Academic Writing Harness

Academic Writing Harness

Use this skill for end-to-end academic writing, revision, review, or export in Claude Code. Codex and OpenClaw use the matching repo skill under .agents/skills/academic-writing-harness/.

Required reading

1. Read .copilot-mode.json and AGENTS.md before repository operations.
2. Read docs/harness/academic-writing-workflow.md completely. It is the platform-neutral workflow and artifact contract.
3. Read docs/harness/evaluation-contract.md before running or changing a quality evaluation. It defines solver/scorer separation, evidence locators, budgets, frozen fixtures, and release evidence.
4. Read the relevant source skill for the requested phase:
   • ../project-management/SKILL.md for setup and output type.
   • ../literature-review/SKILL.md for evidence discovery.
   • ../draft-writing/SKILL.md for section drafting.
   • ../manuscript-review/SKILL.md for structured review.
   • ../word-export/SKILL.md for DOCX delivery.
   • ../auto-paper/SKILL.md only for full autopilot runs.

Operating contract

• Resume recorded workspace and project state before starting a new phase.
• Select an output profile before planning. Do not force every academic output into IMRaD.
• Build a claim-evidence ledger before prose and keep source identifiers near every material claim. Evidence locators must include a stable source revision, section/page/paragraph/span, offsets when available, and a hash of the cited span; a model summary is not a locator.
• Register exemplars separately from evidence sources. Extract only declared structural or stylistic features and record provenance.
• Write section-by-section, then score the persisted artifact in a separate, read-only pass. A solver must not award its own gate result or alter scorer fixtures, thresholds, or expected outcomes.
• Declare per-stage tool/model/time/retry budgets and enforce source-dominance limits. Exceeding a budget or letting one source dominate is a recorded gate outcome, not a reason to hide or truncate audit evidence.
• Run deterministic hooks and satisfy each hard gate before advancing. Release claims require frozen positive, negative, mutation, degraded, and resume fixtures plus the command, environment, counts, hashes, and failure details.
• Treat human acceptance as an external authority boundary. An Agent/MCP may inspect or revoke a host-issued Ed25519 confirmation receipt, but must never mint one by labeling its own input as human; receipts bind the project, reviewed artifact/state hashes, named reviewer, confirmation identifier, and time, and are verified only against host-configured public keys.
• Before inserting an image, re-run the read-only C2PA and pinned visible/open-DWT checks against its current bytes. Preserve the original, require documented review for raster uncertainty, and never equate a negative detector result with proof that no watermark exists.
• Keep verified source data, agent interpretation, and user-authored notes in separate trust layers.
• Finish with cross-section review, reference verification, export validation, audit artifacts, and Memory updates.

Tool adaptation

Tool names can differ by runtime. Map capabilities by contract: workspace state, project management, literature search, full-text ingestion, reference storage, draft read/write, quality hooks, pipeline gates, and export. If a required capability is unavailable, record the degraded path rather than inventing evidence or silently skipping a gate.

The default MCP entrypoint is the compact 12-tool facade; the 118-tool full surface remains an explicit compatibility and diagnostics profile. Both must exercise the same domain rules and audit contracts.