bigeasyfreeman/skills/stop-slop

Stop Slop v2: Code + Content Pattern Removal

Dual-mode slop guidance for generated-output surfaces. Mode 1 names deterministic code checks that may run when project config enables them. Mode 2 defines rubric-based content and product-output evaluation. ADLC core gates the Build Brief contract with bin/adlc slop-gate; project-specific linters or content checkers can plug in behind that contract.

---

Output Eval Loop

Stop Slop is an output-side quality gate, not a prompt-improvement checklist. A better prompt can still produce bad output; ADLC only trusts output after it has been scored against a benchmark.

For any generated-output surface, the benchmark is:

{
  "slop_quality_gate": {
    "applicability": "required | not_applicable",
    "reason": "string",
    "mode": "code | content | product_output | agent_output | mixed",
    "threshold": 0.7,
    "metrics": ["rubric_score | exact_match | schema_validity | semantic_similarity | test_strength"],
    "eval_cases": [
      {
        "id": "SLOP-001",
        "source": "golden | human_edit | council_rejection | runtime_failure | production_sample | incident | support_ticket | analytics_drop | other",
        "input": "string",
        "expected_quality": "string",
        "golden_output": "optional string",
        "rubric": "optional string",
        "metric": "string",
        "threshold": 0.7
      }
    ],
    "baseline_score": 0.82,
    "regression_tolerance": 0.03,
    "failure_action": "block | revise | human_approval | monitor",
    "case_promotion_sources": ["human_edit", "council_rejection", "production_sample"]
  }
}

Run the loop in three places:

1. Pre-ship regression: prompt, model, content-template, agent-role, or product-output changes re-run saved eval cases and compare against the baseline.
2. Delivery guard: output below threshold is blocked or revised before it reaches a user, ticket, issue, PR, document, or release.
3. Post-ship sampling: live generated-output systems sample real executions and score them with the same benchmark.

Default thresholds:

• Code: hard failures block; warnings follow the code slop scoring below; verification quality delegates to test-strength (coverage >= 0.8, mutation_kill_rate >= 0.6, no material surviving mutants).
• Content, product output, and agent output: score >= 0.70 unless the Build Brief sets a stricter threshold.
• Any score regression larger than regression_tolerance blocks unless failure_action = human_approval and the human approval is captured.

Every failure should produce a candidate eval case before it becomes a new style rule. Candidate sources include human edits, council rejections, runtime failures, production samples, incidents, support tickets, and analytics drops tied to generated-output quality.

See docs/specs/slop-eval-loop.md.

Mode 1: Code Slop

Detects structural code problems that indicate incomplete, lazy, or AI-generated code. This mode runs only when project configuration or the task's slop_quality_gate opts into code-output checks.

Detection Patterns

1. Placeholder Detection

Regex rules are path-aware. A match is a hard failure only in shipped executable source where the pattern leaves incomplete behavior. Tests, fixtures, examples, and docs may use TODO/FIXME or placeholders when the surrounding verifier proves they are intentional.

pass\s*(#.*)?$          # bare pass statements
TODO\b                  # TODO comments in shipped executable source only
FIXME\b                 # FIXME comments in shipped executable source only
raise NotImplementedError
\.\.\.\s*$              # ellipsis as function body

AST rules (Python):

• Function body is a single pass or Expr(Constant(Ellipsis)) node
• Class body is a single pass node with no docstring

AST rules (TypeScript/JavaScript):

• Function body is empty block {}
• Function body throws new Error("not implemented")

2. God Function Growth

Ceiling: 50 SLOC per function (default). Configurable via .stop-slop.yml:

code_slop:
  max_function_sloc: 50

Measurement: count non-blank, non-comment, non-decorator lines inside the function body. Exceeding the ceiling triggers a warning. Exceeding 2x the ceiling is a hard failure.

3. Near-Duplicate Code Blocks

Flag any two code blocks (3+ lines) with >80% token similarity within the same file or across files in the same commit. Uses normalized token comparison (strip whitespace, normalize variable names to positional placeholders).

Threshold: 80% similarity = warning. 95% similarity = hard failure (copy-paste detected).

4. Identity Transforms in Comprehensions

Regex/AST detection for comprehensions that return the loop variable unchanged:

# Flagged:
[x for x in items]
[item for item in collection]
{k: v for k, v in d.items()}

# Not flagged (these transform):
[x.name for x in items]
[x for x in items if x.active]

5. Unnecessary Defensive Comparisons

Detect patterns where the comparison is redundant:

# Flagged:
if x == True:
if x == False:
if x == None:
if len(items) > 0:
if len(items) != 0:
if bool(x):

# Fix:
if x:
if not x:
if x is None:
if items:
if items:
if x:

6. Hardcoded URLs, Ports, Timeouts Outside Config

Regex patterns for values that belong in configuration:

https?://[^\s"']+            # URLs (except in config/, .env, tests/)
:\d{4,5}[/"'\s]             # port numbers
timeout\s*=\s*\d+           # hardcoded timeouts
sleep\(\s*\d+               # hardcoded sleep values

Exemptions: files matching **/config/**, **/.env*, **/test*/**, **/fixture*/**, **/mock*/**.

7. Missing Integration Wiring

Detect functions/classes that are defined but never imported or called elsewhere in the project. Applies to:

• Exported functions not imported anywhere
• Route handlers not registered in a router
• Event handlers not subscribed to any emitter
• CLI commands defined but not added to a command group

8. Missing TDD Protocol

For any new function added in a commit, check that a corresponding test file or test function exists. Naming conventions checked:

• test_<function_name> in tests/ or test_*.py
• <function_name>.test.ts or <function_name>.spec.ts
• describe("<FunctionName>" block in test files

Missing test coverage for new functions = warning. Missing test file entirely = hard failure.

Code Slop Scoring

Hard failures block the commit/delivery. Warnings accumulate:

• 0 warnings: clean
• 1-3 warnings: proceed with notice
• 4+ warnings: block until addressed

---

Mode 2: Content Slop

Detects AI writing patterns in prose, product output, and agent output when the Build Brief marks a generated-output surface active. It does not run on every .md or .txt file by default.

Content Mode Flow

1. Run the deterministic regex / banned-pattern pass first.
2. If regex finds a hard-fail pattern, stop immediately.
3. If regex clears, run slop-judge to catch residual generic filler, passive evasion, and tautology.
4. Score only the text that clears both the deterministic pass and the slop judge.

Slop Judge Input

{
  "mode": "general | outreach | product_output | agent_output",
  "regex_screen": "pass",
  "content": "string",
  "audience": "internal | external",
  "rubric": ["specific criterion"],
  "threshold": 0.7,
  "baseline_score": 0.82,
  "regression_tolerance": 0.03
}

Slop Judge Output

{
  "verdict": "pass | revise",
  "score": 0.0,
  "threshold": 0.7,
  "criterion_scores": [
    {
      "criterion": "specificity",
      "score": 0.0,
      "reason": "string"
    }
  ],
  "regression_delta": 0.0,
  "rationale": "string",
  "signals": ["generic_filler", "passive_evasion", "tautology"],
  "new_eval_case_candidate": {
    "source": "council_rejection | human_edit | runtime_failure | production_sample | other",
    "input": "optional string",
    "bad_output": "string",
    "expected_quality": "string",
    "metric": "rubric_score",
    "threshold": 0.7
  }
}

Rubric Dimensions

These dimensions guide scoring. They are not blanket bans in ADLC core.

1. Active voice. Human subjects. Find the actor. Make them the subject. "The team shipped it" not "It was shipped." "You read the data and concluded" not "The data tells us."

2. Reduce filler adverbs. Flag repeated filler adverbs when they weaken specificity. Do not fail a piece because one legitimate adverb appears.

3. No throat-clearing openers. State the point. Not "Here's the thing:" or "Let me be clear:" or "The uncomfortable truth is." Start with the content.

4. Avoid empty binary contrasts. Flag "Not X. But Y." only when the contrast adds no information.

5. No rhetorical setups ("Here's what I mean:"). "Here's what I mean:" becomes the meaning itself. "Think about it:" becomes trust the reader. "What if I told you" becomes just tell them.

6. No vague declaratives. "The implications are significant" says nothing. Name the implication. "The stakes are high" says nothing. Name the stake.

7. No dramatic fragmentation (staccato lists). "Speed. Quality. Cost." becomes "Speed, quality, cost." Complete sentences. No staccato drama.

8. Specificity over abstraction. Put the reader in the room. Concrete scenes over narrator-from-a-distance. "You sit down on Monday and realize your pipeline is empty" beats "Teams often struggle with pipeline consistency."

Banned Phrases (Immediate Removal)

Throat-clearing

• "Here's the thing:" / "Here's what X" / "Here's why X"
• "The uncomfortable truth is" / "The real X is"
• "Let me be clear" / "I'll say it again:" / "I'm going to be honest"
• "Can we talk about" / "It turns out"

Emphasis crutches

• "Full stop." / "Period." / "Let that sink in."
• "Make no mistake" / "This matters because"

Business jargon (replace with plain language)

• navigate → handle / address
• unpack → explain / examine
• lean into → accept / embrace
• landscape → situation / field
• game-changer → significant
• synergy → (delete)
• double down → commit
• deep dive → analysis
• circle back → return to
• moving forward → next

Adverbs

Repeated filler adverbs can lower the rubric score when they replace concrete detail.

Filler phrases

• "At its core" / "In today's X" / "It's worth noting"
• "At the end of the day" / "When it comes to" / "In a world where"
• "The reality is" / "What this means is"

Meta-commentary

• "As we'll see..." / "Let me walk you through..." / "In this section..."
• "I want to explore..." / "Plot twist:" / "Spoiler:"

Vague declaratives

• "The implications are significant"
• "The reasons are structural"
• "The stakes are high" / "The consequences are real"
• "This is genuinely hard" / "This is what X actually looks like"

Structural Patterns to Remove

Binary contrasts → state directly

| Pattern | Fix | |---------|-----| | "Not X. But Y." | Say Y. | | "X isn't the problem. Y is." | "Y is the problem." | | "The answer isn't X. It's Y." | "Y." | | "Not just X but also Y" | "X and Y" or just "Y" | | "stops being X and starts being Y" | "becomes Y" |

False agency → name the human

| Pattern | Fix | |---------|-----| | "the data tells us" | "we read the data and concluded" | | "the culture shifts" | "people changed how they work" | | "the market rewards" | "buyers pay for" | | "the decision emerges" | "X decided" | | "the conversation moves toward" | "they steered toward" |

Sentence structure

• Prefer commas or periods over decorative em dashes.
• Three-item lists are allowed when they carry real distinctions.
• Sentences may start with What/When/Where/Which/Who/Why/How when the question is doing real work.
• Paragraphs may start with "So" when the transition is useful.
• Vary sentence length. No three consecutive equal-length sentences.
• No every paragraph ending with a punchy one-liner.

5-Dimension Scoring (1-10 each, total out of 50)

| Dimension | What it measures | Signs of failure | |-----------|-----------------|-----------------| | Directness | How much filtering/softening exists | Qualifiers, hedges, throat-clearing before the point | | Rhythm | Sentence length variation | Three equal-length sentences in a row; staccato fragmentation | | Trust | Authenticity; no false sincerity | "I promise," performative emphasis, manufactured drama | | Authenticity | Specificity and concrete detail | Abstract claims, narrator-from-a-distance, vague declaratives | | Density | Compression; no filler | Phrases that add words without adding meaning |

Score bands:

• 43-50: Ship as-is
• 35-42: Minor revision before human review
• 25-34: Significant revision required — do not submit to human
• Below 25: Rewrite from scratch

Thresholds:

• General content (Magnus, internal): 35/50
• Outreach and external-facing (Sander, proposals): 38/50

Below threshold = revise internally. At or above threshold = proceed to human review.

Before/After Reference

Before: "Here's the thing: building products is hard. Not because the technology is complex. Because people are complex. Let that sink in." After: "Building products is hard. Technology is manageable. People aren't."

Before: "In today's fast-paced landscape, we need to lean into discomfort and navigate uncertainty with clarity. This matters because your competition isn't waiting." After: "Move faster. Your competition is."

Before: "Speed. Quality. Cost. You can only pick two. That's it. That's the tradeoff." After: "Speed, quality, cost — pick two."

Before: "What if I told you that the best teams don't optimize for productivity? Here's what I mean: they optimize for learning. Think about it." After: "The best teams optimize for learning, not productivity."

Before: "It turns out that most teams struggle with alignment. The uncomfortable truth is that nobody wants to admit they're confused. And that's okay." After: "Teams struggle with alignment. Nobody admits confusion."

---

Integration

Optional Project Brand Foundation

After stop-slop scoring, a project may optionally verify against a local brand foundation, for example charters/<project>-brand-foundation.md:

• No vocabulary from "never use" list
• Sentence structure matches Eric's voice patterns
• Side-by-side test: does this read as Eric or as AI?

Stop-slop clears AI-output quality patterns. Brand checks are project-local overlays, not ADLC core requirements.

Configuration

Optional .stop-slop.yml in project root:

code_slop:
  max_function_sloc: 50
  duplicate_threshold: 0.80
  exclude:
    - "**/migrations/**"
    - "**/generated/**"
    - "**/vendor/**"

content_slop:
  threshold_general: 35
  threshold_outreach: 38
  exclude:
    - "**/CHANGELOG.md"
    - "**/LICENSE*"

CLI Interface

# Build Brief contract check, implemented in this repo
bin/adlc slop-gate --build-brief ./build-brief.json --json

# Optional project-provided Mode 1: Code slop check
stop-slop code --input ./src/ --output ./code-report.json

# Mode 2: Content slop score
stop-slop content --input ./draft.md --output ./score.json

# Content slop score and revise (up to 2 passes)
stop-slop content --input ./draft.md --revise --passes 2 --output ./revised.md

# Outreach mode (higher threshold: 38/50)
stop-slop content --input ./outreach.md --mode outreach

# Optional project-provided full check
your-stop-slop-runner all --commit HEAD