first-principles-review — скилл для Claude Code: разбор решений

---
name: first-principles-review
description: "Use when asked for first-principles or Occam's-razor review, or when high-risk decisions involve competing constraints, fallback growth, duplicate owners, or architecture direction risk. Ordinary bug fixes stay on the fast path."
---

# First Principles Review

## Purpose

Use this as a lightweight decision review before another Aegis workflow makes a
directional choice. It is a compositional skill, not a standalone workflow.

Do not replace `brainstorming`, `systematic-debugging`, `writing-plans`,
`requesting-code-review`, or `verification-before-completion`. Use it to clean
the decision surface those skills will act on.

When this review materially changes the direction, surface `Aegis Visibility`
in natural prose: name the first principle, dropped assumption, smallest
sufficient path, or owner / retirement falsifier that changed the decision.
Keep it advisory and task-specific; do not turn the lens into approval
authority or a generic skill trace.

## Use When

- The user asks for first principles, first-principles thinking, or Occam's
  razor.
- A design, plan, or fix has multiple plausible paths and unclear selection
  criteria.
- The task has ambiguous goals, competing constraints, or product/architecture
  direction risk.
- Debugging is drifting into repeated fixes, fallback growth, duplicate owners,
  consumer-side patches, or "just add another branch" reasoning.
- A review finds that the implementation may be locally correct but directionally
  wrong.

## Do Not Use

- Simple Q&A, status checks, tiny wording/config edits, or clearly bounded
  single-owner changes.
- Mechanical execution of an approved plan unless a new directional conflict
  appears.
- As a required step for every task, every turn, or every TDD cycle.

## Five-Line Review

Check the agreed goal for omissions and unsupported additions, including
quantities, exclusions, relationships and conditions. Review criteria come from
authorized requirements; this lens cannot add them or veto an authorized
transformation merely because its form differs.

Answer only what is needed, usually in five short lines:

```text
First Principle: What irreducible outcome must this satisfy?
Non-negotiables: What constraints cannot be broken?
Assumptions to Drop: What is habit, inherited shape, or unproven preference?
Smallest Sufficient Path: What is the least complex path that satisfies the first principle?
Escalation Signal: What finding would require spec/design/architecture review?
```

When the direction depends on a new mechanism or an unfamiliar domain, insert
one optional line between path and escalation:

```text
Known Prior Art: proven external pattern worth adopting or adapting to project
constraints (cite source), or `unknown` when precedent cannot be verified here
```

For repairs and optimization, require evidence for the cause or bottleneck
before retaining a solution. Resolve unknowns that could change the choice;
experiments may precede that decision. "Smallest" means the least intervention
covering relevant causal paths while preserving required behavior, not the
smallest diff or fewest owners. Independent causes may need separate repairs;
a shared carrier does not prove a shared cause. Use systematic debugging when
causal coverage is unresolved; no extra checklist is required for a settled case:

```text
Minimality Check:
- Smallest textual diff:
- Correct owner:
- Bug class fixed:
- New branch/fallback added:
- Old path retired or scheduled:
- Verdict: sufficient repair | local patch | needs first-principles review
```

## Decision Hygiene Review

Use this escalation only when a design, fix, or plan needs endorsement before it
is written into a spec or implementation plan.

Escalate from the five-line review when any of these risk signals appear:

- multiple plausible paths and no clear selection criteria
- a new owner, duplicate owner, fallback, adapter, or compat-only carrier
- an old path that may need delete-first handling or a retirement trigger
- an unverified assumption that the proposal depends on
- user language such as "more elegant", "long-term stable", "first
  principles", or "Occam"
- a plan could encode the wrong owner, abstraction, compatibility boundary, or
  retirement schedule
- an existing object, behavior, responsibility, contract, or relationship may
  be reinterpreted, narrowed, replaced, or retired, and the proposal could lose
  a legitimate role or explicit reference while removing invalid authority

Use this compact shape:

```text
First-principles invariants:
- Non-negotiable goal:
- Non-negotiable constraints:
- Historical assumptions to delete:

Bounded preservation reminder:
- Evidence-backed behavior that must remain correct:
- Highest-risk counterexample:
- Material unknown / uninspected surface:
- Known explicit anchors / upstream-downstream refs and disposition:
- Role-before-value ambiguity, if any:

Owner / retirement matrix:
- New canonical owner:
- Old owner:
- Compat-only carrier:
- Delete-first / retirement trigger:

Falsification matrix (evaluate in order; stop at the first failing gate):
- Gate 0 - Premise evidence (stage-graded): design stage accepts spec/logic
  refs; implementation/runtime stage requires log, telemetry, or test
  evidence. No evidence -> park as watch-listed, do not enter value
  evaluation.
- Gate 1 - Decidability: no obtainable evidence to judge it -> park as a
  watched falsifier, do not act.
- Gate 2 - Value (any one): lowers future fix probability / shrinks the
  solution set / exposes a missing actionable acceptance criterion / the same
  broken assumption is reused elsewhere. Passing here still requires
  gates 3-5.
- Gate 3 - Cost match: change cost vs business priority, costed at the
  current stage; a valid but deferred counterexample is recorded in the owner
  doc, never parked orally.
- Gate 4 - Goal regression: the constrained solution still meets the original
  goal; re-read the accepted-constraint list at existing checkpoints (design
  review, plan approval, pre-completion verification).
- Gate 5 - Minimize then classify: boundary-shaped destructive claims must be
  minimized first; premise attacks classify as destructive directly.
  Classes: destructive (refactor/retirement track) | supplementary (boundary
  constraint) | watch-listed (falsifier). Conflicting counterexamples are
  decided by business-goal ranking, never by stacking constraints.
- Gate 6 - Adoption trace: the cheapest checkable form (regression test for
  destructive; contract, assertion, or checklist line for supplementary),
  written into the existing owner doc.

Verdict:
- Adopt / revise / reject / needs evidence:
- Blocking gaps:
- Next evidence:
```

## Architecture Integrity Lens

Use this narrower lens when a proposal is executable but may still encode the
wrong owner, abstraction, contract boundary, or retirement path. It is advisory
method-pack output and may be embedded inside `Decision Hygiene Review` when
that is enough.

Trigger it when any of these appear before approach selection, task
decomposition, review, or completion-risk reporting:

- responsibilities may overlap or a canonical owner is unclear
- the smallest diff adds a caller-side fallback, guard, adapter, or compat-only
  carrier
- an existing source-of-truth or contract could solve the class of problem at a
  higher level
- a stale owner, fallback, or old path may keep carrying real logic
- the work makes a long-term stability, "cleaner architecture", or
  higher-level simplification claim

Use this compact shape:

```text
Architecture Integrity Lens:
- Invariant: What must remain true for the system to be coherent?
- Canonical owner / contract: Which owner, contract, or source-of-truth should carry the behavior?
- Responsibility overlap: What duplicate owner, caller-side patch, fallback, or stale path might still carry real logic?
- Higher-level simplification: Can the problem be solved at the owner / contract / source-of-truth layer instead of by another local branch?
- Retirement / falsifier: What old path retires, or what evidence would disprove this architecture judgment?
- Responsibility / capability boundary: Which invalid authority retires, and
  does the same carrier still serve a separately evidenced legitimate role?
- Verdict: proceed | revise design | split owner | return to baseline | needs ADR/baseline sync
```

`Bounded preservation reminder` is a risk-triggered reasoning aid, not a
universal artifact or an exhaustive behavior inventory. Inspect the smallest
relevant contract, consumer, test, or history evidence. Preserve, rebind,
retire, or reject each known explicit reference; state unresolved relationships
as unknown instead of re-inferring them or claiming semantic completeness. The
Method Pack does not build an authoritative relationship graph, prove
referential integrity or input lineage, calculate complete behavior coverage,
or issue a runtime gate.

Do not run this lens for every low-risk task. If it does not change the
decision surface, return to the active workflow immediately.

## Composition

- With `brainstorming`: run before approach selection when the request is broad,
  ambiguous, likely to inherit a poor product shape, or involves owner /
  retirement / fallback / adapter risk. Use `Decision Hygiene Review` or the
  narrower `Architecture Integrity Lens` before recommending or selecting an
  approach when those signals appear.
- With `systematic-debugging`: run after evidence shows repeated fixes, fallback
  growth, duplicate owners, or consumer-side patching.
- With `writing-plans`: run before task decomposition when the plan could encode
  the wrong owner, abstraction, compatibility boundary, fallback, adapter, or
  retirement schedule. If the approved spec did not already cover this, use
  `Decision Hygiene Review` or the `Architecture Integrity Lens` before writing
  tasks.
- With `requesting-code-review`: run when review should check direction and
  owner integrity, not just code quality.
- With `verification-before-completion`: use only to name residual directional
  risk. It does not grant completion authority.

## Boundaries

- Prefer evidence from current project files, baseline docs, tests, logs, and
  user requirements. If evidence is missing, mark the line as unknown rather
  than inventing a principle.
- Keep the result advisory. This skill may recommend escalation, but it does
  not create authoritative `GateDecision`, `PolicySnapshot`, or completion
  authority.
- If the five-line review does not change the decision surface, return to the
  active workflow immediately.