majiayu000 classify-change
The single most important rule for AI-assisted coding in production: confine the AI's mass output to leaf nodes, where the inevitable bit of tech debt won't propagate through the system. This skill makes the leaf-vs-core call concrete and consistent, so individual judgment doesn't drift across the team.
When to use
Use whenever the leaf-vs-core distinction matters. Triggers include:
- "Is this leaf or core?"
- "What classification is this change?"
- "Should I vibe code this / can AI write this?"
- "How carefully do I need to review this?"
- Before starting non-trivial work, even when the user didn't ask
- Before or during
plan-featureorpr-prepare, when the leaf/core call should be explicit and consistent with this framework
Skip only for changes that are obviously trivial (typo fixes, etc.).
The two categories
Leaf node. Code that nothing else depends on — feature endpoints, UI components, scripts, reports, one-off migrations. Failure is local. Tech debt here doesn't spread. Safe for high AI involvement, lighter review.
Core code. Code many things depend on — auth, payment, data schema, public APIs, shared frameworks, orchestrators. Failure is system-wide. Tech debt here compounds. Requires human leadership and line-by-line review, even if AI assists.
When a change spans both, it counts as core (the stricter rule wins).
The six diagnostic questions
Walk through these in order. Get user input where the answer isn't obvious from the codebase.
Q1 — Propagation range
If we change this, what depends on it?
- Few callers (one or two specific places) → leaf-leaning
- Many callers / many modules → core-leaning
Q2 — Expected change frequency
Will this part of the codebase keep evolving, or is it close to "done"?
- Likely stable for the foreseeable future → leaf-leaning
- Will keep evolving and needs to stay extensible → core-leaning
Q3 — Tech-debt tolerance
If this code becomes a little ugly, will future work suffer?
- A bit of debt is fine, contained → leaf-leaning
- Debt here will block other work → core-leaning
Q4 — Examples that fit
Which of these does the change look most like?
- Reports, dashboards, single endpoints, UI components, scripts, one-off jobs → leaf
- Auth, payments, data schema, public APIs, shared utilities, orchestration → core
Q5 — Failure cost
If this code goes wrong, how far does the failure spread?
- Region of the app, easy rollback → leaf-leaning
- System-wide, affects users / data → core-leaning
Q6 — Review intensity needed
Can the change be trusted via interface + tests, or must someone read every line?
- Interface + tests are enough → leaf-leaning
- Line-by-line read required → core-leaning
Decision logic
After running through the questions:
- All / mostly leaf-leaning → classify as leaf
- All / mostly core-leaning → classify as core
- Mixed / spans both → classify as core (strictest rule wins; consider splitting the change)
Output format
## Classification: [LEAF NODE | CORE CODE]
## Why
- Q1 (propagation): <answer> → <leaf/core>
- Q2 (frequency): <answer> → <leaf/core>
- Q3 (tech debt): <answer> → <leaf/core>
- Q4 (examples): <answer> → <leaf/core>
- Q5 (failure cost): <answer> → <leaf/core>
- Q6 (review intensity): <answer> → <leaf/core>
## Recommended approach
### If LEAF
- Safe for high AI involvement
- 1 reviewer is enough
- Trust the tests + interface; you don't need to read every line
- 3 e2e tests minimum (1 happy + 2 errors)
- Plan still required, but plan.md can be lightweight
### If CORE
- AI may assist, but a human must lead the design and own the review
- 2+ reviewers, including the module owner
- Line-by-line review of the diff
- PR description must mark which files were human-reviewed line-by-line
- If the change is large, split into a human-led core change + a leaf feature change
Edge cases
"It's a leaf, but it touches core code in passing"
If the change reads core code (imports, calls) but doesn't modify core code, it's still leaf. The classification is about which lines change, not which lines exist.
"It's a script / one-off / migration"
These are leaf even if they read core data — failure is bounded by a single run. But: if the script writes to production data, the verification of the script needs to be core-grade (don't trust it from spot-checks alone).
"The code is in a leaf path but the failure mode is system-wide"
E.g., a script that accidentally writes to every user record. Reclassify based on failure cost, not file location. The point of the framework is failure containment.
"Modeling the boundary will keep moving as models improve"
True. Today's core might be tomorrow's leaf as models become more reliable. Tech leads should re-examine the leaf/core boundary quarterly. But don't move the line based on hope — move it based on track record.
Common mis-classifications
- A new endpoint is leaf, even if "the API" is core. A single new endpoint is local. Modifying the framework that all endpoints use is core.
- A new UI page is leaf. Modifying the design system or shared component library is core.
- A new report is leaf. Modifying the BI infrastructure / data warehouse schema is core.
- A bug fix is whatever the file it's in is. Don't auto-classify bug fixes as low-risk; if the bug is in core, treat the fix as core.
Stopping condition
End the skill when the user has a clear classification + the recommended approach for it. Carry the classification into plan-feature or pr-prepare to set the review rigor there — this skill is the canonical owner of the leaf/core definition those workflows should follow.