Four-mode analytical toolkit for documents, books, web pages, decisions, and untrusted skills. SUMMARIZE produces an even-handed one-page digest of any text (author motivation, intended audience, tone, key quotes). PROOFREAD makes light spelling/typo corrections on full-length documents (including entire books, chapter by chapter), driven by the Harper grammar checker (local, offline, rule-named findings), and produces a separate proof report with suggestions and repeated-phrase findings. DECIDE picks three archetypal personas from a built-in roster and runs a transcripted debate over a question, complex decision, or 'what should I do next' — returning a recommendation plus the full debate. AUDIT vets a third-party skill, readvisor, or paradigm before the user trusts it: a deterministic payload scan of the bundled scripts plus an LLM audit protocol against a threat model (prompt injection, privilege escalation, epistemic contamination, executable payload, audit evasion), producing a plain-English explainer, a pass/flag/fail verdict, and a report. Use for 'analyze this', 'summarize this', 'one-pager', 'tl;dr but smart', 'read this for me', 'proofread this', 'copy-edit this book', 'fix typos in', 'spot repetition', 'help me decide', 'pros and cons', 'what should I do about', 'three perspectives on', 'debate this', 'scan this skill', 'is this skill safe', 'audit this', 'vet this', 'check this before I enable it', 'what does this skill actually do'. Works with plain text, markdown, html, pasted text, web pages, free-form questions, skill folders or .skill zips, and any document enough converts to a markdown twin — Word documents, PDFs, EPUB ebooks, PowerPoint decks, Excel workbooks (read the twin, or read_file the original and enough converts it; PDFs, decks and workbooks need the PDF extra).
A lightweight, four-mode analytical workbench for text, decisions, and things
you haven't decided to trust yet. Designed to run comfortably on 16 GB
machines — the dispatch lives here, the heavy guidance lives in references/
and is loaded only when a mode fires.
| Mode | What it does | Reference to load |
|---|---|---|
summarize | One-page even-handed digest of any text — author motivation, biases, intended audience, tone/lexicon, the point, key quotes. | references/summarize.md |
proofread | Full-document copy-editing pass (typos/spelling) driven by Harper (optional; see references/proofread.md), with a separate proof report covering suggestions and repeated-phrase notes. | references/proofread.md |
decide | Three archetypes from the roster below debate the user's question; output is a recommendation plus the full debate. | references/decide.md |
audit | Vets a third-party skill, readvisor, or paradigm before you trust it: deterministic payload scan (scripts/payload_scanner.py) + the LLM audit protocol against the threat model; verdict + report. | references/audit.md |
summarize.proofread.decide..skill zip, readvisor, or paradigm in — "scan this skill",
"is this skill safe", "audit this", "vet this", "check this before I
enable it", "what does this skill actually do" → audit.references/*.md file completely. Do not improvise the
mode from the SKILL.md alone — the reference contains the workflow, the
output template, calibration notes, and edge cases.If the user request is ambiguous (e.g. they hand you a draft and say "look at this"), ask one clarifying question: "Do you want a summary, a proofread, a decision/debate on something in the text, or a safety audit?" Don't guess.
One exception to the ask-first rule: when the user is about to enable,
install, or trust something that came from outside — a downloaded skill, a
shared readvisor, a paradigm someone sent them — audit is the right mode even if
they didn't use the word. Say which mode you picked and why, then run it.
These ten archetypes are shared across modes. decide picks three of them per
request (whichever three are most relevant to the question). proofread and
summarize may invoke individual personas when forming suggestions or
characterizing a text's likely reader. audit does not use the roster —
an audit is forensic, and findings are evidence, not opinions.
Each persona is described in one line here. The reference files quote and expand them as needed; do not re-describe them inline elsewhere.
When decide selects three, prefer combinations that will actually
disagree. Skeptic + Editor + Stylist tend to converge; Skeptic +
Contrarian + Pragmatist tend to argue productively.
All artifacts from analyzer go to:
rness/io/output/analyzer/Mirror any subfolder the user named (e.g. if they asked you to analyze a
manuscript in rness/io/input/manuscripts/foo/, output to
rness/io/output/analyzer/manuscripts/foo/).
Filename conventions:
summarize → <source-slug>-summary.mdproofread → <source-slug>-proof-report.md (separate from any corrected
manuscript copy; see references/proofread.md for the corrected-copy
filename rule)decide → <question-slug>-decision.mdaudit → audits/<skill-name>/<YYYY-MM-DD>-audit.md, plus a
machine-readable audits/<skill-name>/verdict.json sidecar next to it.
The sidecar is a contract with the harness — its exact keys and vocabulary
are specified in references/audit.md; do not improvise them.If rness/io/output/analyzer/ does not exist, create it.
audit's pass / flag / fail is a category backed by named findings,
not a number — and it stays advisory. The user decides.)references/audit.md restates these as hard rules
because they are the mechanism by which the audit avoids becoming the thing
it's looking for.Four lenses. One workbench. The user decides.
enough-tooltip-text: "use analyzer to summarize a document or webpage, proofread any text including full length works, help you decide between possible options, or audit a skill you got from somewhere else before you trust it."
0gsd/enough
3cf8a60
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.