Imported from dvansonsbeek/3d (.claude/skills/functionality-complexity-tradeoff/SKILL.md). Install upstream with npx skills add dvansonsbeek/3d --skill functionality-complexity-tradeoff. Copyright stays with the author.

Functionality pruner

This skill governs decisions about whether functionality justifies its existence. It runs in two stages: a necessity gate (does the problem this code addresses actually occur in this context?) followed by a worth ledger (does the value justify the cost?). It applies equally to unimplemented features (accept / reject / minimize) and to existing code (keep / simplify / delete / remove-as-obsolete). For measuring complexity itself, see structural-simplification. For the upstream principles (YAGNI, scope control, proportional solutions), see architecture-guidelines.

Core Directives

1. Necessity precedes worth. Before scoring value and cost, verify the problem the code addresses can actually occur in this context. Code guarding against architecturally impossible states has no product value for that failure mode. Skip the worth ledger and emit OBSOLETE unless the code is serving as the canonical executable invariant (§1c).
2. Separate the ledger. Value and cost are distinct axes. Score each independently; never collapse into a single number.
3. Cost compounds, value decays. Value is realized per use; cost accrues on every future change, test run, review, and incident. Always evaluate over the feature's expected lifetime.
4. The default is No. If worth is not clearly positive, reject or minimize. YAGNI is the null hypothesis.
5. Build and audit share a model. The same axes apply whether deciding what to add or what to remove. A feature that would fail as a proposal today should fail as existing code today.
6. Remove over refactor, refactor over rewrite. A retrospective audit that finds negative worth — or that fails the necessity gate — prefers safe removal or deprecation to elaborate justification. Removal still follows migration, rollback, and compatibility constraints.
7. Outcome evidence informs worth; it is not the verdict. Consume current linked outcome evidence when available. Completion, deployment, or adoption alone cannot prove downstream value, and no hypothesis state automatically dictates a worth decision.

1. The Necessity Gate

The Worth Model (§2) assumes the code under review is solving a real problem. Before scoring V and C, confirm that the problem itself exists in this stack. If it does not, the worth ledger does not apply: emit OBSOLETE in retrospective mode, or DROP with a necessity-failure rationale in prospective mode. For retrospective removals, apply the safety constraints in §7b before changing code.

[!IMPORTANT] A monorepo single-page application deployed as one artifact cannot run client and server at different versions; a "client version check" in that stack guards against an impossible state. It has no V for that failure mode — not low V — because the failure mode it prevents cannot occur. Worth scoring would mis-classify this as low-V / low-C "DEFER" or "KEEP." The necessity gate catches it.

1a. Categories of non-problem-solving code

1b. Detection heuristics

Run these BEFORE scoring V or C. A high-confidence positive result routes the verdict to OBSOLETE (retrospective) or DROP-as-non-problem (prospective), subject to the invariant-documentation and load-bearing exceptions in §§1c/8e.

[!IMPORTANT] The invariant audit is the highest-yield necessity check. Most non-problem-solving code is defending against violations of invariants the surrounding stack already guarantees. Enumerate those invariants explicitly before reading the code, then walk the branches with the list in hand.

1c. Necessity findings that are not deletions

[!WARNING] Some "impossible-state" code is documenting an invariant rather than enforcing one — an assert version_match whose purpose is to fail loudly if a future contributor changes the deployment topology. That has small but real value as machine-checkable documentation. The fix is usually to convert it to a comment, an ADR reference, a build-time check, or a test — not silent deletion. If the code is the canonical record of an invariant nothing else captures, SIMPLIFY (downgrade to documentation) rather than OBSOLETE.

Inverse failure mode: see §8e. Some complexity that resembles cargo-culting or over-engineering is in fact load-bearing because the simple version was measured to be too slow, too unsafe, or too fragile. Read the original rationale before voting OBSOLETE on anything that merely looks like a non-problem.

1d. Necessity vs. low worth

The distinction matters for the audit record.

Record the distinction so a later audit does not reintroduce the same code under new conditions. "We removed the client version check because it cost more than it returned" invites debate about thresholds; "we removed it because client/server version skew cannot occur in a single-artifact deploy" closes the question.

2. The Worth Model

Once the necessity gate (§1) passes, the question becomes: does the value delivered justify the cost imposed? Worth is the relation between Value (V) delivered and Cost (C) imposed over lifetime L. Both sides are multi-dimensional.

Value axes

Aggregate product value ≈ U × F × R × I. If any axis is zero, ordinary product value is zero; external floors, keystone cost, and safety exceptions are handled separately in §8.

[!IMPORTANT] A feature loved by 2% of users, used once a year, with a trivial workaround, has near-zero total value no matter how elegant it is. Score honestly — especially R and F, which are routinely inflated.

Cost axes

Structural cost is delegated to structural-simplification: the Component-kinds Δ, Dependency-edges Δ, Max-chain-depth Δ, Module-count Δ introduced (prospective) or already present (retrospective). See the Reporting Vocabulary in structural-simplification for the symbol mapping. This skill adds three ongoing-cost axes that structure alone does not capture:

Aggregate cost over lifetime, in axis-symbol form (one-time structural delta plus ongoing maintenance × lifetime):

Aggregate cost ≈ (ΔD + ΔK + ΔP + Δn) + (M + X + E) × L

— where ΔD, ΔK, ΔP, Δn are the structural deltas from structural-simplification (see its Reporting Vocabulary: Component-kinds Δ, Dependency-edges Δ, Max-chain-depth Δ, Module-count Δ).

The worth inequality

Worth > 0   ⇔   V × L   >   C_structural + (M + X + E) × L

For short-lived code, the structural footprint dominates. For long-lived code, M + X + E dominates. Most production features are long-lived; plan for the ongoing term.

[!WARNING] Evolution tax (E) is the most-underestimated axis because it is invisible in the current code review. It shows up later, as the PR that "should have been small but touched twelve files."

3. Two Modes

The model is the same; the inputs differ. Both modes run the necessity gate (§1) before scoring worth.

3a. Prospective — evaluating proposed functionality

Applied to tickets, specs, PRDs, loose ideas, or PR scope before implementation. All inputs are estimates; record confidence explicitly.

1. State the functionality in one sentence: "This allows [who] to [do what] so that [outcome]."
2. Run the necessity gate (§1). Confirm the failure mode addressed is reachable in the target stack and is not already owned by another layer. A prospective necessity failure is rare but consequential: it stops a build that would have produced dead code on day one.
3. Score V axes with evidence: user interviews, request tickets, analytics of the workaround, competitor behavior. Opinions are not evidence. Unsupported opinions are not enough evidence for high-confidence build decisions. Cite linked outcome hypotheses when present; before release they express expected value, not observed impact.
4. Score C axes against a concrete implementation sketch: files touched, new abstractions or dependencies introduced, tests required, failure modes created.
5. Apply the Decision Protocol (§6). Verdicts are prospective (§7a).

3b. Retrospective — auditing existing functionality

Applied to code, modules, features, capabilities, or flags that already exist. Inputs are observable; bias toward measurement over judgment.

1. Define the boundary: files, symbols, entry points, feature flags, routes, or callers.
2. Run the necessity gate (§1). Walk the heuristics in §1b before any worth scoring. A positive finding short-circuits the rest of the audit to OBSOLETE.
3. Score V from usage data:
   • Telemetry hits per time window, per user cohort.
   • Reach: unique users or flows that enter this code path.
   • Irreplaceability: does an alternative path exist? Do users already use it?
   • If V cannot be measured, that itself is a finding — instrument, identify an external floor (§8c), or keep confidence Low.
   • Consume current outcome-evidence records from requirements-traceability. Preserve the canonical hypothesis version, cohort, threshold, window, and guardrails. stale or inconclusive evidence cannot support High value confidence; rejected evidence lowers the supported value claim but does not by itself prove zero value.
4. Score C from current observable state:
   • Structural: measure D, K, P, n per structural-simplification.
   • M: dedicated tests, doc pages, recent commit churn, dependency drift.
   • X: bug ticket history, incident postmortems, security/perf hotspot reports.
   • E: count of PRs / design docs where this feature caused scope expansion, workarounds, or delays.
5. Apply the Decision Protocol (§6). Verdicts are retrospective (§7b).

[!NOTE] A retrospective audit with no telemetry available should first return an instrumentation task, not a verdict — unless the necessity gate has already produced a finding, in which case telemetry is not needed (you cannot measure usage of a code path that cannot be triggered). Deciding to delete a feature merely because you cannot see it being used is survivorship bias in reverse; deciding to remove it because the failure mode it guards against cannot occur is structural reasoning.

4. Heuristic Checks

Fast signals — not substitutes for measurement. The necessity-gate heuristics in §1b run first; the table below covers worth-related signals that apply once necessity has passed.

[!IMPORTANT] Churn × complexity is a strong empirical signal for "code that costs more than it returns" (Tornhill, Your Code as a Crime Scene). Files that change often AND score high on cyclomatic or cognitive complexity are disproportionately responsible for defects and maintenance spend. Run this check before any subjective judgment in retrospective mode.

[!NOTE] Usage silence and zero-everything signature look identical from outside. The difference: usage silence assumes the path is reachable but unused (low V); zero-everything signature, combined with an invariant audit, suggests the path is unreachable (failed necessity). The verdicts diverge — DELETE vs. OBSOLETE — and the rationales close the question with different durability.

5. Forcing Questions

Each question exposes a common failure mode. Answers MUST be written, not implicit.

Necessity interrogation

Apply BEFORE value interrogation. If any answer is "no" or "we cannot construct one" after checking the relevant callers and runtime paths, the code is a candidate for OBSOLETE / DROP-as-non-problem.

• Can the failure mode this guards against actually occur given the deployment topology, type system, and runtime guarantees of this stack?
• Construct one concrete real-world sequence that activates this code without violating an architectural invariant. Can you?
• Is the concern owned by another layer (framework, middleware, type system, deployment topology, network boundary)? Is that layer already enforcing it?
• Do the prerequisites of the pattern this implements hold here (long-lived process, multiple implementations, non-idempotent dependency, mutable shared state, etc.)?
• Does the original rationale still apply, or has the world it described — the dependency, the platform, the client class, the ongoing migration — changed?
• If this is documenting an invariant rather than enforcing one, is there a cheaper place for that documentation (comment, ADR, build-time check, test)?

Value interrogation

• Who specifically needs this? Roles, counts, cohorts — not "users".
• What do they do today without it? If nothing, the value may be imagined.
• What is the simplest alternative that would satisfy 80% of the need? (CLI, config, docs, external tool, manual process, nothing at all.)
• What evidence — not opinion — supports the V estimate?
• What is the smallest useful slice we could ship and still claim the win?

Cost interrogation

• What new vocabulary — concepts, abstractions, types — does this add? (Component-kinds Δ)
• What currently-independent parts does this link? (Dependency-edges Δ)
• How long is the dependency chain a typical change traverses once this exists? (Max-chain-depth Δ)
• How many tests — including error paths, edge cases, and integration — will this require? (M)
• If this breaks, what else breaks with it? What is the blast radius? (X)
• What future change does this make harder, slower, or more dangerous? (E)

Counterfactual

• If we delete this in 12 months, what is the removal cost?
• If we never build it, what is the realistic worst outcome?
• Is there a non-code solution (docs, training, config, external tool, process change)?

[!WARNING] If the removal cost in 12 months exceeds the build cost today, this is a one-way door. Apply §8 asymmetric trade-offs before committing. One-way doors demand higher V and greater confidence.

6. Decision Protocol

1. Run the necessity gate (§1). Walk the heuristics in §1b. If the code addresses a problem that cannot occur in this context, emit OBSOLETE (retrospective) or DROP with a necessity-failure rationale (prospective). Skip remaining steps.
2. Score V axes (U, F, R, I) on a 0–3 scale with one-line evidence per axis. When outcome evidence exists, cite its hypothesis ID, state, freshness, and observation identity; do not replace its threshold or guardrails with a more favorable interpretation.
3. Score C axes:
   • Delegate D, K, P, n to structural-simplification (deltas for prospective; absolute measured values for retrospective).
   • Score M, X, E on 0–3 with one-line evidence per axis.
4. Record confidence (Low / Medium / High) for each side independently.
5. Compare across both ledgers without summing.
6. Classify using the Worth Matrix (§6a) and apply the confidence gate (§6b).
7. Emit the Output Contract (§9).

6a. The Worth Matrix

Read the matrix identically in both modes. Prospective verdicts are accept / reject; retrospective verdicts are keep / simplify / delete. The matrix only applies when the necessity gate (§1) has passed; necessity failures bypass it entirely.

6b. Confidence gate

A verdict carries the confidence of its weakest input. If either V or C confidence is Low:

• Prospective → default to DEFER. Gather evidence before committing to high-cost action.
• Retrospective → default to QUARANTINE. Add instrumentation, revisit after N weeks with measured data.

Do not commit to irreversible verdicts (BUILD, DELETE) on low-confidence estimates. OBSOLETE is exempt from the confidence gate when the necessity finding is itself High confidence — a structural impossibility does not become more or less impossible with more data.

An unmeasured, inconclusive, or stale outcome assessment keeps the affected value claim Low unless independent current evidence supports it. supported may raise confidence only within the measured cohort, window, and guardrails. Authoritative floors in §8c remain source-driven and do not require an empirical outcome hypothesis.

7. Verdicts

7a. Prospective verdicts

7b. Retrospective verdicts

[!WARNING] "Interesting", "clever", and "elegant" are not verdicts. Cleverness imposes cost but rarely contributes measurable value. If a reviewer's rationale reduces to "it's nice that we have this," require value evidence before KEEP. If it reduces to "it's defensive — just in case," apply the necessity gate before defaulting to KEEP.

8. Asymmetric Trade-offs

Cases where the Worth Matrix gives the wrong answer on its own.

8a. Optionality premium

A low-V / low-C feature may be worth keeping or building if it preserves concrete future optionality — a known next feature whose path becomes cheap because of it.

Test: is the next feature named and probable, or is the optionality speculative? Speculative optionality fails YAGNI; the null hypothesis wins. Note the overlap with generality without instantiation (§1a): an abstraction whose anticipated variation never materialized fails both this test and the necessity gate.

8b. Irreversibility tax

A feature that is hard to remove once shipped — public API, persisted schema, user-visible behavior, wire format — must clear a higher bar. Raise the required V by one tier, or require High confidence.

8c. Regulatory / contractual / accessibility floor

Some features deliver V that cannot be observed from usage telemetry: audit logs, accessibility paths, legal holds, compliance records, safety interlocks. Assign a fixed-high U regardless of F × R; C is still measured normally. These features are kept even when "unused" when the applicable external requirement, jurisdiction, contract, or safety case is identified. They pass the necessity gate only after that requirement is mapped to this code path.

8d. Keystone cost

Some features have high local C because they are the seam holding a correct abstraction in place. Removing them would raise global complexity elsewhere. Measure net Component-kinds Δ, Dependency-edges Δ, Max-chain-depth Δ, Module-count Δ across the whole system before committing to DELETE or SIMPLIFY. A local reduction that increases global complexity is not a simplification (see structural-simplification Core Directive 5).

8e. Hot-path performance or safety

Some complexity exists because the simple version was measured to be too slow, too unsafe, or too fragile. C appears inflated but is structurally load-bearing. The audit must read the original rationale (commit message, ADR, benchmark) before voting SIMPLIFY.

[!IMPORTANT] §8e is the inverse of the necessity gate. Necessity failures look load-bearing but aren't; §8e cases look cargo-culted but are. Origin archaeology is the shared diagnostic — the difference is whether the rationale's premises still hold today (necessity passes, §8e applies) or have lapsed (necessity fails, OBSOLETE applies). Lost history is not permission to remove load-bearing complexity; lapsed history supports removing obsolete code after the normal safety checks.

9. Output Contract

Every application of this skill MUST produce a coder-facing decision record. Keep fields concrete enough for Codex to choose the next edit, test, telemetry task, or rejection:

Subject:        <feature / module / ticket / path under review>
Mode:           Prospective | Retrospective
Necessity:      Pass | Fail
Necessity note: <if Fail: which §1a category, which invariant violated /
                 prerequisite missing / premise lapsed; one line.
                 If Pass and non-trivial: brief note on what made it pass.>
V scores:       U=<0-3>  F=<0-3>  R=<0-3>  I=<0-3>       (1-line evidence each; OMIT if Necessity=Fail)
C scores:       Component-kinds Δ=<±n>  Dependency-edges Δ=<±n>
                Max-chain-depth Δ=<±n>  Module-count Δ=<±n>      (prospective: deltas; retrospective: measured absolutes; OMIT if Necessity=Fail)
                M=<0-3>  X=<0-3>  E=<0-3>                 (1-line evidence each; OMIT if Necessity=Fail)
Confidence V:   Low | Medium | High                       (OMIT if Necessity=Fail)
Confidence C:   Low | Medium | High                       (OMIT if Necessity=Fail)
Outcome evidence: <hypothesis IDs, states, freshness, observation links, or none / not applicable>
Decision:       <BUILD | BUILD-minimal | NEGOTIATE | DEFER | DROP | KEEP | SIMPLIFY | QUARANTINE | DEPRECATE | DELETE | OBSOLETE>
Rationale:      <2–4 sentences tying scores → decision, or necessity finding → OBSOLETE>
Next action:    <build minimal slice, delete path, add telemetry, write test, update lint rule, or stop>
Verification:   <command / telemetry / caller check / Not run + reason>
Minimal alt:    <smallest slice preserving most V, if applicable>
Revisit when:   <measurable trigger or calendar date>

[!IMPORTANT] Revisit when is non-optional for DEFER, QUARANTINE, BUILD-minimal, and DEPRECATE. Every such decision MUST have a measurable trigger (usage threshold, date, dependency version, adjacent feature shipping) or it will rot into a permanent maybe.

[!NOTE] OBSOLETE does not require a Revisit when, because the trigger for revisiting is implicit: a change in the stack's invariants. If the deployment topology ever splits, the type system loosens, the upstream layer's guarantee is removed, or the lapsed dependency returns, the necessity finding becomes invalid and the case reopens automatically.

10. Common Patterns

11. Composition with Sibling Skills

• requirements-grounding — owns outcome-hypothesis meaning, thresholds, cohorts, guardrails, and revisit intent. This skill consumes that definition; it does not rewrite it.
• requirements-traceability — owns measurement links, evidence state, and freshness for the exact hypothesis version. This skill consumes its current assessment and alone issues the functionality-worth verdict.
• structural-simplification — source of the complexity measurement (D, K, P, n). This skill consumes those deltas; it does not redefine them.
• architecture-guidelines — upstream principles (YAGNI, scope control, proportionality, deletion over patching). This skill is the applied protocol through which those principles bind to individual decisions. The necessity gate (§1) is the most direct expression of YAGNI applied to existing code: "you ain't gonna need it" generalizes to "you never needed it; the problem was never in this context."
• continuous-improvement — when this skill's verdicts repeatedly contradict current practice or sibling skills, that is a signal to update the skills themselves, not to override the verdicts case-by-case. Repeated OBSOLETE findings in a single area, in particular, are a signal to update architecture-guidelines with the relevant invariant so future contributors do not re-introduce the same non-problem-solving code.

[!NOTE] This skill deliberately does not define its own complexity metric. Cyclomatic complexity, cognitive complexity, Halstead volume, and maintainability index are all input signals to the C side of the ledger, surfaced through structural-simplification and the churn × complexity heuristic. Keeping the measurement in one place preserves the single-source-of-truth discipline across the skill library. Likewise, the necessity gate (§1) does not redefine architectural invariants — it consumes the invariants documented in architecture-guidelines and the stack's own ADRs, and uses them as the basis for impossibility findings.