High Quality Content Writer

High Quality Content Writer

Use this skill as a closed-loop content system: draft or clean text, score it with deterministic gates, evaluate it in isolation, then retry with concrete fix instructions until quality passes or the retry budget is exhausted.

Operating Model

Default to this loop:

1. Prepare an isolated workdir under .tmp/high-quality-content-writer/<timestamp>/.
2. For websites, always separate scoring from extraction:
   • For trope scoring, run slop_score.py --url <url> first so the result is calibrated against the public AI Vetter action when available.
   • For token-efficient content review/rewrite, extract compact markdown through Jina Reader and save it as source.md:
   • curl -L -s 'https://r.jina.ai/http://https://example.com/path' -o source.md
   • If Jina fails or returns an error page, fall back to direct fetch/HTML extraction.
3. Save pasted text or extracted markdown as source.md and the current candidate as candidate.md.
4. Run deterministic scoring:
   • scripts/slop_score.py --file candidate.md --pretty
   • scripts/quality_gate.py --file candidate.md --pretty
5. If available, run an isolated evaluator subagent using references/evaluator-flow.md. Give it only the candidate text, scoring JSON, user brief, and the skill path.
6. If any gate fails, run an isolated writer subagent using references/writer-flow.md. Give it source.md, candidate.md, previous findings, and fix instructions. Ask it to write the next candidate.md.
7. Repeat until pass or max_attempts is reached. Default max_attempts = 4.
8. Return the final text first, then a compact QA report with scores, attempts, remaining risks, and what changed.

Do not stop after analysis when the user asked for final content. The expected output is improved content plus evidence that it passed the gates.

Gates

Use three complementary gates:

• Trope score: slop_score.py catches tropes.fyi-style AI-writing patterns. Pass target: score <= 25 and verdict Human or AI-assisted; Suspicious is allowed only if the user wants light editing.
• Slop-check score: quality_gate.py implements the bundled quote/article slop-check criteria. Pass target for articles: clean or good; acceptable is allowed only with explanation; probable_slop and obvious_slop hard fail.
• Quality evaluator score: isolated evaluator scores 1-10 on clarity, specificity, fact density, voice, slop risk, and brief adherence. Pass target: >= 8.

Hard fail if the draft invents facts, drops required claims, adds unsupported numbers, or changes the user's intended meaning.

Common Commands

python3 scripts/slop_score.py --file candidate.md --pretty
python3 scripts/quality_gate.py --file candidate.md --pretty

URL scoring remains available:

python3 scripts/slop_score.py --url https://example.com/article --pretty

For website content checks, prefer explicit Jina extraction when the user wants token-efficient review:

python3 scripts/slop_score.py --url https://example.com/article --pretty
curl -L -s 'https://r.jina.ai/http://https://example.com/article' -o .tmp/high-quality-content-writer/source.md
python3 scripts/quality_gate.py --file .tmp/high-quality-content-writer/source.md --type article --pretty

Output Contract

For a grading-only request:

Verdict: <clean/good/acceptable/probable_slop/obvious_slop + trope verdict>
Score: <quality score>/10
Trope score: <0-100>
Slop-check score: <points>

Findings:
- <issue>: <evidence>

Fix instructions:
- <sentence-level or structural instruction>

For a write/fix request:

<final improved content>

QA:
- Attempts: <n>
- Quality gate: <pass/fail>, <score>/10
- Trope score: <score>, <verdict>
- Slop-check: <score>, <verdict>
- Main fixes: <short list>
- Residual risk: <only if any>

References

• references/tropes.md: full uncut tropes.fyi-style rubric supplied by the user.
• references/slop-check.md: short quote criteria imported from $slop-check.
• references/slop-check-full-criteria.md: full 24-criteria article rubric imported from $slop-check.
• references/c8c-loop-analysis.md: distilled mechanics from content-improve-loop.chain and c8c evaluator implementation.
• references/writer-flow.md: isolated writer subagent prompt.
• references/evaluator-flow.md: isolated evaluator subagent prompt.
• references/api-example.md: public Vetter action id, request shape, and response shape.