Story Splitting

Story Splitting

Split work so the team can deliver, test, learn, and change direction in small batches. A good split preserves a whole user-visible capability while reducing variation, uncertainty, scope, or quality level.

The goal is not "make smaller tickets." The goal is N% of the system 100% done and demonstrable, not 100% of the system N% done.

Attribution

This skill is based on Tim Ottinger's Splitting Stories - A Resource Listicle and synthesizes the linked story-splitting writing by Tim Ottinger, Bill Wake, Joshua Kerievsky, Gojko Adzic, Neil Killick, George Dinwiddie, Mike Cohn, Richard Lawrence, Peter Green, J. B. Rainsberger, Rachel Davies, and others. Load resources/source-notes.md for source-by-source provenance.

Deep-dive resources are in the resources/ directory. Load them on demand:

How This Fits With Other Skills

Use story-splitting before planning when the input is a large story, epic, feature idea, roadmap item, or backlog item. This skill discovers small valuable child stories. The planning skill then sequences one selected child story into implementation slices with evidence and delivery details.

Keep story boundaries vertical regardless of the repository's implementation or pull-request workflow. Never turn database, API, UI, and tests into separate backlog outcomes.

Use find-gaps after a split or plan exists when you need to tighten missing states, acceptance criteria, edge cases, or unverifiable language. If find-gaps discovers that the plan is still horizontal or too large, return here and split again.

Use grill-me when installed for unresolved product or design decision-making rather than splitting mechanics. Otherwise use specification and ask one focused question at a time before or after this skill proposes slices.

Use storyboard when the work spans multiple UX surfaces or mock states; the storyboard can reveal missing screens and flow gaps that become child stories. Use design skills such as shape, critique, and polish to improve the mocks themselves, not to replace product slicing.

This skill stops at the product split. Defer implementation workflow, TDD, mutation timing, feature-flag policy, pre-PR gates, branch topology, and delivery mechanics to project instructions and the relevant installed skills. Story splitting neither requires nor prohibits a local mutation run or a feature flag. Use domain and architecture skills only when the selected story reaches implementation; do not split stories by those technical layers.

Requirement Refinement Pipeline

Use the earliest skill that matches the uncertainty:

Do not use story-splitting to interrogate every product decision from scratch; use the optional grill-me skill or the specification fallback when the decision tree is the real work. Do not use story-splitting to produce implementation tasks; use planning after a child story is selected. Do not use find-gaps before there is an artifact to inspect; use it to harden a split, plan, AC set, or mock spec.

Core Principles

Stories Are Conversations, Not Requirements

Treat a story as a placeholder for collaboration between people with the problem and people who can solve it. Do not optimize for perfect written requirements or a handoff from Product to Engineering.

When a story seems too large, bring product, engineering, design, QA, and operations perspectives into the split. Product knows value and trade-offs. Engineers know where effort and uncertainty hide. Splitting without both viewpoints usually creates either low-value fragments or technical dependencies.

Slice Capabilities, Not Components

Every child story should describe a change in behavior from a real user's or external system's point of view. If the "user" in the story is an internal component, layer, queue, database, API, or service, it is probably a task, not a story.

Prefer:

Buyer can pay one invoice by card and receive a confirmation.

Avoid:

Backend exposes payment endpoint.
Frontend calls payment endpoint.
Database stores payment rows.

Tasks can exist inside a story. Do not relabel tasks as stories to make a plan look incremental.

Stay In The Problem Space First

Ask:

What are the simplest options for delivering value to the customer as soon as possible?

Do not start with:

What is the simplest way to build this already-imagined solution?

The first question reveals customer-capability options. The second often jumps straight to portals, apps, services, vendors, and architecture before the team has chosen the narrow valuable capability.

Start With A Primitive Whole

Prefer a tiny, end-to-end whole over a polished part. Useful first versions include an under-featured vertical slice, a structurally uncomposed path, an internal-only capability, or an explicitly labelled disposable tracer.

This first slice should exercise the real behavior path as far as practical. State its release constraint honestly. Whether it uses a feature flag, another release mechanism, or no hiding mechanism at all is a project policy decision. Disposable work must be explicitly labelled as a spike or tracer experiment.

Hunt For Bargains

A bargain is high value at a fraction of full cost. Look for slices where one-fourth of the imagined feature gives most of the value or learning. Treat the parent story as a draft idea; be willing to discover a cheaper or better version through conversation.

Child Story Quality Bar

Use INVEST as a quality check, not a template:

Also apply the feedback triad: a good slice works, delivers value, and can generate user feedback. If it only makes internal progress, it is probably a task. If it works but produces no possible feedback or learning, ask whether a thinner or different slice would be more useful.

Workflow

1. Reframe The Parent

Before splitting, write the parent in plain language:

• Actor: who receives value
• Need or capability: what they can do or learn
• Outcome: why it matters
• Current constraint: why it is too large, risky, vague, or uncertain

If the parent is a solution phrase like "build online banking portal," reframe it as capabilities like "customers can pay bills," "customers can transfer money," or "customers can view balances."

If the parent is a technical task like "add payment database," find the user behavior it serves. If no behavior exists, keep it as a task and attach it to the first story that proves it matters.

2. Surface Examples And Variations

List concrete examples before inventing tickets:

• happy-path examples
• alternate paths and error paths
• user roles or customer segments
• data shapes, sizes, and edge cases
• business rules and policy variants
• interfaces, channels, devices, or integrations
• quality attributes: performance, scale, reliability, security, accessibility, auditability
• unknowns or assumptions that could invalidate the plan

Acceptance examples are often natural splits. Group related examples into separate stories when each group is independently understandable and testable.

3. Find Splitting Dimensions

Use these dimensions as prompts, then choose the split that creates the most value and optionality:

Load resources/pattern-catalog.md when a story needs more prompts or when applying named methods such as SPIDR, Hamburger, Humanizing Work's patterns, or Bill Wake's twenty ways.

4. Use The Hamburger Method When The Team Is Stuck

If the team keeps drifting into technical decomposition, use the Hamburger Method from resources/pattern-catalog.md: list the technical layers, list simpler-to-richer options per layer, then take a first "bite" across every layer so the result is still end-to-end. The point is to avoid eating only one layer.

5. Choose The First Slice

Choose the first slice using these rules of thumb:

• Prefer the split that lets the team deprioritize or delete at least one follow-up story.
• Prefer the split that produces more similarly sized children, if value is comparable.
• Prefer the slice with the highest value-to-cost ratio.
• Prefer an end-to-end slice that burns down integration or architecture risk early.
• In complex domains, pick one or two useful learning slices instead of pretending to enumerate the whole backlog up front.
• If the first slice cannot be useful to anyone, make it useful for learning, ops, support, or validation, and state that explicitly.

6. Write Child Stories

For each child story, capture:

• Title: actor + action + narrow scope
• Value: who benefits and how
• Scope: what is included
• Intentional deferrals: what is not included yet and why
• Acceptance examples: concrete precondition → trigger → observable-outcome examples (or equivalent testable examples)
• Release constraint: shippable, internal-only, demo-only, disposable, or otherwise constrained according to project policy—and why
• Follow-ups: stories that add paths, interfaces, data, rules, or quality

Use the user's domain vocabulary. If they say "buyer," write "buyer," not "user."

Validation Checklist

Before returning a split, check every child story:

• Does it name a real actor or external system?
• Does it have a trigger and observable outcome?
• Does it run through the real production path needed for that behavior?
• Could it be tested without waiting for every sibling story?
• Could it generate feedback, learning, or a visible decision if shipped, demoed, or exercised?
• If no later story happens, did this still deliver value or learning?
• Are dependencies explicit and small enough to manage?
• Are deferred paths, data, rules, and qualities visible?
• Is it safe to ship, or is the release constraint explicit?
• Does it avoid prescribing unnecessary implementation details?
• Is the eventual full capability still visible, so useful follow-up slices are not forgotten?

Red flags:

• child stories split by process activity: design, code, test, document, deploy
• "As a database/API/frontend/service..."
• "Build the backend" and "build the frontend" as separate stories
• "Collect data" or "build integration" without a user-visible or feedback-generating outcome
• every child must finish before any child can be tested
• a spike for every story
• acceptance criteria that say "works" or "fast" without observable detail
• slices that are so tiny they only specify a solution detail
• all business rules forced into the first slice without discussion

Output Format

When asked to split a story, return a compact artifact that can feed directly into planning:

## Parent
[Reframed parent capability]

## Recommended First Slice
[One sentence story]

Why this first: [value, risk, learning, or bargain]

## Split Candidates
| Slice | Value | Includes | Defers | Acceptance Examples | Release Constraint |
|---|---|---|---|---|---|
| ... | ... | ... | ... | ... | ... |

## Parking Lot
[Explicit follow-ups, questions, or intentionally unsplit tasks]

## Warnings
[Component splits, unsafe deferrals, unclear ownership, or missing examples]

## Next Step
[Use the project's planning and delivery instructions for the selected first slice, run `find-gaps` on the split when available, or ask one decision question. Do not prescribe TDD, mutation timing, feature flags, or pre-PR gates here.]

If the user wants an interactive refinement session, ask one high-value question at a time rather than dumping a questionnaire. Start with the question that most changes the split: usually actor, outcome, release constraint, highest-value customer segment, or biggest risk.

For source provenance and article-by-article concept mapping, load resources/source-notes.md.