mimesis

mimesis

μίμησις: imitation, representation.

Two opposite jobs for Claude Code. Strip AI tells from writing and design so they read as human-made, and compile dense, token-efficient instructions for a machine reader. The plugin is named after what the first job neutralises: the machine imitating a person.

The thesis is one line: the guidelines outrank the model. A versioned tell corpus and a deterministic linter carry the quality, not whichever model happens to be writing. When a future model regresses, the corpus and the linter still produce clean, em-dash-free text.

/plugin marketplace add blakecyze/mimesis
/plugin install mimesis

The skills

Skill	Domain	What it does	Invocation
mimesis-human	A	Remove the machine fingerprint from prose and add human texture. Audit, rewrite, or generate.	/mimesis-human [text or file]
mimesis-concise	A	Cut prose to what does work: throat-clearing, dutiful summaries, hedging, puffery, redundancy.	/mimesis-concise [text or file]
mimesis-tone	A	Write or recast prose in a named voice: persuasive, practitioner, warm, blunt, plain. Still de-slopped.	/mimesis-tone [text] --tone <name>
mimesis-design	A	Read-only audit of markup, CSS or a described UI for AI design tells: purple gradients, Inter 700, three-card rows.	/mimesis-design [path]
mimesis-compile	B	Compile dense, token-efficient instructions for an LLM. Bypasses the humanisation linter by design.	/mimesis-compile [intent]
mimesis-principles	A	Standing anti-tell rules so prose comes out clean by default. Loaded automatically.	auto

How it works

Domain A runs a three-layer stack; Domain B runs none of it. The linter is the line between them: it cleans every human-facing output and never touches a machine-facing one. Sharing machinery across that line is a bug, not a feature.

flowchart TB
  subgraph A["Domain A · human-facing (de-slop)"]
    direction LR
    a1([text]) --> a2["L1 corpus"] --> a3["L2 linter"] --> a4([clean text])
    a5["L3 generator<br/>(optional)"] -. into .-> a2
  end
  subgraph B["Domain B · machine-facing (compile)"]
    direction LR
    b1([intent]) --> b2["compile.md"] --> b3([prompt for an LLM])
  end

Layer	What	Role
L1 corpus	tells, craft, design-tells, tones	primary, model-independent, never skipped
L2 linter	em dashes, kill-list, parallelism, hedges	runs on every Domain A output
L3 generator	Claude by default, Codex via --gen codex	optional, swappable, never auto-routed

The kill list in L1 is versioned by model era, because the tells drift. Domain B has its own reference, compile.md, version-keyed to the 2026 model families with stale advice flagged.

What the linter flags

A single Python 3 stdlib script, no dependencies. Detection only: it reports, the skill restructures.

Category	Catches	Severity
em-dash	em dashes, plus the en-dash, double-hyphen and spaced-hyphen workarounds	blocker
blacklist	kill-list words, versioned by model era, with inflections	important / polish
negative-parallelism	"not X but Y", "it's not X it's Y"	important
participle-tail	clause-final "highlighting...", "underscoring..."	important
hedging	reflexive hedge phrases	polish
smart-quotes	curly quotes and apostrophes	polish
repeated-opener	the same opener on three or more sentences	polish
low-burstiness, low-lexical-diversity	flat rhythm, narrow vocabulary (length-gated heuristics)	advisory

It spares ---, --flags and table separators, so it does not fire on real markdown or CLI text.

Two absolute rules

Zero em dashes, ever. The skill restructures into commas, full stops, parentheses or two sentences. A restructure, not a swap. The linter catches the en-dash, double-hyphen and spaced-hyphen workarounds too.

Catch the merely perceived. Tricolons, participle tails, dutiful sign-offs and antithesis get flagged and rewritten even when grammatically fine, via the final-pass checklist in craft.md.

Run the linter on its own

No plugin required:

python3 linter path/to/draft.md        # readable report
python3 linter --json path/to/draft.md # machine-readable findings
cat draft.md | python3 linter          # from stdin

Exit status is 0 when clean, 1 when anything is flagged.

Limits

• Detectors false-positive. mimesis writes for signal and voice. It does not promise to beat any given detector, and swapping synonyms to game a score changes nothing detectors measure.
• Meaning preservation is mandatory in rewrite mode. No claim changes, no fabricated anecdotes or metrics.
• The kill list drifts as models drift. It is versioned, not eternal.

Licence

MIT. See LICENSE.