Agnostic Planning Skills
Ten skills. Four agents. One non-negotiable rule: no code without an approved plan. Turn your AI from eager coder into disciplined product collaborator — regardless of your tech stack.
The Problem
By default, AI coding assistants jump straight to code. No PRD. No task breakdown. No risk assessment. They generate plausible-looking output that passes a glance but lacks structure, test coverage, and traceable scope.
The Mental Model
Every feature travels through stages before code is written. Each stage produces an artifact and gates block progress until the artifact is reviewed. You can enter and exit at any stage.
The full pipeline: create-prd → review-prd → [Gate] → generate-tasks → estimate-tasks → identify-risks → prioritize-backlog → plan-tickets → plan-sprint → generate-status-report → create-retrospective
The Rule
This one rule is encoded into every skill and agent. The PRD is the single source of truth for scope. Every HARD-GATE blocks the AI from proceeding until you explicitly approve the artifact.
HARD-GATE: PRD Approval
The AI must present the PRD draft and wait for explicit approval. It cannot generate tasks, create tickets, or write a single line of code until you say yes.
HARD-GATE: Ticket Approval
Tickets are always draft-only by default. The AI will not create issues in your tracker unless you explicitly ask it to. No ghost tickets.
HARD-GATE: Sprint Commitment
The agent does not assume sprint IDs, capacity, or team availability. Sprint placement must be explicitly confirmed by you.
The Tools
Every skill has the same structure: frontmatter, quick reference, HARD-GATE rules, step-by-step process, output template, and integration table. They auto-detect your project's tooling — no config needed.
PRD
Generate a structured Product Requirements Document from a feature description: goals, user stories, functional requirements, success metrics, and open questions.
Evaluate PRDs for completeness, testability, clarity, and feasibility. Findings classified as Critical, Suggestion, or Note. Every finding cites specific document evidence.
Task Management
Break a feature into TDD task checklists. Task 0.0 is always "create feature branch." Uses TDD quadruplet: write test → watch it fail → implement → watch it pass.
Draft tracker-ready tickets with classification by type, area, execution order, and dependency level. Default: draft-only. Never creates tickets without explicit permission.
Assign effort estimates using story points (Fibonacci: 1,2,3,5,8,13), t-shirt sizes (XS–XL), or time ranges. Every estimate includes a confidence level: High, Medium, or Low.
Backlog
Rank items using RICE (Reach × Impact × Confidence / Effort), MoSCoW, value-vs-effort matrix, or WSJF. Must state rationale for every ranking decision.
Ceremony
Select tickets from a prioritized backlog. Define a single sprint goal. Cap capacity at 80% of velocity. Uses team_size × sprint_days × 6h × utilization_factor.
Generate structured retrospectives: What Went Well, What Didn't, Action Items (with Owner + Timeline), Metrics, and Kudos. Every "what didn't" must have a linked action item.
Execution
Scan plans for 6 risk patterns: dependency chains, single points of failure, ambiguous requirements, external dependencies, capacity issues, and technical uncertainty. Every risk cites specific evidence from the plan.
Produce stakeholder status reports: Executive Summary, Accomplishments, In Progress, Blocked, Risks, Next Steps. Never fabricates progress — marks unknown items as "needs update."
The Orchestrators
Agents don't just run one skill. They walk through a lifecycle: Discovery → Draft → Review → Estimate → Ticket → Sprint. Each agent has approval gates that pause the flow until you confirm.
Discovery → PRD Draft → Review & Revise → Task Estimation → Ticket Generation → Sprint Placement
Estimation → Risk Assessment → Tracking Setup → Status Reporting
PRD Review → Feasibility Assessment → Estimation Quality Review → Technical Risk Report. Produces Go/No-Go/Go-with-Conditions.
Scope → Plan → Prioritize → Sprint → Execute → Retrospect. The meta-agent — orchestrates all 10 skills across the complete delivery lifecycle.
Walkthrough
Here's what happens when you use three skills in sequence on a real feature.
You start with a feature idea
The AI reads the project context, then drafts a structured PRD in markdown — saved to /tasks/prd-google-oauth-login.md. It covers goals, user stories (login with Google, account linking, error handling), functional requirements, non-goals (no password fallback), design considerations, and success metrics.
You review and approve the PRD
The AI stops and asks: "Does this PRD capture the right scope? Anything to add or remove?" The gate holds. No tasks, no tickets, no code — until you say "approved." You can also invoke @review-prd for a structured quality check with findings classified as Critical, Suggestion, or Note.
The AI generates a TDD task checklist
Each feature turns into a task quadruplet:
feat/google-oauth-login
/auth/google/callback route
The AI auto-detected your test command (npm test) from package.json and source directory (src/).
The AI drafts tracker-ready tickets
Each ticket gets classification — type (feature), area (auth), execution order (1), dependency level (none), target bucket (current sprint). The five-section format includes title, description, acceptance criteria, technical notes, and definition of done. Still draft-only — the AI waits for your go-ahead before creating issues in GitHub, Jira, or Linear.
You're now ready for development.
What was one vague sentence is now a reviewed PRD, a TDD task checklist with exact file paths, and draft tickets classified for your tracker. You have clear scope, test-first tasks, and no ambiguity about what to build next.
No Config Required
Skills auto-detect your environment by inspecting the files already in your repository. No manual configuration, no setup file, no framework lock-in.
Get Started
GitHub CLI
Requires gh v2.90.0+
skills.sh
Requires skills.sh CLI
Your first command
The AI drafts a PRD, stops for your approval, and only proceeds when you say yes. That's the gate working. That's the whole point.
Quick Reference
You don't need to run the full pipeline every time. Pick the path that matches your need.
New feature from scratch
Prioritize existing work
Full delivery cycle
Orchestrates all 10 skills across 6 phases
Remember This
"No implementation without an approved PRD."
If you take one idea from this page, take that. Every skill, every agent, every gate in this library exists to enforce that one rule. The AI becomes a planning partner, not a code cannon. Scope is explicit. Tests are written first. Nothing ships by accident.