More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Auto Paper Improvement Loop: Review, Fix, Recompile

    1

    Improve a LaTeX manuscript through controlled review-fix-recompile cycles: freeze a baseline, get an independent review, apply bounded fixes, diff-review the changes, and recompile.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Auto Paper Improvement Loop: Review, Fix, Recompile

    Auto Paper Improvement Loop: Review, Fix, Recompile

    Example session with this skill installed

    Run two improvement rounds on a 12-page ML methods draft. The build is green, results numbers are frozen, prose and figures are editable. Round 1 review found 2 blockers (a table number mismatch, 0.78 in text vs 0.72 in the table; the abstract claims "outperforms all baselines" but one baseline is missing), 5 majors, 4 minors.

    • Read your context and instructions
    • Compiled the auto paper improvement

    LOOP COMPLETE — 2 rounds, blockers 2 → 0, majors 5 → 2, build green throughout, draft handed to a human reader.

    Setup: baseline snapshotted as draft-v1/, build confirmed green from clean. Fix budget written: results numbers frozen; prose, figures, and related work editable.

    Round 1 — review (fresh context, no knowledge of any prior discussion): 11 findings, 2 blockers, 5 majors, 4 minors. Blocker 1: text says 0.78, the table says 0.72. Blocker 2: the abstract claims "outperforms all baselines" but one baseline is missing from the comparison table. Majors include a muddled methods paragraph and two unsupported "demonstrates" claims. Minors are typos and a mislabeled axis.

    Triage: both blockers fixed. Four of five majors fixed; the fifth (add a related-work subsection) backlogged — it would add a section and the budget says no new sections without a blocker demanding it. Two minors fixed (one-line); two backlogged.

    Fixes: Blocker 1 — text corrected to 0.72 (the table was right; the prose was stale). Blocker 2 — abstract narrowed to name the baselines actually compared. Change log records every finding number, change, and file/line.

    Diff review + recompile: the diff shows only the triaged changes — no surprise edits. Recompile from clean: green, no new warnings. Snapshot as draft-v2/.

    Round 2 — independent review (fresh context, no knowledge of round 1's fix list): 6 findings, 0 blockers, 2 majors, 4 minors. Blockers at zero, majors shrinking — the loop is working. Triage and fix, snapshot draft-v3/, recompile green.

    Loop ends here: blockers zero, majors nearly gone, budget spent. The draft goes to a human reader with the change logs and finding counts attached. No new claims introduced, no numbers changed except the one a finding unfroze.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Unstructured "revise this paper" requests produce drift: each round rewrites different things, introduces new errors, and nobody can say what changed. The loop fixes that with three controls: a frozen baseline, an independent review, and bounded fixes with diff review.

    Setup freezes the baseline first: copy the manuscript to draft-v1/ (or tag it in git), confirm the build is green, and write down the fix budget — which sections are editable and which are frozen. Each round then has exactly four steps. Step 1, independent review: a fresh reviewer (a fresh model instance or a colleague) reads the draft cold with a fixed rubric — correctness (do numbers in text match tables and figures?), claims (is every "shows" or "proves" supported?), structure, and clarity — returning a numbered findings list with severity and location, capped at about 15. Step 2, triage: you decide what gets fixed — all blockers, most majors, minors only if one-line; the rest goes on a backlog with a reason. Step 3, bounded fixes: only the triaged findings, one at a time, no new claims, no new sections unless a blocker demands it, with a change log (finding → change → file/line). Step 4, diff review and recompile: read the diff, not the full paper, then recompile from clean — new warnings are regressions.

    Between rounds, snapshot the output and compare finding counts: blockers should go to zero, majors should shrink. Two to three rounds is the normal budget; if round 3 has as many blockers as round 1, the problem is the draft's structure — stop looping and restructure.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse All Claude Skills.

    • It improves an existing draft; it cannot write a first draft or fix results that are still changing.
    • More than ~15 findings in a round means the draft needs restructuring, not more rounds.
    • Two to three rounds is the normal budget; looping past that has diminishing returns and rising drift risk.
    • A red build before round one stops the loop — the build must be fixed first.
    • It cannot replace a human reader at the end; the loop hands the draft over, it doesn't ship it.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    A reviewer who saw the previous fix list will grade its own homework — findings shrink for the wrong reason. Fresh context every round keeps the review honest.

    The diff is the review surface. Reading the full paper hides what changed; the diff shows exactly which lines moved and whether each fix broke its neighbors.

    Stop looping and restructure. The problem is the draft's structure, not its prose — more rounds will only add drift. Fix the structure, then loop again if needed.