Engineering writing

Engineering Writing

Use this skill for English manuscript creation and structural revision. It is for writing the paper's argument, not merely polishing sentences.

Core Stance

• Final manuscript prose defaults to English.
• Non-English notes are source material only; do not preserve them as the final manuscript language unless the user explicitly asks.
• Evidence comes first. Do not invent methods, datasets, experiments, metrics, numbers, references, mechanisms, novelty, limitations, or conclusions.
• Write the argument before writing sentences.
• Every claim must have a visible route to method detail, figure/table evidence, experiment evidence, or an explicit boundary.
• If evidence is missing, expose the gap or write a scaffold with placeholders.
• Do not finalize Abstract or Conclusion claims before Methods/Experiments evidence and boundaries are available; produce a scaffold instead.
• For an existing paper, read the current live draft and relevant files before rewriting.

Boundaries

• Use engineering-polishing for pure language polish after the section logic is already stable.
• Use engineering-figure-table for captions, table design, visual roles, and figure/table consistency.
• Use engineering-response for reviewer, editor, advisor, or senior-author comments.
• Use engineering-validation for build checks, readiness claims, citation checks, and final submission QA.

Request Modes

Classify the user request before choosing an output shape.

WRITE

Use this mode when the user asks to draft, write, compose, generate, or turn notes, modules, figures, tables, results, or outlines into manuscript prose. Examples include abstract drafting, Introduction opening, Related Work positioning, Methods overview, Results paragraph, ablation interpretation, robustness paragraph, Discussion, or Conclusion.

For WRITE requests, output real manuscript prose first. Put structure, evidence boundary, and safety notes after the draft. Do not begin with Verdict, Story spine, Source-note triage, or a full claim-evidence matrix.

WRITE Quality Floor

For non-minimal WRITE requests, the draft must be a complete manuscript paragraph, not a compressed safety note or a list converted into sentences. Evidence boundaries still matter, but the boundary sentence must not become the whole paragraph.

Showcase-Grade Writing Procedure

Use this procedure for showcase-grade, public-demo, or non-minimal WRITE requests. The final answer should show the revised manuscript prose first, not the internal notes.

1. Extract argument: identify the task, bottleneck, method object, evidence, and boundary.
2. Draft manuscript prose using section-specific logic.
3. Self-revise once before final output to remove generic frames, improve mechanism interpretation, and integrate the boundary naturally.

For showcase-grade or non-minimal WRITE requests:

• Do not output the first safe draft directly. Internally revise it once before final output.
• Prefer concrete failure modes over generic openings.
• Prefer mechanism-specific verbs over generic improve, address, or support when the supplied evidence allows a more precise verb.
• Use boundary as scientific scope, not apology.
• Do not write without claiming... in manuscript prose unless the user asks for audit wording.
• Do not let the final sentence become a disclaimer.
• For Results, if diagnostic notes are supplied, use them to explain why the ranking occurs.
• For Ablation, recover contribution roles from major deltas.
• For Methods, explain why the reader path is ordered that way.
• For Related Work, write technical-axis positioning, not meta-commentary.
• For Chinese notes, do not translate sentence by sentence; reconstruct the English paper argument.

Silent self-edit checklist before final WRITE output:

• Is the first sentence concrete?
• Does the paragraph have a section job?
• Is there at least one mechanism or failure-mode interpretation when supplied evidence permits it?
• Does every number come from supplied evidence?
• Is the boundary written as scope, not apology?
• Are generic frames removed?
• Is there any meta-commentary that should be rewritten as manuscript prose?

For showcase or public-demo WRITE requests, one paragraph should normally be 120-220 words unless the user asks for a shorter answer. Drafts below 80 words are acceptable only for smoke checks, captions, or explicitly minimal prompts.

A showcase-grade paragraph should normally contain at least four sentence roles:

1. setup, question, or bottleneck
2. method, comparison object, or technical axis
3. evidence, transformation, or result
4. interpretation or component role
5. boundary, operating scope, or do-not-claim

• Abstract: name the concrete failure mode, method core, evaluation boundary, and strongest supplied evidence.
• Introduction: move from task difficulty to core bottleneck, why existing routes are insufficient, and why the proposed formulation is motivated.
• Related Work: write by technical axes and nearest-neighbor distinction; do not write paper-by-paper summaries unless the user asks for notes.
• Methods: give a reader path: input, intermediate object, transformation, handoff, execution or safety boundary. Do not write a module directory.
• Results: write ranking, key number, mechanism-level interpretation, and operating boundary. Do not only read table cells.
• Ablation: tie each major delta to a component role or failure mode, then state what the ablation does not prove.
• Robustness/failure: name the stress axis, trend, hardest regime, and current operating boundary.
• Conclusion: recover method, strongest evidence, bounded takeaway, and future work derived from the failure regime.

Public Recorded Demo Blockers

Public recorded writing demos must be rejected or rerun if manuscript prose contains any of the following:

• an unsupported physical mechanism that is not present in the source notes, such as lateral jamming, jamming, slip, compliance, deformation, fatigue, binding, impact, resonance, collision, vibration, backlash, wear, fracture, buckling, instability, drift, or physical failure statistics
• prompt or process meta-language inside manuscript prose, such as supplied notes, supplied positioning, supplied failure modes, available evidence, dominant failure source, pending verification, pending verified citations, unsupported, not verified, Unsupported / not verified, Evidence Needed, or without claiming
• a final sentence that reads like an audit disclaimer rather than scientific scope
• Related Work prose that discusses citation verification inside the manuscript paragraph
• Conclusion prose whose second paragraph starts as a limitation inventory instead of a concrete operating boundary

Manuscript prose may use scope-aware wording, but not audit wording. Evidence boundary notes belong after the draft unless they can be phrased as scientific scope. If a mechanism is plausible but not supplied, do not name it. Do not write without claiming... in manuscript prose. Do not write supplied notes, supplied positioning, supplied failure modes, available evidence, dominant failure source, pending verification, pending verified citations, unsupported, not verified, Unsupported / not verified, or Evidence Needed inside the Draft.

Avoid repeated generic openers such as X remains difficult, To address this limitation, and These results indicate when a section-specific opening can name the actual failure mode, experiment question, or component role.

The boundary should usually be one concise final sentence or clause. If the boundary or do-not-claim content is longer than the manuscript claim in WRITE mode, rewrite so evidence-based prose is primary and the boundary is concise.

PLAN

Use this mode when the user asks for an outline, paper plan, section plan, story map, contribution design, or writing order. In PLAN mode, it is appropriate to show the thesis, section job map, and contribution-evidence map before prose.

AUDIT

Use this mode when the user asks to review, diagnose, check, critique, find weaknesses, identify unsupported claims, or explain why a draft is not working. For full manuscript audits, prefer engineering-paper-auditor.

POLISH

Use this mode when the user asks for local wording, concision, sentence flow, or translation of already stable section logic. For pure polish, prefer engineering-polishing.

When to Open Extra Files

File	Open when
references/paper-workflow.md	Starting from scratch, rebuilding a draft, or deciding paper-writing order
references/contribution-evidence.md	Defining the paper thesis, contributions, claim boundaries, or evidence map
references/title.md	Drafting or auditing the title
references/abstract.md	Drafting or revising the abstract
references/introduction.md	Drafting Introduction, positioning, gap, formulation, or contribution list
references/related-work.md	Drafting Related Work, taxonomy, nearest-neighbor distinction, or folded literature positioning
references/methods.md	Writing Methods, overview, formulas, system roles, algorithms, execution, or safety/deployment logic
references/methods-worksheet.md	A Methods draft risks becoming a module directory or formula dump
references/experiments.md	Planning or writing Experiments/Results, baselines, metrics, main results, ablations, failure analysis
references/experiments-worksheet.md	Results need setup, baseline, metric, ablation, category, stress, or failure-envelope structure
references/discussion.md	Writing Discussion, limitations, interpretation, or implications
references/conclusion.md	Writing a bounded conclusion and future work
references/section-boundaries.md	Auditing whether content belongs in the right section or when a section is drifting
references/section-budget.md	Checking whether a section is too thin, too dense, or taking space from evidence
references/page-budget-war-plan.md	Cutting manuscript length without damaging evidence anchors
references/source-learning.md	Learning structure from 3-5 neighboring papers without copying wording or surface format
references/ral-style-writing-guide.md	Drafting engineering prose from notes using RAL-style section skeletons: Abstract, Introduction, Related Work, Methods reader path, Results interpretation, ablation, robustness, or Conclusion
references/positive-patterns.md	Choosing a paper-level or section-level pattern before drafting, or a draft reads structurally generic
references/bad-sentence-repairs.md	Repairing common bad manuscript sentences and section-level failure symptoms
manifest.yaml	Planning which references to load for section, input-state, or failure-repair tasks
references/examples.md	Needing concrete prompt and output behavior examples
references/failure-modes.md	Handling thin evidence, invented-citation requests, or overclaim pressure
references/venue-aware-writing.md	Target venue family may change abstract, limitation, reproducibility, or contribution framing
references/paper-level-narrative-map.md	Full-paper or multi-section writing needs story dependency checks
../_shared/story-spine.md	Building problem-to-evidence-to-boundary logic before drafting
../_shared/evidence-boundary.md	Any task asks for stronger claims, missing evidence, or manuscript facts
../_shared/citation-boundary.md	Related Work or citations are requested without provided sources
../_shared/citation-verification-workflow.md	The user supplies externally discovered citations (deep research, scholarly APIs) that need a verification gate before Related Work positioning
../_shared/claim-strength.md	Calibrating verbs, novelty, robustness, generalization, or causal language
../_shared/ai-assisted-writing-policy.md	AI-assisted prose or venue disclosure risk is relevant
../_shared/list-to-argument.md	Source material is a bullet list, module list, result-row list, or contribution list
../_shared/non-english-source-notes.md	Non-English notes must become English manuscript prose
../_shared/output-mode.md	The user asks for output only
../_shared/sentence-role-and-story-flow.md	Drafting, shortening, or reordering paragraphs where every sentence must justify its role
../_shared/terminology-ledger.md	A writing task may rename methods, metrics, categories, or baselines

Intake

Before drafting, identify:

• target section and venue
• paper type: robotics system, algorithm, method, benchmark, dataset, device, control, perception, learning, or experiment-heavy systems paper
• core object: system, task, method, dataset, mechanism, device, or phenomenon
• problem and gap
• proposed method or formulation
• evidence: figures, tables, metrics, comparisons, ablations, stress tests, qualitative evidence, or deployment evidence
• boundary: what is not shown
• current manuscript paths and page/word constraints

If core claim, evidence, or boundary is absent, state the gap before drafting. You may still provide a scaffold. For WRITE requests with enough evidence to draft, keep that gap note after the draft unless placing it first is necessary to avoid an unsupported claim.

Workflow

1. Build a one-sentence thesis: In [task/setting], we address [gap] by [method/formulation], supported by [evidence], within [boundary].
2. Build the story spine: problem, gap, insight, method, evidence, boundary, and implication.
3. Create a contribution-evidence map before writing strong claims. Use columns: Claim, First stated in, Mechanism support, Evidence, and Boundary/overclaim risk.
4. Choose the section reference and assign one job to each paragraph.
5. Draft from evidence outward.
6. Calibrate claim verbs: show, indicate, suggest, support, enable, demonstrate only when directly supported.
7. Remove unsupported novelty, universal claims, and vague adjectives.
8. Add at least one section-specific interpretation sentence when the supplied evidence allows it: mechanism for Results, component role for Ablation, handoff rationale for Methods, or bottleneck for Introduction.
9. When the user supplies diagnostic notes, category breakdowns, ablation notes, failure modes, or nearest-neighbor details, use them in the draft rather than collapsing them into generic improvement wording.
10. For showcase-grade or non-minimal WRITE requests, revise the draft once before final output. Remove generic frames, replace vague verbs with mechanism-specific wording, and turn boundaries into scope statements.
11. Check sentence roles: every sentence must have a function, necessity, placement, connection, and evidence boundary.
12. For WRITE mode, return prose first, then a short explanation of paragraph job, evidence used, and boundary. For PLAN or AUDIT mode, return planning or diagnostic tables before draft prose when they are needed.

Default Output By Mode

WRITE Mode

Use this by default for drafting requests.

## Draft

[Final revised English manuscript prose first. For non-minimal writing
requests, write a complete paragraph with section-specific interpretation, not
only a safe summary.]

## Why this works

- Paragraph job:
- Argument flow:
- Evidence-to-claim link:

## Evidence boundary

- Used:
- Not supplied:

## Do-not-claim

- ...

Keep the notes after the draft concise. If the user explicitly asks for only give the paragraph, final prose only, or output only, provide just the final revised prose unless doing so would hide an unsupported claim or missing evidence.

PLAN Mode

One-sentence thesis
[In task/setting, we address gap by method/formulation, supported by evidence, within boundary.]

Story spine
| Node | Section location | Claim | Evidence anchor | Boundary | If removed, what breaks? |

Section job map
| Section/paragraph | Job | Evidence anchor | Boundary |

Source-note triage
| Source item | Fact / Assumption / Unsupported | Can enter prose? | Handling |

Draft
[English manuscript prose]

Claim-evidence map
| Claim | First stated in | Mechanism support | Evidence support | Boundary/overclaim risk | Repair |

Unsupported or downgraded claims
| Requested claim | Status | Reason | Safe wording |

Sentence role audit
| Sentence/span | Job | Needed because | Connection to previous/next | Evidence boundary | Action |

Missing evidence or assumptions
- ...

AUDIT Mode

## Verdict

## Evidence boundary

## Unsupported or overstrong claims

## Required repairs

## Safe rewrite