# Lektorat Auto Revise > Autonomously works off a `lektorat audit` findings report by routing each artefact by type — documentation to the `audience-doc-author` agent (fully autonomous), blog posts to the `blog-author` skill (assisted) — composing a per-file briefing that binds the audience artefact and writing-style spec, letting the author revise, then re-auditing until no finding above the severity floor remains and there is no regression. Operationalises `spec/project/lektorat-auto-revise/`. Invoke when the user asks to "arbeite das Lektorat-Audit automatisch ab", "auto-revise the audit findings", "remediate the lektorat findings with the right author", or equivalent EN/DE requests. Don't use for the read-only audit or interactive per-finding fixes (use `lektorat-apply`), or for first authorship (`audience-doc-author`/`blog-author`). Writes to `.audits/lektorat-auto-revise//`. Supports resume per `spec/claude/resumable-work/`. - Skill: `nolte/lektorat-auto-revise` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add nolte/lektorat-auto-revise` - Raw SKILL.md: https://api.skillmd.com/api/skills/nolte/lektorat-auto-revise/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: nolte (https://skillmd.com/u/nolte) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/nolte/lektorat-auto-revise --- # Lektorat Auto-Revise Operationalises `spec/project/lektorat-auto-revise/` for the `nolte-shared` plugin: the autonomous bridge from an existing `lektorat audit` findings report to a finished, re-verified artefact. It consumes the audit, routes each affected artefact to the **author best suited to its type**, composes a per-file briefing, lets that author revise, and re-audits until the artefact converges. This skill binds the spec's contract to an on-disk procedure. It owns **no** editorial rules of its own — severities, dimensions, scope, audience binding, and semantic-preservation guarantees belong to `spec/project/lektorat/`; the writing-style and audience rules belong to the dispatched author's bound specs. When this skill and a spec disagree, the spec wins and this skill needs the update. ## German trigger phrases This skill also triggers on equivalent German-language requests, including: - "arbeite das Lektorat-Audit automatisch ab" - "lass den passenden Author die Findings abarbeiten" - "vollautomatische Lektorats-Remediation" - "revidiere die Audit-Findings automatisch" ## User-language policy Detect the user's language from their message and respond in it. The machine-readable JSON output uses English keys (mirroring `spec/project/lektorat/` §Outputs); the human-readable `summary.md` uses English section headings so downstream tooling can parse it; prose around the report is localised. ## Why this is a skill, not an agent - **The blog route's interactive touchpoint is load-bearing**: `blog-author` is an interactive skill whose briefing inputs (topic-as-thesis, source list, slug, cross-language binding key) can't be reconstructed from findings; surfacing that dialogue back to the operator only fits a skill. - **Externally visible writes**: outputs land under `.audits/lektorat-auto-revise//` and the dispatched author edits in-scope artefacts in place; the skill owns the persistent on-disk state. - **Orchestration role**: the skill dispatches the `audience-doc-author` agent and the `blog-author` skill, re-runs the `lektorat audit` for the convergence gate, and stays in the main thread to compose briefings, run the assisted-blog dialogue, and write outputs. - **Counter-dimension considered**: the documentation route alone is fire-and-forget and would fit an agent, but consolidating both routes plus the convergence loop behind one entry point keeps the operator's mental model coherent and is where the load-bearing interactivity lives. ## Inputs - **Audit report** (required): a path to an existing `.audits/lektorat//findings.json`, or a directive to run a fresh `lektorat-apply` audit first and consume its output. The report's shape is `spec/project/lektorat/` §Outputs; consume it verbatim and never add, drop, or rename a field. - **Severity floor** (optional): inherited from the input run; defaults to addressing every `critical` and `warning` finding. A caller may narrow to `critical` for a de-noised pass. - **Audience-artefact path** (optional): resolved through the same priority chain as `lektorat` §Audience binding; a missing artefact stops per-file remediation with the message pointing at `audience-identify`. ## Procedure Run the steps in order. Each maps to a requirement block in the spec; consult the spec for the authoritative rules rather than relying on this summary. 1. **Resolve the audit report** (§Input contract). Load the `findings.json`. Process only the `findings` array. For any file named in an `inventory_findings` entry, do **not** dispatch an author — surface the infrastructure condition to the operator instead. 2. **Group findings by `file`** so each artefact is handled once with the full set of its findings. 3. **Route each file** (§Artefact-type routing) into exactly one class before any dispatch: - `documentation` (MkDocs page under `docs//`, top-level Markdown, or a GitHub Release/Issue/PR body) → `audience-doc-author` agent. - `blog-post` (carries the consumer's cross-language binding key, e.g. `translationKey`, in its frontmatter) → `blog-author` skill. - `rejected` (any path `lektorat` §Scope and applicability excludes — `spec/`, `skills/**/SKILL.md`, `skills/**/templates/**`, `skills/**/examples/**`, `agents/*.md`, source, generated config) → hard-reject with a one-sentence message naming the owning flow; dispatch no author. - A file matching none of the three classes stops the whole run with an operator-facing message; never silently skip it. 4. **Compose the briefing per file** (§Author briefing composition): the file's findings verbatim, the resolved audience set (via the `lektorat` priority chain, no other rule), the bound writing-style spec(s), and the target D1–D6 dimensions. For any file with a **D1 readability finding**, also pass the LIX target inputs per `spec/project/readability-lix/` §Iterative improvement loop: the current `lix`, the resolved corridor (`aim`/`warn`/`crit` for the file's `content_mode` and language), and the dominant lever (`ASL` or `LWP`), so the author's pass targets the right readability lever. Never dispatch an author without **both** a resolved audience set and a bound writing-style spec — a missing either is a per-file stop condition. 5. **Dispatch the matching author** (§Autonomy and human touchpoints): - Documentation → dispatch `audience-doc-author` **fully autonomously** (no per-finding approval, no diff gate). Pass the audience-artefact path, the doc-type spec, the `prose-style` baseline, and the file under revision. - Blog → run **assisted**: surface the findings-derived briefing to `blog-author` and let its briefing touchpoint stand. Never fabricate the briefing inputs it requires. - Never rewrite prose in this skill; the rewrite is always the author's. 6. **Re-audit convergence gate** (§Re-audit convergence gate). After each author pass, re-run the `lektorat audit` on the revised artefact with the same configuration. A file is **converged** only when it has no remaining finding at or above the severity floor **and** its post-revision finding count is ≤ its pre-revision count. A D1 readability finding converges only when the re-audit shows LIX at or below the file's `warn` corridor (`readability-lix` §Iterative improvement loop); a LIX drop achieved by a forbidden transformation (decompounding an established term, vaguer-shorter swaps, altering a protected term) is a semantic-preservation failure, not convergence. On non-convergence, re-dispatch the author with the residual findings up to **2 author passes** per file, then **escalate** the residual to the operator — never loop further. Flag any **regression** (post-count higher than pre-count, or a risen LIX) and never auto-accept it as converged. Treat any semantic-preservation violation (§Semantic preservation) as a failed pass. 7. **Write the run trail** (§Outputs) under `.audits/lektorat-auto-revise//`: `routing.json`, `run.json` (referencing the source audit run), per-file rewrite diffs with pre/post counts and pass count, and `summary.md`. List escalated and regressed files **first** in `summary.md` so unremediated findings can't be overlooked. ## Hard rules - Never define, re-classify, or re-weight findings, severities, or dimensions; consume the `lektorat` report verbatim. The rules live in the spec, not here. - Never dispatch an author without both a resolved audience set and a bound writing-style spec. "No audience" and "no style spec" are stop conditions, not soft defaults. - Never rewrite prose in this skill — every rewrite is delegated to the routed author, so the author's audience and writing-style competence applies. - Never run the documentation route with a human gate, and never run the blog route without `blog-author`'s briefing touchpoint. - Never mark a file converged while it has a finding at or above the severity floor or a regression; never loop past 2 author passes — escalate instead. - Never route or dispatch an author at a `rejected`-class path; never widen the routing table to a class `lektorat` forbids. - Never silently skip an unroutable file or truncate unremediated findings — stop the run or escalate, and record it. ## Gotchas Per `spec/claude/skill-management/` §Gotchas: concrete corrections to non-obvious environment facts the executing agent would otherwise get wrong. - **Finding IDs aren't stable across a rewrite, so convergence is judged by count and severity — never by ID diff.** The `findings.json` `id` is a hash of file + dimension + line; once an author rewrite shifts line numbers, the re-audit re-keys every finding. Compare the re-audit's per-file finding count and highest severity against the pre-revision snapshot; matching IDs across passes will spuriously miss both regressions and fixes. - **A lower LIX number is not automatically convergence.** A D1 readability finding converges only when the re-audit shows LIX at or below the file's `warn` corridor (`spec/project/readability-lix/` §Iterative improvement loop) *and* the drop was earned by a permitted transformation. A LIX reduction achieved by decompounding an established term, a vaguer-but-shorter swap, or altering a protected term is a semantic-preservation failure — treat it as a failed pass, not a win. - **Routing is decided by frontmatter and path, not by reading the prose.** A blog post is identified by the consumer's cross-language binding key (e.g. `translationKey`) in its frontmatter, not by living under `docs/`; a file under `docs//` without that key is `documentation`. Misrouting a blog post to the autonomous `audience-doc-author` route skips `blog-author`'s load-bearing briefing touchpoint (topic-as-thesis, source list, slug), which can't be reconstructed from findings. ## Examples - Read `examples/01-documentation-route-autonomous.md` when a `docs//` page with a D1 readability finding routes to the `audience-doc-author` agent fully autonomously and converges on the LIX corridor. - Read `examples/02-blog-route-assisted.md` when a file carrying the cross-language binding key routes to the `blog-author` skill and runs assisted, preserving its briefing touchpoint. - Read `examples/03-rejected-path-and-escalation.md` when a finding sits at a rejected-class path (e.g. `spec/`) or a file fails to converge within 2 passes and must be escalated and surfaced first in `summary.md`. ## Resumability Per `spec/claude/resumable-work/`, this skill is `resumable: true`. State is persisted to `.resume/lektorat-auto-revise/.yml` after every successful per-file convergence and after each named phase boundary. On re-invocation, scan that directory for files with `status: in_progress` whose `inputs:` snapshot matches the current invocation; if one matches, prompt the operator with `Resume run from phase (last checkpoint )? [resume / start-new / discard]`. The state-file envelope and fail-closed semantics on schema or YAML errors are defined in the spec; don't duplicate them here.