Safety Critical Patterns Skill for Claude

Safety-Critical Coding Patterns

Guidelines adapted from NASA's Power of 10 rules for safety-critical software.

When to Apply

Full rigor: Safety-critical systems, financial transactions, data integrity code

Selective application: Business logic, API handlers, core algorithms

Light touch: Scripts, prototypes, non-critical utilities

> "Match rigor to consequence" - The real engineering principle

When NOT To Use

• Ordinary application code, where these defensive checks become the

bloat that prefer-invariants-over-fallbacks targets (use

conserve:code-quality-principles)

The 10 Rules (Adapted)

1. Restrict Control Flow

Avoid goto, setjmp/longjmp, and limit recursion.

Why: Ensures acyclic call graphs that tools can verify.

Adaptation: Recursion acceptable with provable termination (tail recursion, bounded depth).

2. Fixed Loop Bounds

All loops should have verifiable upper bounds.

# Good - bound is clear
for i in range(min(len(items), MAX_ITEMS)):
    process(item)

# Risky - unbounded
while not_done:  # When does this end?
    process_next()

Adaptation: Document expected bounds; add safety limits on potentially unbounded loops.

3. No Dynamic Memory After Initialization

Avoid heap allocation in critical paths after startup.

Why: Prevents allocation failures at runtime.

Adaptation: Pre-allocate pools; use object reuse patterns in hot paths.

4. Function Length ~60 Lines

Functions should fit on one screen/page.

Why: Cognitive limits on comprehension remain valid.

Adaptation: Flexible for declarative code; strict for complex logic.

5. Assertion Density

Include defensive assertions documenting expectations.

def transfer_funds(from_acct, to_acct, amount):
    assert from_acct != to_acct, "Cannot transfer to same account"
    assert amount > 0, "Transfer amount must be positive"
    assert from_acct.balance >= amount, "Insufficient funds"
    # ... implementation

Adaptation: Focus on boundary conditions and invariants, not arbitrary quotas.

6. Minimal Variable Scope

Declare variables at narrowest possible scope.

# Good - scoped tightly
for item in items:
    total = calculate(item)  # Only exists in loop
    results.append(total)

# Avoid - unnecessarily broad
total = 0  # Why is this outside?
for item in items:
    total = calculate(item)
    results.append(total)

7. Check Return Values and Parameters

Validate inputs; never ignore return values.

# Good
result = parse_config(path)
if result is None:
    raise ConfigError(f"Failed to parse {path}")

# Bad
parse_config(path)  # Ignored return

8. Limited Preprocessor/Metaprogramming

Restrict macros, decorators, and code generation.

Why: Makes static analysis possible.

Adaptation: Document metaprogramming thoroughly; prefer explicit over magic.

9. Pointer/Reference Discipline

Limit indirection levels; be explicit about ownership.

Adaptation: Use type hints, avoid deep nesting of optionals, prefer immutable data.

10. Enable All Warnings

Compile/lint with strictest settings from day one.

# Python
ruff check --select=ALL
mypy --strict

# TypeScript
tsc --strict --noImplicitAny

Rules That May Not Apply

| Rule | When to Relax |

|------|---------------|

| No recursion | Tree traversal, parser combinators with bounded depth |

| No dynamic memory | GC languages, short-lived processes |

| 60-line functions | Declarative configs, state machines |

| No function pointers | Callbacks, event handlers, strategies |

Integration

Reference this skill from:

• pensive:code-refinement - Clean code and quality dimension
• sanctum:pr-review - Code quality phase
• /harden - composed in the hardening pipeline
• /full-review safety-critical - focused entry point, and an

auto-detection row when assertion density is low, loops are

unbounded, or recursion lacks a termination proof

Violation Output Format

For each rule violation, report:

Rule N: <rule name>
Location: file.py:42
Anchor: `<verbatim source text at line 42>`
Issue: <what violates the rule>
Fix: <concrete remediation>

Verify Findings Are Grounded (safety-critical:findings-verified)

Every finding must cite a real location and a verbatim anchor. Write

findings to .review/findings.json and confirm each citation resolves:

python plugins/imbue/scripts/citation_verifier.py \
  --findings .review/findings.json --repo-root .

Drop or label UNVERIFIED any finding the verifier fails (exit 1); only

verified findings enter the report. See Skill(imbue:review-core) Step 5

and Skill(imbue:structured-output) for the schema.

Exit Criteria

• [ ] Each of the 10 rules has an explicit verdict for the target

(applies / violated / not applicable), not a silent skip

• [ ] Every reported violation cites a concrete file:line and the

rule number it breaks

• [ ] Rules deemed not applicable name the reason (e.g. "no dynamic

allocation in this module") rather than being omitted

• [ ] Loops flagged under Rule 2 are checked for a statically

provable upper bound; unbounded loops are reported

• [ ] Recursion flagged under Rule 1 is reported when it lacks a

termination argument

• [ ] A summary states whether the target is suitable for

safety-critical use, or which rules block that judgment

• [ ] Every reported violation carries a Location + verbatim Anchor

confirmed by citation_verifier.py (exit 0), or unverified

violations were dropped or labeled UNVERIFIED.

Sources

• NASA JPL Power of 10 Rules (Gerard Holzmann, 2006)
• MISRA C Guidelines
• HN discussion insights on practical application