Scoville Code Anti-AI-Slop

The result is the point. The process earns its place by making that result
safer, clearer, or easier to verify.

It usually looks harmless:

• The agent reports "All tests pass." No tests ran. The suite has passed as
  prose.
• A failing test is declared "pre-existing" after one glance at its name. No
  baseline, comparison, or other evidence appears.
• Lint and typecheck pass, so the changed behavior is reported as verified. The
  behavior itself was not consulted.
• A failing assertion or safety guard is weakened until CI turns green.
  Consensus has been reached. Correctness was not invited.

That is coding slop: evidence is narrated instead of observed. The report says
fixed. The behavior remains unknown, or the check that caught the bug no longer
does.

Scoville Code is a goal-first Agent Skill for planning, changing, testing,
reviewing, and removing code or engineering artifacts. It keeps canonical
ownership, scope, risk boundaries, validation, and honest evidence visible. It
can answer or diagnose without editing, and it does not turn every rename into
a release rehearsal merely because a checklist was feeling ambitious.

Why "Scoville"?

The family is named for useful signal that survives dilution. In coding, the
heat is the requested behavior after plans, wrappers, tests, and confident
status prose have all tried to become the feature.

How to use

Name Scoville Code for codebase work where scope, ownership, risk, or evidence
matters:

Use Scoville Code to implement rate limiting in the existing API owner. Keep the diff scoped, preserve public behavior outside the stated limit, and run the repository's relevant checks.

Use Scoville Code to diagnose why this migration sometimes leaves consumers on the old schema. Identify the supported root cause and evidence. Do not change files.

Use Scoville Code to review this patch for correctness, hidden failure paths, ownership drift, and missing validation. Report prioritized findings only.

Explicit $scoville-code-anti-ai-slop invocation also works on hosts that
support named Skill invocation.

Install

Use an Agent Skills-compatible host and Terra 5.6 Medium or a comparably
capable executor such as Opus 4.8. Ask the agent to install:

Install this Agent Skill and refresh the available Skill list:
https://github.com/benjaminstelzer/scoville-code-anti-ai-slop/tree/main/scoville-code-anti-ai-slop
Keep the installed directory name scoville-code-anti-ai-slop. Use Terra 5.6 Medium or a comparably capable executor such as Opus 4.8.

The final path must end in
<skills-dir>/scoville-code-anti-ai-slop/SKILL.md. For Claude Code, use
~/.claude/skills/ globally or .claude/skills/ inside one project. Other
hosts use their supported Skills directory.

What it costs. The historically qualified Core used 2,045 tokens and was
16.59% larger than v1.0.6. Workflow references load only when needed. This
update leaves the current Core unchanged but expands the Change reference.
Historical token measurements remain bound to their qualified package. Use the
Skill where correctness and maintainability matter. See
benchmark evidence.

What it enforces

• Outcome over ceremony. Plans, tests, docs, and refactors support the
  requested behavior. Producing them is not completion by itself.
• Canonical ownership. The change fits the project's existing architecture,
  records, terminology, and workflow instead of creating a second owner.
• Proportionate risk. Small reversible work stays small. Destructive,
  Public, security, data, or release work receives stronger gates.
• Evidence before claims. Checks prove only what they observed. A failed
  tool is not silently promoted to a passing product.
• Root-cause correction. The agent changes approach after repeated failure
  instead of applying patch number three with renewed optimism.
• Navigable code structure. Hand-written source files use a default ceiling
  of 2,000 physical lines with project priority and concrete exceptions. Domain
  ownership, module boundaries, dependency direction, generated sources, and
  resource cleanup remain explicit without forcing one architecture.
• Material questions only. It asks when a missing choice changes behavior,
  authority, cost, reversibility, or scope, not for details the code settles.
• Complete handoff. The final report names changed behavior, relevant
  validation, unresolved failures, and repository state without pretending.

The complete contract is in
SKILL.md.

How it works

The Core selects an internal mode from Advise, Explore, Develop, or Harden, then
loads only the planning, change-workflow, or validation guidance the operation
needs. Project instructions and established owners outrank Skill defaults. The
Skill creates no private plan or decision log and installs no executable
software. The repository remains the source of truth, which saves everyone
from auditing the audit trail's audit trail.

Scoville family

Each Skill works independently. Combine only the concerns the task actually
needs:

• Brainstorm explores
  materially different mechanisms before selection.
• Research turns web,
  GitHub, and scholarly evidence into a decision-ready, claim-traceable result.
• Code owns
  engineering scope, implementation, risk, and validation.
• UI owns
  interface hierarchy, framework fit, accessibility, and rendered evidence.
• Scribe owns
  wording, terminology, factual meaning, and source fidelity.
• Plan owns durable Plans,
  Work Items, Decisions, and lifecycle state.
• Handoff transfers active
  work to another agent or session.

Status

A reliability-first extension of
Microsoft SkillOpt tested the six
Scoville Skills across 1,201 optimization and evaluation runs. The
previously qualified Code package passed 30/30 final cases. Its Core was
16.59% larger than v1.0.6 because reliability coverage expanded. SkillOpt
compressed that strengthened version.

The new maintainability candidate and the v1.0.17 control each passed 8/8
open Train cases. Both passed 3/4 open Validation cases in three independent
runs. Every failure was the same narrow-security-fix case. One conservative
SkillOpt proposal also passed only 3/4 and was rejected. The four-case Holdout
remains sealed, so this release makes no new qualification claim. See
benchmark evidence.
The family run ledger shows the complete count.

Sources

• OpenAI coding-agent best practices
  for explicit outcomes, constraints, permissions, and focused verification.
• Cursor Thermo-Nuclear Code Quality Review
  for ownership, simplification, and complete-change review.
• OWASP Top 10 for LLM Applications
  for prompt injection, unsafe output, information disclosure, and excessive
  agency.
• Martin Fowler on internal quality
  and technical debt.
• Simon Willison on
  vibe coding versus reviewed AI-assisted engineering.
• Google Engineering Practices
  for design, complexity, tests, naming, consistency, and review context.
• DORA code maintainability
  for source discoverability, dependency traceability, and reproducible builds.
• Configurable file-size checks in ESLint
  and Checkstyle, whose
  different defaults are not treated as one universal standard.

License

MIT - see LICENSE.