Start your learning journeyAI Playbook
Log inStart learning
Role playbook

Software developers

The daily loop with coding agents: plan first, make tests the contract, keep diffs small, run tasks in parallel worktrees, get an agent review, and let hooks and routines handle formatting, safety and chores.

Checked 13 Oct 202622 min read17 prompts1 skill2 hooks2 routines

For Software engineers who write, review and ship code every day

Ship small, tested changes while the agent types and you decide what merges

  • A plan, test, implement, verify loop that turns a ticket into a reviewable pull request
  • Tests written first, so the agent cannot call a change done until they pass
  • Two or three tasks running at once in separate worktrees, each with its own small pull request
  • An agent review on every branch before a teammate spends time on it
  • Dependency bumps and flaky-test reports drafted overnight for you to review
Works withClaude CodeOpenAI CodexCursorGitHub CopilotGemini CLI
On this page
  1. Where AI helps in this job
  2. Start in 15 minutes
  3. The tools developers use
  4. Members only: Workflow: from ticket to pull request
  5. Members only: Workflow: parallel tasks in worktrees
  6. Members only: Workflow: an agent review before a human sees it
  7. Members only: Workflow: tickets and docs over MCP
  8. Members only: A skill for developers
  9. Members only: Guardrails and hooks
  10. Members only: Routines that run without you
  11. Members only: Prompt ladder for developers
  12. Members only: Measure it
  13. Members only: Risks and what not to automate

A coding agent reads your repository, runs your tests and edits files. That moves your job from typing to deciding: what to build, what correct means, and whether a diff is safe to merge. This page gives you the loop, the prompts and the guardrails for that job in Claude Code, with the Codex, Cursor, Copilot and Gemini CLI equivalents where they exist.

Where AI helps in this job#

TaskWhat AI does wellWhat you still own
Reading unfamiliar codeTraces a request through files and explains each step with file and line referencesChecking the explanation against what the code does at runtime
Planning a changeLists the files to touch, the order, edge cases and open questionsThe approach, the scope and what you will not do
Writing tests firstTurns acceptance criteria into failing tests in your framework's styleWhat correct means, and tests that assert behavior instead of implementation
ImplementingWrites the smallest change that makes the tests pass, following nearby patternsDesign, naming, and a diff small enough to review
DebuggingReproduces from a stack trace, adds logging, bisects, proposes a root causeConfirming the root cause instead of accepting a symptom patch
ReviewA first pass for logic errors, missing tests and security smellsThe approval, and anything about product or architecture
Refactors and migrationsMechanical edits across dozens of files, in parallelSequencing, rollout and backwards compatibility
Dependency updatesReads changelogs, bumps versions, fixes broken calls, runs the suiteMerging, and the risk call on major versions
Pull request descriptionsSummarizes the diff and pastes the test evidenceAccuracy, and what the reviewer needs to know

Agents are weakest where they cannot check their own work. Without a test, a build or a type check to run, a change is a confident guess and you become the test suite. They also invent APIs and flags from older library versions, drift on large vague tasks, and do worse on performance, concurrency and security-sensitive code than on routine code. Keep tasks small, give every task a check, and read every diff.

Start in 15 minutes#

These three prompts work in any agent with access to your repository (Claude Code, Codex, Cursor, Copilot agent mode, Gemini CLI) and as pasted text in a chat app.

Prompt

Explain a code path before you touch it

Level 1Any assistant
Read [FILE OR FOLDER]. Explain what happens from [ENTRY POINT, FOR EXAMPLE POST /orders] to [OUTCOME, FOR EXAMPLE A NEW ROW IN THE orders TABLE]. List each function in order with its file and line, what it reads and writes, and where it can fail. Do not change any files.

Good output is a numbered call chain with file and line references and the failure points marked. Open two or three of the cited lines and confirm they say what the agent claims.

Prompt

Plan before code

Level 2Any assistant
I need to [CHANGE, FOR EXAMPLE ADD A RETRY TO THE PAYMENT WEBHOOK HANDLER]. Do not write code yet. Propose a plan:
1. The files you will change and why.
2. The tests you will add first, with their names and what each one asserts.
3. Edge cases and failure modes.
4. Anything you are unsure about, as questions for me.
Keep the change under [200] lines if you can. Wait for my approval before you edit anything.

Good output names real files and real test cases and asks at least one question. If the plan touches files you did not expect, ask why before you approve it.

Prompt

Tests as the contract

Level 2Any assistant
Write failing tests for this behavior: [ACCEPTANCE CRITERIA]. Use the test framework and style already in [TEST FOLDER]. Run the tests and show me the output that proves they fail for the right reason: an assertion about the missing behavior, not an import error or a typo. Do not write the implementation yet.

Good output is red tests whose failure messages describe the missing behavior. Run them once yourself. A test that fails for the wrong reason will pass for the wrong reason too.

The tools developers use#

  • Claude Code runs in the terminal, IDEs, the desktop app and the browser. It has plan mode, skills, hooks, subagents, claude --worktree for parallel sessions, routines for scheduled work, and /code-review and /security-review. It reads CLAUDE.md, and reads AGENTS.md when a repository has no CLAUDE.md.
  • OpenAI Codex runs as a CLI, an IDE extension, cloud tasks and inside the ChatGPT desktop app. It reads AGENTS.md, loads skills from .agents/skills, runs hooks from .codex/hooks.json, has /plan and /review, and reviews pull requests when you comment @codex review.
  • Cursor is an editor built around an agent: Plan Mode (Shift+Tab), rules in .cursor/rules or AGENTS.md, hooks in .cursor/hooks.json, parallel agents in worktrees, Cloud Agents, and Bugbot for pull request review.
  • GitHub Copilot has agent mode in VS Code and JetBrains, Copilot CLI, the Copilot cloud agent (assign it an issue and it opens a pull request) and Copilot code review. It reads AGENTS.md, .github/copilot-instructions.md and skills in .github/skills.
  • Gemini CLI is Google's open source terminal agent, with plan mode, skills, hooks and MCP. It reads GEMINI.md; add AGENTS.md to its context.fileName setting to share one file.

Write conventions once: put them in AGENTS.md, and put skills in .agents/skills/, which Codex, Cursor, Copilot and Gemini CLI read. Claude Code reads .claude/skills/, so symlink each skill folder there.

Workflow: from ticket to pull request

Run this for any feature or bug that fits in a day. You spend your time on the plan and the review; the agent does the typing, the test scaffolding and the write-up. You need the ticket, a repository whose AGENTS.md or CLAUDE.md names the ...

Members only below this line

Get the full Software developers playbook free

What you get on this page:

  • 4 step-by-step workflows
  • 14 prompts
  • 1 skill
  • 2 hooks
  • 2 routines

Free. You also get Stratiflux Weekly: one email a week, unsubscribe anytime.

Also in the members part

  1. Workflow: parallel tasks in worktrees
  2. Workflow: an agent review before a human sees it
  3. Workflow: tickets and docs over MCP
  4. A skill for developers
  5. Guardrails and hooks
  6. Routines that run without you
  7. Prompt ladder for developers
  8. Measure it
  9. Risks and what not to automate

Sources

Checked against these on 13 Oct 2026.

  1. Best practices for Claude Code (opens in a new tab)code.claude.com
  2. Claude Code hooks reference (opens in a new tab)code.claude.com
  3. Automate actions with hooks (Claude Code) (opens in a new tab)code.claude.com
  4. Extend Claude with skills (opens in a new tab)code.claude.com
  5. Run parallel sessions with worktrees (opens in a new tab)code.claude.com
  6. Code Review (Claude Code) (opens in a new tab)code.claude.com
  7. Automate work with routines (opens in a new tab)code.claude.com
  8. Claude Code GitHub Actions (opens in a new tab)code.claude.com
  9. Connect Claude Code to tools via MCP (opens in a new tab)code.claude.com
  10. Codex: custom instructions with AGENTS.md (opens in a new tab)learn.chatgpt.com
  11. Codex: build skills (opens in a new tab)learn.chatgpt.com
  12. Codex code review in GitHub (opens in a new tab)learn.chatgpt.com
  13. Cursor hooks (opens in a new tab)cursor.com
  14. Cursor third-party hooks (Claude Code compatibility) (opens in a new tab)cursor.com
  15. Cursor Bugbot (opens in a new tab)cursor.com
  16. About agent skills (GitHub Copilot) (opens in a new tab)docs.github.com
  17. Using GitHub Copilot code review (opens in a new tab)docs.github.com
  18. Gemini CLI skills (opens in a new tab)geminicli.com
  19. Linear MCP server (opens in a new tab)linear.app
  20. Atlassian MCP server: getting started (opens in a new tab)support.atlassian.com
  21. METR: measuring the impact of early-2025 AI on experienced open-source developer productivity (opens in a new tab)metr.org
  22. METR: uplift study update (February 2026) (opens in a new tab)metr.org

Related

GuidePrompting, from easy to hardHow the 17 prompts on this page workGuideSkills: teach an agent a procedure once, reuse it everywhereHow the skill on this page worksGuideHooks: guardrails that run every timeHow the 2 hooks on this page workGuideRoutines: scheduled agents that open pull requestsHow the 2 routines on this page work