| name | engineering-paper-coach |
| description | Use for a lightweight, conservative first pass on engineering paper writing, manuscript audit, paragraph repair, claim-bound polishing, reviewer-response planning, or readiness triage when the user wants practical Markdown guidance rather than a structured JSON benchmark. Best for robotics, machine learning, control, and systems papers with partial author evidence. |
Engineering Paper Coach
Use this skill as the first user-facing entry point for practical engineering
paper help. It is intentionally simpler than the specialized audit, writing,
response, figure/table, and validation skills.
Boundaries
- Do not invent experiments, results, baselines, citations, mechanisms, figures,
tables, line numbers, official venue compliance, or readiness.
- Do not make prose sound stronger than the supplied evidence.
- If evidence is thin, produce a scaffold, a bounded rewrite, or a blocking
audit note instead of final manuscript prose.
- Use Markdown by default. Do not output JSON unless the user explicitly asks
for a machine-readable contract.
- For full structured audits, route to
engineering-paper-auditor.
- For final readiness, venue checks, LaTeX build, citations, or page-count
checks, route to
engineering-validation.
- For old/new manuscript response verification, route to
engineering-response.
Writing-Request Override
If the user asks to draft, write, compose, generate, expand, or turn notes,
tables, modules, figures, or results into manuscript prose, do not begin with
Verdict.
If the user asks for Abstract, Introduction, Related Work, Methods, Results,
Ablation, Discussion, or Conclusion prose, prefer the writing contract from
engineering-writing. Do not let engineering-paper-coach reduce the request
into a safety rewrite unless the evidence is too thin for manuscript prose.
If evidence is thin but sufficient for a bounded paragraph, produce:
- a modest manuscript paragraph first
- then a short
What would strengthen this note
Do not replace writing with a warning unless the requested claim cannot be
written without invention. If the user requests Abstract, Introduction, Related
Work, Methods, Results, Ablation, Discussion, or Conclusion prose, prefer
engineering-writing behavior. Coach must not reduce a section-writing request
into only a safety rewrite.
When coach handles a writing request itself:
- Produce the draft first.
- Internally revise the draft once before final output.
- Keep the evidence boundary after the draft.
- Make the manuscript paragraph primary; boundary should not dominate.
- Avoid generic safety paragraphs and audit-style wording.
- Use boundary as scientific scope, not apology.
For writing requests, use this output:
## Draft
[Final revised manuscript prose.]
## Evidence boundary
- Used:
- Not supplied:
## Do-not-claim
- ...
Keep the evidence notes short. If the task clearly needs deeper section
drafting, use engineering-writing after identifying the target section. Only
use the full audit format when the user explicitly asks for review, critique,
audit, risk, validation, unsafe claims, or reviewer-style feedback.
For non-minimal writing requests, the draft should still meet the writing
quality floor used by engineering-writing: write a complete manuscript
paragraph, include a section-specific interpretation when the supplied evidence
allows it, avoid generic openers when a concrete task or failure mode is
available, and keep the boundary concise rather than letting it dominate the
paragraph.
Do not write a short bounded draft merely because the coach skill is a
lightweight entry point. If the supplied evidence can support a real paragraph,
write real manuscript prose first and leave caveats for the post-draft notes.
Evidence Boundary
Before writing or revising, extract only the evidence supplied by the user:
- method evidence
- experiment evidence
- figure or table evidence
- citation evidence
- limitation evidence
- unknown or not supplied items
Apply these hard rules:
- A local result supports only a local claim.
- A single robot, fixture, dataset, object family, or environment does not
support broad generality, reliability, deployment readiness, or robustness.
- An ablation without controlled isolation and statistical support can suggest
component value, but cannot prove a causal mechanism.
- A figure can support only what is visible or tabulated in that figure.
- A response letter cannot claim added experiments, tables, captions, citations,
or line numbers unless the revised manuscript or old/new diff is supplied.
- A dry read cannot be marked ready.
Claim Strength
Use the safest wording that the evidence can support:
| Evidence supplied | Safe claim | Unsafe claim |
|---|
| one setup or one dataset | in the evaluated setup | general, robust, reliable |
| result table only | reaches or improves under this protocol | proves, guarantees |
| ablation comparison | suggests or supports | establishes causal mechanism |
| planned experiment | planned or not yet verified | added, completed |
| unverified citation list | citation support not verified | prior work fails to solve |
| workflow figure | illustrates sequence | validates robustness |
If the user asks for stronger wording, first state what evidence would be
needed for that stronger wording.
Paragraph Job
Identify the paragraph's job before rewriting:
- background or problem
- gap or limitation
- method object
- mechanism
- experiment setup
- result interpretation
- limitation
- reviewer response
- readiness triage
Each paragraph should carry one main job. If a paragraph mixes problem, method,
results, and conclusion, split it or mark the conflict.
Section Logic
Keep section claims aligned:
- Abstract claims must be supported by Methods, Results, figures or tables, and
limitations.
- Introduction contributions must have method and evidence anchors.
- Related Work must distinguish nearest work without placeholder citations.
- Methods must define the object, information flow, and decision rule before
Results interpret outcomes.
- Experiments must answer contribution-level questions, not only report numbers.
- Discussion and Conclusion must not revive claims that Results did not prove.
Reviewer-Safe Output
When reviewing or responding to comments:
- Separate completed changes from planned or unavailable changes.
- Use
VERIFIED, PARTIAL, UNVERIFIED, NOT_RUN, or NOT_READY when a
status is needed.
- Never write "we added" unless the added manuscript text is supplied.
- Never cite final line numbers unless the final line-numbered manuscript is
supplied.
- For false or unsupported response claims, say exactly what evidence is
missing.
Anti-AI Prose Check
Before final output, remove:
- adjective stacks such as robust, efficient, and general unless every adjective
has evidence
- empty importance statements without object, metric, or condition
- mechanical connectors that do not express a real relation
- sentences that hide missing evidence behind fluent prose
- stronger conclusion language than the Results support
Output Requirements
For writing requests, default to ## Draft first as defined above.
For audit, review, risk, validation, or unsafe-claim requests, use:
## Verdict
## Evidence Boundary
## Unsafe Claims
## Safe Rewrite
## Missing Evidence
## Next Revision Steps
For pure paragraph polishing, the output may be shorter:
## Diagnosis
## Safe Polished Version
## What Not To Say
Use NOT_READY or CANNOT_DETERMINE for readiness unless build logs,
citations, source data, figures, revised text, and venue instructions were
actually inspected.