text-eval

name	text-eval
description	Evaluate prose for AI-generated-text signatures and overbuilt writing, protecting the author's voice. Context-isolated evaluator grounded in a portable AI-tells pattern library. Use to check whether text reads as machine-written, find AI tells, or review drafts for authenticity. Report-only by default.
lane	creative
creative_scrub_ref_version	1
allowed-tools	Read, Glob, Grep, Task

Text Evaluator

Reads prose with a forensic ear for machine cadence and reports AI-generated-text signatures — by clustering, not by single signals — while protecting the author's voice. Scaffolded by /skill-builder audit (Step 4c-bis) on the install-on-absence rule; ships working but conservative (report-only). Tightening any check into a blocking gate is a deliberate user act.

Detection policy is canonical in .claude/skills/skill-builder/references/creative-integrity.md (the Nine Scrub Principles, severity architecture, the cluster-density rule, the voice-protection gate). This skill carries the detection content in references/text-tells.md and the evaluation workflow below. It does not restate policy.

Directives

(Seeded by audit, machine-owned. These are conservative defaults, not your words — ratify your own with /skill-builder inline text-eval [directive], or document a voice profile to activate the voice-dependent checks fully.)

Detection is by clustering, not single signals. No pattern blocks in isolation unless its catalog row is marked [hard]. Default escalation: 1 signal → CONSIDER, 2 → SHOULD FIX, 3+ → MUST FIX. Dedupe by mechanism when counting clusters.

Protect the voice. Calibrate for human presence, not against AI. Any flagged pattern matching a documented voice characteristic demotes to advisory. Absent a documented voice profile, every voice-dependent judgment is advisory-only — never blocking, especially on technical or non-English content. Never fabricate imperfections.

Context isolation. The evaluator runs as a fresh-context subagent that never shared the conversation that wrote the text (principle 6). It receives the pattern library and the text — nothing else.

Report-only by default. This skill evaluates and reports. It never rewrites the author's text and never triggers regeneration. Fix decisions are the author's.

Model-Lane Preflight

CHECKPOINT — Model-Lane Preflight Gate (slim, advisory-only; fires at the TOP of this skill's workflow, before any generative step):

1. SKIP silently IF ANY of: /route dispatched this invocation after running its own lane preflight (a lane advisory OR ask already printed, route otherwise ASSESSED the lane this turn, or it was deliberately suppressed — route's coverage extends to the whole endeavor whether route asked, matched, or the user declined, including skills this one chains); this skill's frontmatter sets model-lane-gate: off; the session is headless / non-interactive; this gate is executing inside a subagent (a subagent is already model-pinned via its own frontmatter); OR an advisory for this LANE already printed this endeavor AND the active model is unchanged (advise at most once per (lane, active-model) pair per endeavor).
2. Resolve this skill's lane (own lane: frontmatter → the Skill→Lane table in .claude/skills/skill-builder/references/model-lanes.md → NO LANE) and the lane's Preferred Model. IF no lane resolves OR the preferred-model cell is empty → silent no-op (correctly absent, not a gap). IF this skill declares per-function lanes AND the invocation's resolved mode has its own row → that function's lane governs; otherwise the skill-level lane covers all modes.
3. Read ACTIVE_MODEL from the session system-context line "The exact model ID is …" (strip [1m]/[200k], lowercase — model-lanes.md § Active-Model Detection; never an env var or Bash probe). IF the Preferred Model is malformed or names a clearly superseded family → emit the one-line stale-mapping advisory and proceed; never advise from a stale mapping, never validate against a hardcoded list.
4. Compare. Match → silent no-op. Mismatch → print ONE advisory line and PROCEED: "Lane advisory: this skill declares the <lane> lane (preferred <preferred>); the active model is <active> — to align, run /model <preferred> and re-invoke; proceeding as-is." Naming the command is INFORMATIONAL (named-command advisory directive, 2026-06-11): NEVER an AskUserQuestion, never a blocking wait (No-Switch-Prompt directive, 2026-06-06). A skill can NEVER switch the session model and never asks the user to.
5. Primary-lane scope. This gate fires ONCE per invocation, for the resolved PRIMARY lane. Cross-lane mid-workflow steps never re-advise — they delegate via this skill's LANE-AGENT-EMBED Delegation Map when present (see .claude/skills/skill-builder/references/lane-delegation.md). Delegation is never a substitute for THIS gate: analytical primary work runs on the analytical main model — never a creative main session orchestrating analytical-pinned agents to dodge the lane.

Commands

Command	Action
/text-eval [path-or-paste]	Evaluate the supplied prose; return a tiered authenticity report

Workflow: Evaluate

1. Provenance guard. Confirm what you are evaluating — AI-drafted in-session content, unknown, or human-authored. Never evaluate calibration/reference texts or non-prose files. Provenance only changes the framing of the report, never whether a tell is named.
2. Calibrate by content type. Short-form (posts, captions, teasers) → single-signal severity, skip corpus checks, apply the wit-shaped-noise single-line exemption. Long-form → the full library. (Define a calibration table here if your project needs one.)
3. Spawn the context-isolated evaluator. Run the ai-text-evaluator agent (context: none) with the pattern library and the text only — never inline in the conversation that produced the text. The agent grounds against references/text-tells.md.
4. Tier by cluster density (per the directives above and the catalog's [hard] rows): report blocking flags (MUST FIX + every [hard] hit) uncapped; cap advisory flags (SHOULD FIX / CONSIDER) at 3–5 so the signal is not buried.
5. Check FOR human presence first. Before requiring any change, count the human-presence markers (references/text-tells.md § Human-presence markers). 4+ strong markers plus a few rhetorical devices is human writing under revision pressure — surface an over-correction warning rather than demanding edits.
6. Report: the tiered findings with evidence + falsifiable test per flag, the human-presence read, and (advisory) any cluster that crossed a blocking threshold. Do not edit the text.

Severity Architecture

Three tiers — MUST FIX / SHOULD FIX / CONSIDER — ordered by cluster density, not by detection. Two non-obvious rules (principle 7): report all blocking flags uncapped but cap advisory flags (3–5), and any flag reflecting a hard project directive is mandatory regardless of its tier. [hard] catalog rows (mechanical defects: pipeline-artifact residue, invisible Unicode; factual-integrity failures: ghost citations, fabricated experience) block at first occurrence — they are not subject to the cluster rule.

Grounding

Before evaluating:

1. Read references/text-tells.md (the pattern library — pattern → tell → falsifiable test).
2. State: "I will apply [PATTERN/TIER] from references/text-tells.md, scored by cluster density per the directives."

• references/text-tells.md — portable AI-text-signature library (version-synced by audit)
• .claude/skills/skill-builder/references/creative-integrity.md — canonical severity / cluster / voice policy + the Nine Scrub Principles