skills.rest
zhaohuanke123
zhaohuanke123COMMUNITY
Explore all skills by this author →
GitHub

task-driven-ai-dev

Convert a repository into a single-task, evidence-backed delivery loop.

3|Updated Apr 9, 2026
Search Tags:#browser testing#git commit#task orchestration#backlog management#validation gates#durable artifacts#iteration logging
One-click install
npx skills add https://github.com/zhaohuanke123/vanko-skill --skill task-driven-ai-dev
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: task-driven-ai-dev
Source: https://github.com/zhaohuanke123/vanko-skill/tree/main/skills/task-driven-ai-dev
Command: npx skills add https://github.com/zhaohuanke123/vanko-skill --skill task-driven-ai-dev

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

This skill prevents unreliable “run everything” automation by turning a repo into an explicit, durable task-delivery loop where each iteration selects exactly one ready task, implements it within a guarded scope, validates it with evidence, and then stops.

Core Features & Use Cases

  • Task-driven single-iteration delivery: Chooses one ready backlog item conservatively and refuses to mark blocked work as complete.
  • Durable artifact contract: Uses repo files such as architecture.md, task.json, progress.txt, and project-config.json as the source of truth.
  • Validation-first completion: Requires real validation (including browser checks when user-facing flows change) before updating task state and recording testing evidence.
  • Subagent orchestration guidance: Provides decision rules for when to delegate selection/worker/verification roles versus staying local.

Quick Start

Ask the AI to run one guarded delivery iteration for the repo’s next ready item by using task.json plus the provided selection and validation scripts, then update task.json and progress.txt only after the validation evidence is recorded.

Frequently Asked Questions about task-driven-ai-dev

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I prevent AI coding agents from faking task completion in multi-task runs?▼

Single-task AI delivery selects exactly one ready backlog item per iteration, implements it within a guarded scope, and stops after validation, preventing the unreliable fake completions common in run-everything automation.

What is single-task AI delivery and how does it manage backlog iterations?▼

A task-driven AI delivery loop converts a repository into a single-iteration cycle that conservatively selects one ready backlog item, implements it within a guarded scope, validates it with evidence, and stops before proceeding.

How do I structure repository artifacts for deterministic AI task selection?▼

Set up durable planning artifacts by creating architecture.md, task.json, progress.txt, and project-config.json in your repository to serve as the source of truth for deterministic next-task selection and strict progress tracking.

Does the AI delivery loop require browser testing for user-facing flow changes?▼

Validate AI code changes by running validate_iteration.py to enforce validation gates, requiring real test evidence and browser checks for user-facing flows before updating task state in task.json and recording progress.

When should I use subagent orchestration instead of local execution for task delivery?▼

Yes, the delivery loop requires browser checks as part of its validation gates whenever user-facing flows change, ensuring that visual and interactive evidence is captured before marking the task as passed.

What are the limitations of single-task AI delivery loops for repository automation?▼

You should avoid single-task iteration loops if your repository lacks durable planning artifacts like task.json, or if you need to run multiple tasks simultaneously without deterministic selection and strict validation gates.

Related Skills

Explore More Software Engineering Skills→
obsfxobsfx

task-sync

Sync Trekker task data with Codex session plans using Trekker as the source of truth.

Community
Advanced
ricky-yoshricky-yosh

gf-resume

Load greenfield progress files and write next-task.json with task details.

Community
Intermediate
zeronx798zeronx798

requesting-code-review

Scan staged git changes for security issues and quality regressions.

Community
Advanced
okenwaokenwa

maestro

Orchestrate independent tasks in parallel and synthesize results.

Community
Advanced