Agentic Software Engineering (ASE): Agentic AI Coding Meets Software Engineering

Agentic Software Engineering

Boost your daily productivity while keeping yourself in the driver's seat!

Fuse Agentic AI Coding with traditional Software Engineering!

Let the agent handle the chores while you stay in control of the craft!

Various operation modes provided — you choose the best for the job!

If you like ASE, please support it by giving a star! Many thanks.

Agentic Software Engineering (ASE) is the opinionated Open Source toolkit of prolific author Dr. Ralf S. Engelschall, for fusing the concept of Agentic AI Coding and traditional Software Engineering. ASE ships as a plugin for the software development tool Anthropic Claude Code CLI — and with reduced support also for the alternative tools GitHub Copilot CLI and OpenAI Codex CLI. It is comprised of agent hooks, parametrizable agent skills, an underlying Model-Context-Protocol (MCP) service and a companion Command-Line Interface (CLI).

Dr. Ralf S. Engelschall in a studio interview on ASE

ASE incorporates reasonable methodology and automation aspects to support the daily, recurring tasks of a Software Developer and Software Architect. The comprehensive functionality spans from brainstorming ideas, searching the Web, asking foreign LLMs, discovering components, through evaluating alternatives, challenging statements, analyzing root-causes, managing tasks, grilling task plans, all the way to analyzing, fixing, refactoring, and crafting code, reviewing changesets, and many more.

Claude Code:

❯

Highlights ↓

Setup ↓

Usage ↓

Workflows ↓

Design ↓

Architecture ↓

Compatibility ↓

-->

⌄

Token and Time Optimization

Switch persona to cut down tokens and reading time

To reduce your overall output tokens and necessary reading time, switch the communication style of the LLM across five built-in intensity levels: from the decorative, eloquent, and explaining writer, the concise, factual, and accurate engineer (default), the layered, pyramid-structured journalist, over the brief, factual, and abbreviating telegrapher, down to the terse, rough, and stuttering caveman.

Claude Code

❯/ase-meta-persona engineer

Parametrized Brainstorming

Diverge, cluster, and shortlist new ideas

Turn a fuzzy topic into a focused shortlist: first diverge into a broad set of ideas, then converge by clustering and scoring them, and finally distill a ranked shortlist with a recommended direction to pursue.

Claude Code

❯/ase-meta-brainstorm --max-clarify=0 --min-ideas=20 clever and useful AI skills for software engineering

Root-Cause Analysis

Trace root-cause with Five-Whys method

Drill past symptoms to the real cause: the Five-Whys method iteratively asks "why?" to trace a problem or claim back to its root, optionally branching into multiple causal paths to widen the investigation.

Claude Code

❯/ase-meta-why --width=2 is the Decibel (dB) unit a logarithmic one?

Component Discovery

Discover the right components for your technology stack

Get methodical support in finding suitable libraries and frameworks to establish or extend your technology stack — grounded, not guessed, by taking into account the downloads per time, age of component, last update time, and GitHub stars.

Claude Code

❯/ase-arch-discover command-line option and argument parsing library

Multi-Criteria Decisions

Evaluate alternatives with a multi-criteria decision matrix

Compare alternatives the rigorous way: a weighted multi-criteria decision matrix turns a fuzzy "which is best?" question into a transparent, defensible verdict.

Claude Code

❯/ase-meta-evaluate Commander vs. Yargs vs. Optique, focus on popularity, features, ease of use, and TypeScript support

Play Advocatus Diaboli

Challenge your decision with an Advocatus Diaboli

Know the antitheses to your decision before others bring them up. Let the decision be challenged by a relentless Advocatus Diaboli and even get a resulting synthesis based on Dialectical Reasoning.

Claude Code

❯/ase-meta-diaboli --count=10 We should migrate the ASE project from TypeScript to Rust

Play Steelman

Strengthen your decision with a Steelman

Further improve the strength of your decision before others challenge it. Let the decision be fortified by a helpful Steelman into a stronger position that consolidates everything that genuinely strengthens it while honestly bounding where it holds.

Claude Code

❯/ase-meta-steelman --count=10 ASE is a useful Claude Code plugin

Document Distillation

Distill a document down to its ranked key points

Cut a long document down to its essence: distill it into a flat, importance-ranked list of key points, each with a salience rank, a rationale, and a verbatim, line-cited evidence snippet.

Claude Code

❯/ase-docs-distill --top=3 @plugin/meta/ase-format-spec.md

Document Proofreading

Analyze documents for spelling, punctuation, and grammar

Scan your documents for problems in spelling, punctuation and grammar. Optionally, let the documents be automatically corrected.

Claude Code

❯/ase-docs-proofread --auto @README.md

Logical Code Analysis

Analyze code for logic and control-flow defects

Scan your source code for problems in its logic, semantics, and control flow, with evidence-grounded, line-cited findings instead of vague hand-wavings.

Claude Code

❯/ase-code-analyze --severity=MEDIUM @tool/src/

Alternative Approach Funnel

Craft through a funnel of approaches

Plan-driven, but not too direct and hence with enough preliminary consideration: ASE first proposes a "funnel" of crafting alternative approaches, lets you pick a suitable one, and only then composes a structured task plan.

Claude Code

❯/ase-code-craft --next IMPLEMENT,DELETE hello: "ase hello" CLI command which prints a nice "Hello World!" in blue to the terminal

Named and Persisted Plans

Implement against named, persisted, structured task plans

Use named, persisted, structured and cross-session task plans in a strict, fixed format — then implemented as a complete, surgical change set you stay in control of.

Claude Code

❯/ase-task-implement hello

Automatic ChangeLog

Auto-generate ChangeLog entries from Git

Keep your CHANGELOG.md file current without the chore: derive concise, human-readable changelog entries straight from your Git history and fold them into the existing file under the right version and category.

Claude Code

❯/ase-meta-changelog

Execute just two commands and you are up and running!

Terminal:Installation

❯npm install -g @rse/ase

❯ase setup install [--tool claude|copilot|codex]

Terminal:Updating

❯ase setup update [--tool claude|copilot|codex]

Terminal:Uninstallation

❯ase setup uninstall [--tool claude|copilot|codex]

Terminal:MCP Installation

❯ase setup mcp list

❯export ASE_MCP_KEY_=""

❯ase setup mcp activate [--tool claude|copilot|codex]

Terminal:MCP Uninstallation

❯ase setup mcp deactivate [--tool claude|copilot|codex]

Agentic AI Levels

From Manual To Fully Autonomous

ASE supports Dr. Ralf S. Engelschall's Agentic AI Level Model, which classifies the gradual sharing of work with AI agents.

Level 0 (manual): you do everything manually and without support from the AI agent at all — think "do-it-yourself", where you obviously know the results and inherently train your own skills.

Level 1 (assisted): you are supported by the AI agent for individual steps only — think "assistant", where you still know all results in detail, but inherently your practical execution skills can partly start to wane in the long term.

Level 2 (supervised): you are working on an entire task with the AI agent as an equal pair — think "co-worker", where your colleague shares the task workload with you, you mutually review each other's results, hence you still know all results in detail, but inherently your practical execution skills can be lost in the long term.

Level 3 (delegated): you are delegating entire workflows to the AI agent — think "offshoring", where your colleague shoulders the entire workload for you, you just review the results, but your own practical execution skill and your ability to quality assure the results will be forever lost if you are not regularly, at least virtually, replaying the delegated work.

Level 4 (autonomous): you are assigning entire process lifecycles to the AI agent — think "third-party", where your vendor does everything for you fully autonomously, you just take the results as is, and your own practical execution skill doesn't matter as it is already out of scope.

ASE focuses on the sweet-spot of the Agentic AI Levels!

Skills & Workflows

Step-Wise Driving, You In The Loop

Within the above Agentic AI Levels, ASE provides its various skills to let the engineer trigger the next steps. Some skills are stand-alone, others are part of a particular operation mode that chains skills into a repeatable workflow.

Each skill is named /ase-xxx-xxx and most of those skills even have useful Unix-style options for adjusting their behavior. Click on the skill name in the following table to open its Unix-style man-page in an in-page popup or run /ase-xxx-xxx --help at run-time.

Standalone Mode 1/2

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

Standalone Mode 2/2

ase-arch-

ase-code-

ase-code-

ase-code-

ase-docs-

ase-docs-

ase-meta-

ase-meta-

ase-meta-

ase-meta-

Task Mode

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

ase-task-

Funnel Mode

ase-arch-

ase-code-

ase-code-

ase-code-

ase-code-

Sync Mode

ase-sync-

ase-sync-

ase-sync-

Operation Modes

Advanced Workflows, Maximum Flexibility

On top of the usual Agentic AI Coding tool modes Ad-Hoc Mode and Plan Mode, ASE adds three operation modes:

Task Mode: The more advanced mode where the user initially crafts and then continuously refines a task plan. The task plan has a fixed format and is persisted and hence is available across agent sessions.

Funnel Mode: The even more advanced mode where the user first sketches the plan, then the agent figures out possible approaches, then the user selects one approach, then a task plan is created for this approach, and then finally this switches over to the regular Task Mode. The task plan has a fixed format and is persisted and hence is available across agent sessions.

Sync Mode: The separated reconciliation mode where the user aligns the code and documentation to the specification and architecture ("forward engineering") or even aligns the specification and architecture to the code and documentation ("reverse engineering").

Get 3 new powerful operation modes in your Claude Code!

Ad-Hoc Prompt

Simple Prompting

The simplest approach is a straight and ad-hoc instruction, as provided by Anthropic Claude Code CLI already without ASE. The pro is its simplicity, the con is its weakness due to missing detail decisions. Use for simple tasks only.

Claude Code

❯Create "hello" CLI command which prints "Hello World" in red to the terminal

Plan Mode

Simple Planning

The advanced approach is to use the standard plan mode, as provided by Anthropic Claude Code CLI already without ASE. The pro is its convenience and possibility to interactively add detail decisions, the con is its weakness because plans cannot easily co-exist and be edited in parallel. Use for medium complex tasks under ad-hoc planning.

Claude Code

❯/plan

❯Create "hello" CLI command which prints "Hello World" in red to the terminal

❯(3) Change output color to blue

❯(3) Use NPM package "chalk" for coloring

❯(1) (Yes, and use auto mode)

Task Mode #1

Advanced Planning

The advanced approach is the ASE Task Mode. The pro is that its plans can easily co-exist and be edited in parallel, the con is its weakness because crafting / resolution / refactoring tenets are not internalized by the LLM and hence cannot guide the operation. Use for complex tasks under explicit planning.

Claude Code

❯/ase-t

[truncated for AI cost control]