Turn engineering notes, result tables, ablation rows, and Chinese notes into bounded manuscript prose.
Engineering Paper Skills is a Codex skill suite for engineering paper writing, manuscript audit, reviewer response, figure/table claim checks, and validation.
What This Is
This beta skill suite is designed for robotics, machine learning, control, systems, and other experiment-heavy engineering papers where claims must stay tied to concrete methods, results, figures, tables, citations, and limitations.
It is conservative by default. It should expose missing evidence, downgrade unsupported claims, and avoid inventing experiments, citations, metrics, tables, line numbers, or venue readiness.
Normal users only need skills/_shared and skills/engineering-*. Repository scripts, tests, and evals are maintainer evidence.
Start With Rough Notes
Most users should begin by giving the skill the section target, task bottleneck, method object, result evidence, diagnostic axes, and boundary. Ask for manuscript prose first and state forbidden claims.
Evidence note: examples are recorded local Codex outputs; detailed provenance is in the Evidence Archive.
Results table with diagnostic axes
Input skeleton:
84% over 180 trials; baselines 69%, 72%, 61%.
Diagnostics: overhead occlusion, side-view depth
uncertainty, open-loop contact error.
Short excerpt from recorded output:
In contact-rich peg insertion, the best-performing
policy is the one that prevents visual ambiguity
from becoming unrecovered contact error.Ablation rows to contribution prose
Input skeleton:
17%, 23%, 44%, 76%, 88% additive rows.
Recover component roles from deltas.
Short excerpt from recorded output:
The ablation indicates that the main bottleneck in
contact-rich action grounding is not merely executing
a hinted motion, but verifying that a visually
plausible hint is grounded to the intended target.Chinese notes to manuscript prose
Input skeleton:
遮挡下 peg-hole 插入; fixed cameras lose cues;
open-loop carries pose error; full system 84% vs
69%/72%/61%; bounded evaluation.
Short excerpt from recorded output:
Occlusion during the final approach, rather than
initial visual recognition, is the central failure
mode in peg-hole insertion under fixed camera views.Methods reader path
Input skeleton:
overhead RGB-D -> pose/confidence -> optional
additional view -> refined pose -> guarded
insertion with force and pose-deviation stops.
Short excerpt from recorded output:
Because pose uncertainty that survives into contact
is difficult to correct safely, the method first
resolves the insertion pose in perception...Recorded local outputs are strong public-beta evidence and top-tier candidate evidence, not top-tier-ready proof. See the Evidence Archive for provenance.
How To Get Strong Prose
- Give the section target and ask for manuscript prose first.
- Name the task bottleneck and the method object, not only module names.
- Provide the result table, diagnostic axes, or failure modes.
- State the boundary and forbidden claims before the draft request.
Skills
Coach
Provides a lightweight conservative Markdown first pass for writing, audit, polishing, response planning, and readiness triage.
Router
Routes mixed paper tasks to the right audit, writing, figure, response, or validation workflow.
Auditor
Checks claim-evidence gaps, story flow, paragraph transitions, section dependencies, and overclaims.
Writing
Drafts or restructures sections from author-provided evidence without expanding the facts.
Polishing
Improves English prose while preserving claim strength and evidence boundaries.
Figure/Table
Audits whether panels, tables, captions, and result prose support visible or tabulated evidence.
Response
Turns reviewer or advisor comments into traceable revision plans without false completion claims.
Validation
Reports readiness using explicit PASS, FAIL, PARTIAL, NOT_RUN, and UNKNOWN states.
Example Outputs
Start with the draft-first writing examples if you want to see manuscript prose. Practical examples and recorded local Codex outputs are in the demo gallery. Structured artifacts and model-run records are kept for reviewers and maintainers.
- Demo gallery
- Recorded showcase v2 results output
- Recorded showcase v2 ablation output
- Recorded Chinese-notes rerun output
- Recorded Methods showcase v2 output
- Recorded showcase v2 rerun manifest
- Recorded showcase v2 rerun2 manifest
- Recorded rough-user evidence manifest
- Recorded realistic-section evidence manifest
- Methods reader-path draft
- Results interpretation draft
- Ablation interpretation draft
- Full-paper realistic audit
- Structured JSON audit
- Sentence-level claim and evidence audit
- AI-smell polishing example
- Figure/table panel-claim audit
- Recorded local model-run artifact
Install
git clone https://github.com/169884902hzl/engineering-paper-skills.git
cd engineering-paper-skills
cp -a skills/_shared skills/engineering-* ~/.codex/skills/Maintainer / Reviewer Evidence
Normal users only need skills/_shared and skills/engineering-*. The repository also includes maintainer QA scripts, golden-output snapshots, response-diff fixtures, venue profiles, a manual rubric, and recorded local model-run artifacts.
These are review evidence, not automatic certification of top-tier paper quality. Behavior regression requires a configured model command template and scored model outputs.
For no-release version tracking, use the known-good commit matrix and changelog.