Build Features End-to-End With a Driver Skill, Not Another Plan Template

Where this fits

You ask your agent to “build a feature end to end,” and it immediately starts editing files. It invents a plan on the fly, skips constraints, and produces code that does not match the system. The build-feature skill solves that by acting as a driver: it loads a design spec phase, an implementation plan phase, and a DAG of leaf skills, forcing approval checkpoints before production code appears.

Why agents benefit

• Separates planning from execution. The agent knows it must not write production code until Plan 1 (design spec) and Plan 2 (implementation plan + DAG) are explicitly approved.
• Preserves existing leaf skills. It names and coordinates leaves like /touch-map or /lego-plan instead of replacing them, so smaller skills stay callable on their own.
• Enforces a single interview method. grill-me is the only interview path, which prevents the agent from drifting into ad hoc question loops.
• Handles refactors cleanly. It is designed for feature work and refactors that are not failure-driven; actual bugs are routed to fix-bug.
• Provides hard gates for ambiguity. Silence is not approval. The agent waits for approved, go, yes, or looks good before moving on.

Practical scenarios

Adding offline mode to a mobile client

The user runs /build-feature Add offline mode. The agent first produces a design spec, gets approval, then builds an implementation DAG with sync, storage, and queue leaves. It never jumps into code before the user confirms the plan.

Refactoring a payment service

A refactor is not a bug, so build-feature is the right entry point. The agent maps the current boundaries, creates a target-state spec, and sequences the refactor leaves in dependency order.

Coordinating multi-file feature work

A user asks for “user notifications” without saying where the logic lives. The skill forces the agent to discover the system, document assumptions, and present a DAG before touching files—especially useful for agents that otherwise optimize for speed over correctness.

Add it to your agent workflow

Install the skill with npx:

npx skills add AlemTuzlak/skills --skill build-feature

Then invoke it directly in a feature request:

/build-feature Add offline mode to the mobile client

Expect two approval blocks: approve the spec, then approve the implementation DAG. Only after the second approval does the agent start producing code.

Compare before adopting

Compare against fix-bug for failure-driven work, and against single leaf skills when the user only needs one step. Look at how actively the pack is maintained, how many real-world issues reference it, and whether the approval gate workflow matches your team’s review process. Quality signals include the clarity of its README, the explicitness of its gates, and whether it composes with skills you already use.

Why it is worth tracking

This skill is a meaningful signal for the agent-skill ecosystem because it tackles orchestration rather than a narrow utility. With 39 stars, it is early but opinionated: the hard gates and driver pattern are exactly what most feature-building agents lack. Check its listing on OpenAgentSkill to track updates and see whether the pack stays fresh before you make it a default part of your agent workflow.