Stats

Actions

Available In

Tags

agentic-ai-features

A Claude Code and Codex plugin for building AI features and AI apps agentically: a foundation council that produces the project README before any code is written, hydrates the project spine into a progressive-disclosure agent index, a feature-planner, an autonomous task-loop orchestrator, three isolated implementer / verifier / reviewer roles, a completeness auditor with reachability + AI-eval-coverage passes, and project-spine templates with AI-specific hard rules.

This is not a generic "do code with Claude" kit. It is opinionated about how AI work differs from CRUD work: prompts need eval suites, models must be pinned, cost / latency are first-class gates, "the tests pass" never equals "the feature works", and — before any of that — the team has to agree what they're building, for whom, and why.

Why use it

Building AI features with an LLM-driven coding agent has a particular set of failure modes:

Code starts before the team agrees on the product — vision, target user, architecture and risks live in the founder's head; the implementer guesses; features ship that nobody can confidently say are "good". The plugin's foundation council exists to surface that disagreement before the first task, not in code review.

Code is added but never wired — the implementer writes the class, the tests pass against injected fakes, the build is green, and production never calls it. The user notices weeks later.

Prompts drift silently — someone edits a system prompt; no test fails because no test snapshots the prompt body; output regresses for one of fifty user phrasings.

Models bump invisibly — a SDK default updates, a feature flag flips, and suddenly the feature is on a different model with different cost / latency / behaviour.

Reviewers rubber-stamp themselves — a single agent writes, verifies, and reviews its own output in one context window; it convinces itself the work is fine.

This plugin's whole shape is built around those failure modes.

What it gives you

Claude Code command / Codex skill

What it does

/agentic-ai-features:init / init

Drops a platform spine into the current repo: CLAUDE.md for Claude Code or AGENTS.md for Codex. Idempotent — refuses to overwrite.

/agentic-ai-features:foundation / foundation

Runs a 5-perspective council (user-advocate, investor-advocate, architect, security-auditor, skeptic) on a short pitch, then a synthesizer that produces README.md, docs/foundation/PERSPECTIVES.md, and a gating docs/foundation/OPEN-DECISIONS.md, then a spine hydrator that updates CLAUDE.md/AGENTS.md and creates docs/refs/. Mandatory before any feature work — the three implementation skills below refuse to start until OPEN-DECISIONS is clean.

/agentic-ai-features:feature-planner / feature-planner

Produces a decision-complete plan (tasks with goal, scope, acceptance criteria, verification, AI-eval criteria, cost / latency budget) before any code is written.

/agentic-ai-features:implement-task / implement-task

Implements one task. Dispatches implementer -> verifier -> reviewer as three separate Agent invocations when the platform supports isolated agents. Stops after one task.

/agentic-ai-features:task-loop / task-loop

Walks a multi-task plan file autonomously. Same three-role pipeline per task, plus commit between tasks, stop at human gates, journal to docs/work-log.md for cross-session recovery.

/agentic-ai-features:check-completeness / check-completeness

Audits claimed work against current repo evidence. Includes a reachability pass (catches "added but never called from main()") and an AI-eval-coverage pass (catches prompts without eval files).