Coding with AI Agents

AI workshop · 2026

Coding with AI Agents

One workflow for Claude Code and Copilot

Ivan Dachev · October 2026

What changed

New models: less prompting, not more

medium

Start at medium effort. Go higher only with a measured gain

hours

Long autonomous runs, given a clear finish line

500K / 1M

Max context, depending on the model. Thinking built in

Delete from prompts:

think carefully

double-check your answer

use a subagent to verify

Source: Anthropic prompting guides for current models (platform.claude.com)

Try this first

Three habits for every session

01

Whole task. Clear finish line.

“Done means: every endpoint migrated, old client deleted, tests pass.”

02

Turn the effort dial. Skip “think hard”.

Modern models always think. Effort decides how much.

03

Read “needs your decision” first.

Short IDs on every question, so you answer “Q2 yes, OP1 no”.

Source: claude.dev playbooks · Best practices for Claude Code

The workflow

Ten steps, one loop

1

01

Context files

short, with a stop rule

2

02

Skills, MCP, plugins

on demand, few tools

3

03

Jira and GitHub

ticket to PR, least privilege

4

04

Spec and review

SPEC.md with a finish line

5

05

Agents and advisor

parallel, observer watches

6

06

Build

effort dial, tests first

7

07

Verify

green ≠ verified, layered review

8

08

Learn

repeated fix becomes a rule

9

09

Marketplace

owner and version per plugin

10

10

Copilot parity

AGENTS.md as the shared base

Step 8 feeds step 1

Step 01 · Context files

Context files: short and sharp

Keep

Commands Claude can’t guess

Rules that differ from defaults

Architecture rules and “never do”

Test runner and repo etiquette

Cut

What the code already shows

Long docs: link them with @import

“Write clean code”

“Think carefully”

Test each line: would removing it cause a mistake? No: cut it.

Source: Best practices for Claude Code (code.claude.com/docs)

Step 01 · The one rule to add

Tell it when to stop

# CLAUDE.md · stop rule

When a step doesn’t need my input, keep going.

Status notes go with your next action.

Stop and ask only when you can’t go on without me,

or before anything destructive: deleting data,

force-pushing, touching files outside this repo.

Report: Needs your decision (Q1, Q2…) · FYI

Progress as % of plan. Commands alone in code blocks.

Keep going

When nothing needs you

Stop and ask

Blocked, or destructive

Unattended

Prod, deploy, push: your explicit yes

Source: claude.dev playbook · Anthropic prompting guides, unattended agentic runs

Step 02 · Skills, MCP, plugins

Context is the scarce resource

CLAUDE.md

Every session

Skills

Loaded when relevant

Hooks

Every time, enforced

Subagents

Their own context

MCP

Tool defs use context

Prefer CLI tools (gh, gcloud) · Vet every plugin like a package

Source: Best practices for Claude Code · Extend Claude Code (code.claude.com/docs)

Step 03 · Jira and GitHub

From ticket to pull request

Jira ticket

SPEC.md

Branch

PR + AI review

Ticket updated

Agent may

Read tickets and code, comment, draft the PR

CLI first (gh). MCP where there is no CLI.

Human approves

Merge, status changes, anything destructive

Least-privilege tokens. Read-only by default.

Source: Best practices for Claude Code, CLI tools and MCP · claude.dev, How the Claude Code team does code review

Step 04 · Planning and specification

Spec before code

1 · Explore

2 · Plan

3 · Implement

4 · Commit

Plan mode: reads, asks, changes nothing

1 · Claude interviews you

2 · It writes SPEC.md

3 · A fresh session builds it

One-sentence diff? Skip the plan.

SPEC.md

Goal and finish line: “Done means…”

Files, interfaces, layers involved

Out of scope

Tests and an end-to-end check

When to stop and ask

Source: Best practices for Claude Code, Explore, plan, code and Let Claude interview you

Step 05 · Agents, subagents, advisor

Delegate big work, check evidence

Advisor

Second opinion at decisions

/advisor opus

Fast main model + stronger advisor · same-tier advisor for high stakes

Observer

A second agent watches

Catches silent decisions and false success

Unasked behavior change = a question

Source: Anthropic prompting guides, subagent spawning · claude.dev playbook · Claude Code docs, advisor tool

Steps 05–06 · Model settings

Effort dial, not “think hard”

Level

Use it for

low

Quick edits, first review pass

medium

Default. Most coding work

high

Hard, ambiguous changes

xhigh / max

Only with a measured gain

Less thinking? Lower effort first.

Model by task: strongest for hard code and review, cheaper for tests and scripts

Source: Anthropic prompting guides, Calibrate effort and Efficiency at lower effort

Step 06 · Implementation

Give it a check it can run

Tests first

No test edits without approval

TASKS.md

Survives compaction. No TODO without evidence

Hooks and /goal

Per change: lint and touched tests. Full suite per batch

Fresh session

After two misses or a full context. Point to docs

Source: Best practices for Claude Code, Give Claude a way to verify its work · claude.dev playbook

Step 07 · Verification and code review

Review in layers, not one gate

2×

More merged pull requests

−40%

Share of PRs reverted

Claude Code team, June to September 2026

Source: How the Claude Code team does code review (claude.dev, Oct 2026)

Step 07 · Human review

People where the risk is

0

humans, once checks are green

Low risk

Docs, copy, additive changes behind a flag

1

human approval

Medium risk

Ordinary product changes

2

human approvals

High risk

The riskiest paths you name: auth, billing, CI

The change sets the tier, not the author

Whoever prompted the change owns it

A bot approval never replaces a human one

Source: How the Claude Code team does code review (claude.dev, Oct 2026)

Step 08 · Self-learning

Fix it once, at the source

Repeated correction

“I said this twice already”

Fix at the source

Trap · why · how to apply

Prune and test

Did the behavior change?

Put it in

CLAUDE.md

a skill

a hook

a lint rule

REVIEW.md

Incidents and reverts become evals

/doctor proposes CLAUDE.md cuts

Source: Best practices for Claude Code · How the Claude Code team does code review (claude.dev)

Step 09 · Private plugins and marketplace

Package what works, share it

acme-workflow/

Skills

Hooks

Subagents

MCP servers

Commands

Mods

Marketplace

git repo + marketplace.json

Team

/plugin install x@acme

Owner + version

claude plugin validate · test

Review it like a package

Source: Claude Code docs, plugins · claude.dev, Getting started with Claude Code mods (Oct 2026)

Step 10 · Copilot parity

One base for every tool

AGENTS.md

Shared rules and commands

Claude Code

CLAUDE.md: @AGENTS.md

Copilot

copilot-instructions.md

Modules

AGENTS.md per package

Same for both: spec, finish line, TDD, layered review

Claude Code extras: advisor, subagents, hooks, /goal, plugins

Source: Claude Code docs, CLAUDE.md imports · claude.dev, Claude Code mods (AGENTS.md support)

Take-home

Your checklist

Ask

Whole task with a finish line

No “think hard” lines

Model and effort set up front

Long runs

Stop rule in CLAUDE.md

Context and permissions set before you leave

Parallel by default, full suite per batch

Check

A check it can run

An observer agent beside the main one

Green ≠ verified

Team

Repeated asks become written rules

No TODO without evidence

AGENTS.md as the shared base

Based on the claude.dev checklist and lessons from our own sessions

In one line

Hand over the whole task. Give it a way to check. Own what ships.

Questions?