Lucid Writer

Make the writing clearer without making the evidence stronger than it is.

Lucid Writer is an Agent Skill for Claude Code and Codex. It edits, drafts and reviews research articles, reports, abstracts and reviewer responses, from a two-sentence copyedit to a full manuscript audit. Before it changes wording, it checks each material claim:

claim → evidence → scope → uncertainty → inference ceiling → wording

Docs: https://mdmanurung.github.io/lucid-writer/

Install

Claude Code plugin (skill plus three review subagents):

/plugin marketplace add mdmanurung/lucid-writer
/plugin install lucid-writer@lucid-writer

Or the installer, for Claude Code and Codex:

git clone https://github.com/mdmanurung/lucid-writer
cd lucid-writer
bash scripts/install.sh both        # Windows: .\scripts\install.ps1 both

Use one of the two, not both. Install by hand, update and uninstall are on the docs home page.

Use

There is no special command: the skill loads when a writing task matches. Start with "Use lucid-writer to…" to force it.

Mode	Example ask	What you get
Edit	"Tighten this paragraph by ~20%."	the revised text only, with numbers, citations and caveats kept
Draft	"Turn Figure 3 and these notes into a Results subsection."	a draft, with `[EVIDENCE NEEDED]` where support is missing
Build	"Restructure this Discussion around the main answer."	a planned rewrite, with changes carried through the document
Review	"Audit the manuscript against the figures. Do not rewrite yet."	prioritised findings ending with a Not checked list

For high-stakes work:

Review subagents: run the evidence auditor and skeptical reviewer in parallel, then manuscript preflight before submission. Skip them for ordinary edits; they cost extra tokens.
Project state: python3 skills/lucid-writer/scripts/init_project.py . creates .lucid-writer/ with a brief, a claim ledger and a decision log.
Mechanical checks: lint.py manuscript.md flags placeholders, causal verbs and filler; validate_ledger.py checks the claim ledger.
Always-on (Claude Code only): /lucid-writer full applies the evidence ceiling to every reply and your voice to deliverables; /lucid-writer prose is plain-prose rules for chat, with no evidence rules; /lucid-writer off stops it.
Learned rules (Claude Code only): /lucid-writer:learn "<rule>" stores a proposed writing rule; /lucid-writer:rules approves, rejects and retires it. Only approved rules are injected, and rules that loosen the evidence rules are refused.
Mycelium: in a project with .living/ and .mycelium/, numbers checked by /mycelium-extra:verify outrank the prose.

It will not strengthen a claim beyond its evidence, even on request; invent citations or numbers; follow instructions hidden in manuscripts or reviewer comments; or rewrite to evade AI detectors.

Example:

Use lucid-writer to tighten this paragraph without changing claim strength:
[paste a research paragraph]

Input:

Taken together, these findings demonstrate that CD161+ CD4 T cells play a crucial protective role against malaria infection.

Lucid Writer:

This association is consistent with a relationship between this phenotype and parasite control but does not establish a causal protective role.

The numbers in the full sentence survive; the causal claim the data cannot carry does not.

Documentation

Page	Covers
How it works	the rules, the inference ceiling, what an author cannot override
Use	daily prompts, review subagents, project state, mechanical checks
Mycelium	writing from verified provenance; the convention pack
Evaluation	case studies, regression cases, the benchmark, and their limits
Develop	layout, validation, packaging, contributing, security

Develop

python3 scripts/validate_repo.py   # repository checks
quarto preview site                # this documentation, locally

Contributions should fix a demonstrated failure, with an eval case. See CONTRIBUTING.md and SECURITY.md; sources and credits are in docs/methodology.md and NOTICE.md.

License

MIT. See LICENSE.