Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
If labels are missing, find corresponding figure/table code in figures/*.tex and embed into the correct section file.
Also check compile_utils.sh output for "UNEMBEDDED" warnings — each one means a figure or table from figures/ is not in any section.
MANDATORY FIX LOOP — do NOT proceed to compilation until all figures AND tables are embedded:
UNEMBED=0
for pdf in figures/*.pdf; do
[ -f "$pdf" ] || continue
bn=$(basename"$pdf")
grep -rq "$bn" paper/sections/*.tex paper/main.tex 2>/dev/null || { echo"UNEMBEDDED PDF: $bn"; UNEMBED=$((UNEMBED+1)); }
donefor tbl in figures/TABLE_*.tex; do
[ -f "$tbl" ] || continuefor lbl in $(grep -oh '\\label{[^}]*}'"$tbl" 2>/dev/null); do
grep -rq "$lbl" paper/sections/*.tex paper/main.tex 2>/dev/null || { echo"UNEMBEDDED TABLE: $lbl"; UNEMBED=$((UNEMBED+1)); }
donedoneecho"Total unembedded: $UNEMBED"
If UNEMBED > 0, fix ALL before compiling. For each unembedded item, copy the figure/table block into the appropriate section with lead-in text + analysis. Re-run until UNEMBED = 0.
Step 3: Compile (manual steps, no latexmk)
cd paper/
rm -f main.aux main.blg main.log main.out main.toc 2>/dev/null
pdflatex -interaction=nonstopmode main.tex 2>&1 | tee compile_pass1.log
bibtex main 2>&1 | tee bibtex.log
pdflatex -interaction=nonstopmode main.tex 2>&1 | tee compile_pass2.log
pdflatex -interaction=nonstopmode main.tex 2>&1 | tee compile.log
[ -f main.pdf ] && echo"main.pdf $(wc -c < main.pdf) bytes" || echo"PDF not generated"
Step 4: Error diagnosis and fix loop (MANDATORY)
After each compilation, check main.log for CRITICAL errors. You MUST fix ALL errors before declaring compilation complete.
MATH_ERR=$(grep -c 'Bad math environment delimiter\|Missing \$ inserted\|begin{document} ended by' paper/main.log 2>/dev/null || echo 0)
LR_ERR=$(grep -c 'Not allowed in LR mode' paper/main.log 2>/dev/null || echo 0)
echo"Math errors: $MATH_ERR, LR mode errors: $LR_ERR"
[ $((MATH_ERR + LR_ERR)) -gt 0 ] && grep -B2 'Bad math\|Missing \$ inserted\|Not allowed in LR mode' paper/main.log | grep -E '^\./|^l\.' | head -20
Iterate up to MAX_COMPILE_ATTEMPTS times, each with full 4-step compilation. For each error:
Math errors: read the error location from main.log, open the file, fix broken $...$ delimiters individually. Do NOT use broad sed patterns.
LR mode errors: add \par or blank line before float environments.
Missing packages: install them.
BibTeX failures: fix LaTeX errors first (BibTeX fails when LaTeX errors exist upstream), then recompile.
After each fix, recompile and recheck. ⛔ Do NOT proceed until MATH_ERR = 0 and LR_ERR = 0.
When fixing errors in main.tex, only fix the specific error. Do not rewrite or restructure main.tex — the template's preamble, page margins, and formatting settings must remain unchanged.
Body pages = Introduction through Conclusion, excluding references and appendix.
Body pages must be ≥ MAX_PAGES. If insufficient, return to paper-write to expand content.
"
import json, re, os
try:
with open('figures/all_results.json','r',encoding='utf-8') as f: results=json.load(f)
except Exception:
print('⚠ all_results.json unreadable'); exit(0)
def extract(obj, p=''):
n = {}
if isinstance(obj, dict):
for k, v in obj.items(): n.update(extract(v, f'{p}.{k}'))
elif isinstance(obj, list):
for i, v in enumerate(obj): n.update(extract(v, f'{p}[{i}]'))
elif isinstance(obj, (int, float)) and not isinstance(obj, bool):
if 0.001 < abs(obj) < 1e10: n[p] = obj
return n
jn = extract(results)
pn = set()
section_dir = 'paper/sections' if os.path.isdir('paper/sections') else 'paper'
for tf in sorted(os.listdir(section_dir)):
if not tf.endswith('.tex'): continue
try:
with open(f'{section_dir}/{tf}','r',encoding='utf-8',errors='ignore') as f: t=f.read()
except: continue
for m in re.finditer(r'(?<![a-zA-Z])(\d+\.?\d+)(?![a-zA-Z_{}])', t):
try: pn.add(float(m.group(1)))
except: pass
key_patterns = ['accuracy','acc','rmse','mae','r2','f1','auc','loss','precision','recall','bleu','rouge','perplexity','ppl','speed','latency','throughput','map','ndcg']
key_values = {k: v for k, v in jn.items() if any(w in k.lower() for w in key_patterns)}
miss = sum(1 for k, v in key_values.items() if not any(abs(p - v) < abs(v) * 0.01 + 0.001 for p in pn))
total = len(key_values)
if total == 0:
print(' (no key metrics in JSON to check)')
elif miss > 3:
print(f'❌ {miss}/{total} key values not found in paper — check numerical consistency'); exit(1)
else:
print(f'✅ Numerical consistency: {total-miss}/{total} key values present in paper')
"
1
else
echo
" (no all_results.json to check against)"
fi
# Unrealistic values check (detect fabricated or overfitted values)
echo
"--- Unrealistic values check ---"
if
then
"
import json
with open('figures/all_results.json','r',encoding='utf-8') as f: data=json.load(f)
suspicious = []
def check(name, val):
if not isinstance(val,(int,float)) or isinstance(val,bool): return
key = name.lower()
if any(w in key for w in ['r2','r_squared','accuracy','acc','precision','recall','f1','auc']) and val > 0.999:
suspicious.append(f'{name}={val:.4f} too perfect (>0.999), likely overfitted or leaked')
if any(w in key for w in ['rmse','mae','mse','loss']) and val == 0:
suspicious.append(f'{name}=0 perfect error, nearly impossible')
if ('p_value' in key or 'pvalue' in key) and val == 0:
suspicious.append(f'{name}=0 perfect significance')
if any(w in key for w in ['improvement','speedup','gain']) and val > 10:
suspicious.append(f'{name}={val} (+{val*100:.0f}%) unrealistically large')
def walk(obj, path=''):
if isinstance(obj,dict):
for k,v in obj.items(): walk(v, f'{path}.{k}')
elif isinstance(obj,list):
for i,v in enumerate(obj): walk(v, f'{path}[{i}]')
else: check(path, obj)
walk(data)
if suspicious:
print(f'🚩 {len(suspicious)} suspiciously perfect values (check for data leakage or fabrication):')
for s in suspicious[:5]: print(f' {s}')
print(' If values are genuinely valid, discuss why in the paper')
else:
print('✅ No suspicious values')
"
"❌ $META meta content leaks — remove references to internal files"
1
# Overclaiming
echo
"--- Overclaiming ---"
for
in
do
"$f"
continue
for
in
"首次提出"
"首次发现"
"完美"
"无可比拟"
"前所未有"
"开创性"
"revolutionary"
"unprecedented"
"groundbreaking"
do
"$w"
"$f"
echo
done
done
"$OC"
echo
"✅ No overclaiming"
echo
"⚠ $OC overclaiming instances"
# Claims-Evidence backfill (does paper cover all planned claims?)
echo
"--- Claims coverage ---"
if
then
"
import re, os
try: plan = open('PAPER_PLAN.md','r',encoding='utf-8').read()
except: exit(0)
rows = re.findall(r'\|\s*([^|]+?)\s*\|\s*([^|]+?)\s*\|', plan)
claims = [c.strip() for c, e in rows if c.strip() not in ('Claim','---') and '---' not in c and len(c.strip()) > 8]
if not claims:
print(' (no claims-evidence matrix in PAPER_PLAN.md)'); exit(0)
# 读取所有论文章节
section_dir = 'paper/sections' if os.path.isdir('paper/sections') else 'paper'
paper_text = ''
for tf in sorted(os.listdir(section_dir)):
if tf.endswith('.tex'):
try: paper_text += open(f'{section_dir}/{tf}','r',encoding='utf-8',errors='ignore').read()
except: pass
# 对每个 claim，提取关键词，在论文里找
missing = []
for c in claims[:15]:
keywords = [w for w in re.findall(r'[a-zA-Z_]{4,}|[\u4e00-\u9fff]{2,}', c) if len(w) > 3][:3]
if keywords and not any(kw.lower() in paper_text.lower() for kw in keywords):
missing.append(c[:50])
if missing:
print(f'❌ {len(missing)}/{len(claims)} planned claims not covered in paper:')
for m in missing[:5]: print(f' - {m}')
exit(1)
else:
print(f'✅ All {len(claims)} planned claims covered in paper')
"
