CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

auto-paper-improvement-loop

Autonomously improve a generated paper via GPT-6-Astra xhigh review → implement fixes → recompile, for 2 rounds. Use when user says "改论文", "improve paper", "论文润色循环", "auto improve", or wants to iteratively polish a generated paper.

Quality

76%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./skills/auto-paper-improvement-loop/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

56%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

An exceptionally concrete, validation-heavy orchestration workflow whose execution guidance is strong, but it is significantly overlong for a SKILL.md body. The duplicated reviewer prompt, five restatements of the same bias-guard invariant, and a fully inlined whitelist specification inflate token cost, and the absent bundle means none of it is progressively disclosed.

Suggestions

Replace the Step 5 review prompt with a reference to the Step 2 prompt plus only the Round-2 deltas (fresh zero-context instruction and the 'ignore any prior review rounds' lines), saving ~40 duplicated lines.

Move the edit-whitelist specification (schema, resolution rules, glob semantics, forbidden-operation detectors, round behaviors) into a references/edit-whitelist.md and keep a short summary with the abort-on-load-failure rule plus a link in SKILL.md.

Fix the sequencing defects: deduplicate the two 'Step 9' headings, place Step 5b after Step 5.5 (or renumber), and implement the actual comparison loop in the Step 4.5 heredoc instead of the '# Compare normalized theorem blocks...' comment placeholder.

DimensionReasoningScore

Conciseness

The Step 2 and Step 5 reviewer prompts are ~45 lines of near-verbatim duplication, the reviewer-independence invariant is restated five times (header blockquote, Constants, dedicated protocol section, Step 5, Key Rules), and the ~100-line edit-whitelist spec is fully inlined. Not 1 because nothing explains concepts Claude already knows — the padding is skill-specific operational detail, but it is clearly padded rather than 'some tightening'.

2 / 5

Actionability

Gives exact bash commands, MCP invocation configs (model + reasoning effort), complete reviewer prompts, regex detectors for forbidden operations, and YAML/JSON schemas — mostly copy-paste ready. Not 5 because the Step 4.5 heredoc defines normalize() but leaves the actual main-vs-appendix comparison as a comment placeholder, and several steps ('Parse user response same as /auto-review-loop', reviewer-slot resolution) defer to material not present in the skill.

4 / 5

Workflow Clarity

Steps 0–9 are clearly sequenced with explicit validation checkpoints (0 undefined refs after recompile, restatement regression test, format-check stop criteria with location-aware hard blocks, whitelist rejection logging, NOT_APPLICABLE/BLOCKED fallbacks for kill-argument). Not 5 because there are two duplicate 'Step 9' headings (Document Results and Summary) and Step 5.5 is ordered before Step 5b, which muddies the sequence despite validation being thorough.

4 / 5

Progressive Disclosure

Section headers are clear and external references (../shared-references/*.md, tools/extract_paper_style.py, skills/kill-argument) are one level deep and clearly signaled, but no bundle files exist and every referenced helper lives outside the skill where it cannot be resolved from it, while ~100 lines of whitelist parameter documentation and two full review prompts are inlined that clearly belong in separate reference files. Not 4 because the amount of inlineable content in a 700-line body outweighs the good signaling.

3 / 5

Total

13

/

20

Passed

Description

91%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A model description: concrete what, explicit bilingual when-clause with enumerated trigger phrases, and a clearly delineated niche. The only gap is that it undersells a few secondary capabilities and could be marginally more distinct from sibling review skills.

DimensionReasoningScore

Specificity

"improve a generated paper via GPT-6-Astra xhigh review → implement fixes → recompile, for 2 rounds" names the domain and several concrete actions with model, effort level, and round count specified. Not 5 because it stays at loop-level actions and omits secondary capabilities the skill actually has (style-ref steering, edit whitelist, format compliance checking).

4 / 5

Completeness

It explicitly answers both questions: the 'what' is the concrete review → fix → recompile loop over 2 rounds, and the 'when' is an explicit 'Use when user says…' clause with concrete trigger phrases — matching the anchor-5 example structure exactly. Not 4 because the 'when' is as explicit as it gets.

5 / 5

Trigger Term Quality

"Use when user says \"改论文\", \"improve paper\", \"论文润色循环\", \"auto improve\", or wants to iteratively polish a generated paper" provides comprehensive natural-term coverage with synonyms across two languages and both command-style and prose triggers. Anti-drift check against 4 ('a few natural terms missing'): nothing significant is missing for this domain.

5 / 5

Distinctiveness Conflict Risk

"generated paper", "GPT-6-Astra xhigh review", and "论文润色循环" carve a clear niche with distinct triggers. Not 5 because generic triggers like "improve paper" carry minor overlap risk with general paper-review or polish skills — the body distinguishes /auto-review-loop but the description does not.

4 / 5

Total

18

/

20

Passed

Validation

75%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 12 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

skill_md_line_count

SKILL.md is long (714 lines); consider splitting into references/ and linking

Warning

allowed_tools_field

'allowed-tools' contains unusual tool name(s)

Warning

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

relative_links

Relative link issues: 2 suspicious

Warning

Total

12

/

16

Passed

Repository

wanshuiyin/Auto-claude-code-research-in-sleep

Path

skills/auto-paper-improvement-loop/SKILL.md

Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.