6, GeminiSonnet 4.6 (default) / Opus 4.6
Runs code?Yes — agent mode runs terminalYes — shell, tests, git, MCP tools
Plans ahead?Yes — Plan mode (review before code)Yes — explicit planning mode
ContextModel-dependent (128K–1M)1M tokens (Opus 4.6 / Sonnet 4.6)
Pricing$10–$39/mo subscriptionUsage-based (API tokens)

Model selection is a cost-quality tradeoff, not a branding debate.

The workshop gives a clear recommendation: use Sonnet 4.6 as the daily driver, reserve Opus 4.6 for deeper reasoning and coordinated agent work, and treat GPT-5.4 as a strong Copilot-side option when you want premium request quality inside the IDE.

Daily Driver

Claude Sonnet 4.6

1M-token context and strong coding performance at a much lower price point than Opus. The deck explicitly recommends it for most day-to-day development sessions.

Use for: routine coding, refactors, planning, and most agent runs.
Heavy Lift

Claude Opus 4.6

Also 1M-token context, but priced for the hard cases: deeper reasoning, complex refactors, and multi-agent coordination where mistakes are expensive.

Use for: hard architectural work, complex reviews, and agent teams.
IDE Premium

GPT-5.4 in Copilot

The deck calls out GPT-5.4 as a strong premium model in Copilot for computer-use style tasks, tool search, and higher-end assisted coding inside the editor flow.

Use for: Copilot-heavy teams that want a strong premium model without leaving the IDE.
1M Tokens for Sonnet / Opus

The workshop treats long context windows as a practical advantage only when the team manages that context deliberately.

272K Tokens for GPT-5.4 in Copilot

Still large enough for serious work, but not a license to keep every file, tab, and transcript open forever.

1/5 Sonnet vs Opus cost guidance

The deck’s economic point is blunt: save Opus for the sessions where extra reasoning actually pays for itself.

Anthropic Claude (Claude Code)
Model Context Max Output Pricing (in/out) Best For
Claude Opus 4.61M tokens128K output$15 / $75 per MTokFrontier reasoning, agent teams, complex refactors
Claude Sonnet 4.61M tokens64K output$3 / $15 per MTokDefault in Claude Code. 98% of Opus coding perf.
Claude Haiku 4.5200K tokens8K output$0.80 / $4 per MTokFast subagents, exploration, lightweight tasks
GitHub Copilot (multi-model)
Model Context Pricing Notes
GPT-5.4272K tokens$10 / $30 per MTokDefault for premium requests. Computer use, tool search.
Claude Sonnet 4.61M tokens$3 / $15 per MTokAvailable in Copilot Pro+. Strong agentic coding.
Gemini 2.5 Pro1M tokensVariesAlternative for large-context tasks.
Key insight: Sonnet 4.6 scores 79.6% on SWE-bench (vs Opus 80.8%) at 1/5 the cost. Use Sonnet as your daily driver; reserve Opus for deep reasoning, agent teams, and complex refactors.

The deck’s prompt framework is simple: give the model the right shape.

CRISP is the workshop’s shared structure for prompting coding agents. It keeps requests concrete, bounded, and aligned to the codebase instead of drifting into vague “fix this” conversations.

C

Context

Name the stack, framework, constraints, and the relevant part of the codebase. “TypeScript/React 19 monorepo with Vitest” is usable context. “Fix the bug” is not.

R

Role

Tell the model how to think about the task: senior backend engineer, security reviewer, test engineer, release owner. The role shapes judgment and review criteria.

I

Intent

State the desired outcome explicitly: refactor to async/await, fix the 401 in middleware, generate unit tests, explain the architecture. Intent keeps the agent from inventing its own target.

S

Scope

Tell the agent what to touch and what not to touch. The deck emphasizes negative constraints because they prevent “helpful” surprise edits across unrelated files.

P

Pattern

Reference the style, conventions, or examples to follow. If the repo already has an error-handling pattern, naming scheme, or test style, point the model at it directly.

Prompt takeaway

The workshop’s “bad vs good” example is making a narrow point: better prompting is mostly better scoping. A strong coding prompt describes the codebase, the goal, the boundaries, and the pattern to imitate before the agent writes anything.

Weak prompt
Fix the bug.
Why it fails
  • No context — which bug?
  • No scope — what files?
  • No role or pattern constraints
Strong CRISP prompt
[Context] TypeScript/Express API, Node 22, vitest.
[Role] Act as a senior backend engineer.
[Intent] Fix the 401 error in src/auth/middleware.ts.
[Scope] Only middleware.ts. Don’t touch test files.
[Pattern] Use the same try/catch style as utils/errors.ts.
Tips & Tricks

If the task prompt keeps repeating repository rules, move those rules into CLAUDE.md or a skill file. Keep the live prompt focused on the current job, not boilerplate the model has already been taught.

A large context window is useful only if the session stays focused.

The workshop is explicit about what fills the window: system prompts, instruction files, opened files, chat history, tool results, and the new request itself. Bigger windows help, but noisy sessions still degrade output.

Wasteful Session

Context gets consumed by noise

  • Irrelevant source files stay open even after the task changes.
  • Long chat history accumulates without compaction or reset.
  • Raw tool output gets pasted back in full instead of summarized.
Clean Session

Context stays aligned to the current problem

  • Only the files and snippets relevant to the task remain in play.
  • Older history is compacted or replaced with summaries.
  • The developer keeps refreshing scope instead of assuming the model “still remembers the right thing.”
1

Open only what matters

Large files, unrelated tabs, and stale examples cost tokens and attention. The deck treats focused file selection as a quality control mechanism, not just an optimization.

2

Compact or reset between tasks

Claude Code’s compaction feature exists because every prior turn otherwise keeps shaping the next answer. New task, new summary, or new session.

3

Summarize results instead of dumping logs

The model needs the key error lines and stack trace, not a two-thousand-line terminal transcript. Distillation preserves budget and improves signal.

Page takeaway
The workshop’s first lesson is operational discipline: better context beats longer context. Tool quality starts with model choice, prompt structure, and aggressively scoped sessions.

Where this fits in the full pipeline.

Phase 03 — Design + Build. Development tooling and agent workflows live in the implementation phase of the SDLC.

Phase 03 — Design + Build
AI-Powered SDLC pipeline diagram with Phase 03 Design + Build highlighted.

Scroll right to see full pipeline →