simple-skills

Claude Skills for structured, adversarial evaluation of ideas, decisions, and research papers, and for writing without the signs of AI-generated text. Each skill is a SKILL.md that Claude loads when a request matches its description. Three of them are that single file and nothing else; humaner and humanertr each add a scanner script (Python 3 standard library only) and a reference file. Each replaces a single-pass opinion with a written, repeatable method.

Skills at a glance

Skill	Method	Needs	Produces
darpa-test	Heilmeier Catechism: DARPA's nine-question proposal screen, run as seven stages	One concrete idea already in hand	A stage-by-stage interrogation ending in a one-word verdict: go, fix-this-first, or stop
humaner	Wikipedia's catalogue of AI-writing signs, rewritten as functional rules with physics and LaTeX exemptions, plus a scanner script	A draft to clean, or a writing task that must not read as AI-generated, in English or Turkish (LaTeX, Markdown, plain text or email)	The revised text in its original format, a change log by rule family, open items that need a number or source, and a scan report with line numbers
humanertr	The same functional rules for Turkish: a tiered lexicon of calques and fillers, suffix-aware matching with İ/ı case folding, register handling for academic, official and general text, plus a scanner script	A Turkish draft to clean, or Turkish writing that must not read as AI-generated (XeLaTeX, Markdown, plain text or email)	The revised text in its original format, a change log by rule family, open items that need a number or source, and a scan report with line numbers and density metrics
llm-council-advisor	Five independent personas, a blind peer review, and a Chairman synthesis	A decision, plan, strategy, or draft with room for defensible disagreement	Verdict summary, dissent map, synthesis matrix, and an execution protocol with success metrics
research-paper-analyzer	Thirteen evidence-grounded modes on a shared claim-and-evidence ledger	One paper or a set of papers (full text gives the best results)	Mode-specific tables, for example claims against evidence, overreach flags, genuine gaps, or a study design

Quick start

1. Get the files: git clone https://github.com/codekyha/simple-skills.git, or download the repository as a ZIP from GitHub's Code menu.

2. Install one skill (full steps under Installing):

   • Claude.ai: ZIP the skill's folder and upload it under Settings → Capabilities → Skills.
   • Claude Code: copy the skill's folder into ~/.claude/skills/.

3. Ask for it in plain words, or by name:

   Run the DARPA test on this idea: ...

   Humanize this abstract: ... (or /humaner in Claude Code)

   Bu özeti doğallaştır: ... (or /humanertr in Claude Code)

   Council this: should we ...

   Analyze this paper for [topic]. (attach the PDF)

The skills

darpa-test

A structured interrogation of one idea (a project, proposal, product, hire, or purchase) that surfaces the question its owner cannot answer without changing the subject. The nine Heilmeier questions run as seven stages: plain-language restatement; current practice and its limits; what is actually novel; who is affected and what changes; risk ranked by probability; the full cost; and checkpoints that each carry a date, a number, and a decision fixed in advance.

A verdict closes the run. Each stage is scored out of 5 on honesty and specificity, not polish; the weakest answer is named; the first thing a motivated skeptic would attack is stated; and the answer ends in exactly one of go, fix-this-first, or stop with a one-sentence reason. A "go" authorizes only the cheapest test from the risk list, not the full build.

Built-in push-back rules reject vague answers: a beneficiary category in place of a named role and moment, risks sorted by drama instead of likelihood, and a checkpoint without a number or a date.

• Try: "run the DARPA test on this," "poke holes in this plan," "is this worth building."
• Not for: brainstorming. The skill interrogates a proposal that already exists.

humaner

Rules for writing without the signs of AI-generated text, and a procedure for removing them from an existing draft. The rules come from Wikipedia's Signs of AI writing page, a field guide that WikiProject AI Cleanup maintains for spotting chatbot prose. The page says that removing a surface marker while keeping the problem behind it only makes detection harder, so each rule names a function: an evaluation with no fact behind it, a rhetorical template, chat register in a deliverable, formatting used as decoration, or an unverified reference. The word lists are search strings, and a synonym doing the same job counts as the same violation.

The rules are adjusted for academic physics writing. Technical senses stay (an energy landscape, an essential singularity, spins aligned with an axis, a classical turning point), and so do en dashes in eponymous names such as Kerr–Taub–NUT, the serial comma, hyphenated compound modifiers and passive voice in methods passages. Turkish text gets the same functional rules plus a short list of candidate calques, labeled as extrapolation. Edits never invent support: a claim that needs a number or a source keeps a [needs number] or [needs source] marker and goes on a list of open items.

The bundled scripts/scan.py (Python 3, standard library only) masks code, math, comments and citation commands in Markdown, LaTeX, plain text and email, then reports em dashes, flagged vocabulary, negative parallelisms, participle tails, inline-header lists, chatbot residue, placeholders and tracking parameters with line and column. Without code execution, Claude runs the same checks by hand.

• Try: "humanize this abstract," "check my referee response for AI tells," "/humaner on this cover letter," "make this read less like ChatGPT."
• Not for: deciding whether someone else's text was AI-written. The page itself says no single sign proves authorship, and the scanner sees surface markers only.

humanertr

The Turkish counterpart of humaner. Wikipedia's page covers English only, so the rules keep the same functions (an evaluation with no fact behind it, a rhetorical template, chat register in a deliverable, formatting used as decoration, an unverified reference) and the Turkish word lists are extrapolations, tiered by confidence. Tier A entries are Turkish calques of catalogued signs (önemli bir rol oynamaktadır, altını çizmektedir, sadece X değil, aynı zamanda Y). Tier B entries are Turkish-specific candidates, such as -erek/-arak converb tails that interpret a sentence. Tier C entries are common in careful human Turkish and are reported as densities only.

The skill separates three registers. Academic text keeps the passive, -maktadır and the roadmap sentence of an introduction. Official correspondence keeps its fixed formulae (Bilgilerinize arz ederim, Konu:, İlgi:). General text gets plain active sentences. Bilingual abstracts are edited as a pair so that Özet and Abstract keep the same claims and numbers.

The bundled scripts/scan_tr.py (Python 3, standard library only) normalizes text to NFC, folds case with the Turkish pairs İ→i and I→ı, matches stems with suffix tails, and masks code, math, comments and citation commands. It reports hits with line and column, plus the share of sentences that carry -maktadır or open with a connector. Thresholds are uncalibrated, and the only Turkish studies found for the sources file give no lexical markers; references/sources.md states what is and is not supported.

• Try: "Türkçe metni doğallaştır," "yapay zekâ izlerini temizle," "/humanertr on this özet," "bu yazı ChatGPT gibi duruyor."
• Not for: spelling or TDK orthography, or deciding whether someone else's text was AI-written.

llm-council-advisor

Five personas analyze the same raw input independently: the Contrarian (attack vectors and failure modes), the First-Principles Thinker (rebuild the problem from its actual constraints), the Expansionist (asymmetric upside), the Outsider (zero-context reading), and the Executor (sequencing and the next physical action). Their outputs are anonymized as Advisors A to E, and each persona reviews the other four for valid points, critical flaws, and blind spots.

A Chairman then weighs the views instead of averaging them and issues a verdict summary; a dissent map that rates each objection Critical, Secondary, or Discarded; a synthesis matrix (surviving risks, unexploited leverage, first-principles reality); and a numbered execution protocol with a success metric per step. The run is sequential inside one response, and "the proposal is sound as stated" is a legitimate verdict: the aim is rigor, not manufactured disagreement.

• Try: "council this," "run an advisor review on this strategy," "give me a no-sycophancy evaluation of this proposal."
• Not for: single-fact lookups or questions with one reasonable answer.

research-paper-analyzer

Thirteen ways of reading a research paper critically, on one paper or a set. Every output is anchored to something specific in the text (a claim, a number, a limitation, a conflict between papers) with a location, so a reader can open the paper and check it. Work starts from a shared ledger of claims, evidence, numbers, limitations, and assumptions, so a claim keeps one ID across modes. Evidence is judged by what the paper's type actually offers (empirical, theoretical, computational, qualitative, or review), and each claim receives a tier with a one-line reason: Strong, Moderate, Weak, Unsupported, or Inherited (resting on cited work rather than shown here).

A bare "analyze this paper" runs a compact single-paper pass, with a table of this shape followed by unresolved questions, overreach flags, and three follow-up questions:

| ID | Claim | Evidence offered | Location | Tier |

For a specific analysis, ask by name or number:

#	Mode	Input	Produces
1	Argument Map	one paper	Thesis, sub-claims, warrants, logical gaps, load-bearing assumptions
2	Evidence Strength Audit	one or more	Evidence tier for each paper's headline claim, with a ranking
3	Comparison Matrix	two or more	Side-by-side table of question, sample, method, variables, findings, limitations
4	Consensus vs Contention	three or more	Settled, contested, and open questions, counted by independent evidence
5	Real Gap Finder	one or more	Up to five anchored, non-generic gaps with their novelty status
6	Conflict Explainer	two or more	Verdict on whether conflicting findings are genuine or definitional or methodological, plus a decisive test
7	Numbers Extractor	one paper	Every key quantity with its location, plus consistency checks
8	Claim-to-Evidence Tracer	one paper	Each claim mapped to its evidence and location; orphan claims and orphan evidence
9	Action Brief	one or more	One-page brief whose implications are tied to evidence tier
10	Next Study Designer	a gap	Study design sketch with a falsification criterion
11	Literature Themer	four or more	Five to eight themes with agreements, disagreements, methods, limits
12	Overreach Flag	one paper	Conclusions that exceed the evidence, with severity and a defensible rewording
13	Follow-up Questions	one paper	Three questions tied to specific findings and methods

Modes chain in dependency order: 3 -> 2 -> 4 -> 6 -> 11 -> 5 -> 10 -> 9 for a set of papers, 8 -> 7 -> 12 -> 13 for deep reading of one paper.

• Try: "is this conclusion justified," "build a comparison matrix for these five papers," "mode 5: find the real gaps," "extract every number and check them against each other."
• Limits: given only an abstract, it runs the single-paper pass at abstract level and says plainly that the evidence cannot be audited. It is analysis only; it does not write papers, referee reports, or rebuttals, though its output is suitable raw material.

Using the skills

Claude keeps each installed skill's description in context and loads the full instructions only when a request matches. A plain request is usually enough; naming the skill is the more reliable trigger for ambiguous requests. In Claude Code every skill is also a slash command, for example /research-paper-analyzer.

Installing

The examples use research-paper-analyzer; substitute another skill's folder name (darpa-test, humaner, humanertr or llm-council-advisor) for the others. Commands run from the repository root. A skill's folder name must equal the name in its SKILL.md frontmatter, as it does here. A ZIP or copy of the folder carries any bundled scripts/ and references/ with it.

Claude.ai

Skills need a Pro, Max, Team, or Enterprise plan with code execution enabled (Enterprise owners must first enable skills in Admin settings → Capabilities).

1. Open Settings → Capabilities and turn on Code execution and file creation.

2. Build a ZIP whose top-level entry is the skill folder, not the whole repository. Any ZIP tool works; this command is cross-platform (Python 3):

   python -m zipfile -c research-paper-analyzer.zip research-paper-analyzer

3. In Settings → Capabilities → Skills, choose Upload skill, select the ZIP, and switch the skill on.

Some documentation labels the settings page Features rather than Capabilities. Uploaded skills are private to your account, so each team member uploads separately; a skill uploaded here is not available through the Claude API, which has its own upload.

If a .skill file is presented in a Claude conversation, its file card offers a Save skill button (where your organization allows skill creation) that installs it without the steps above. The packaging script bundled with Anthropic's skill-creator example skill produces the same kind of file (a ZIP archive) and validates the frontmatter first; run it from the skill-creator directory:

python -m scripts.package_skill path/to/research-paper-analyzer

Claude Code

Skills are read from the filesystem, so nothing is uploaded. Copy the folder into the personal directory for availability in every project, or into a project's .claude/skills/ to limit it to that repository (commit it there to share it with collaborators):

# Available in every project
mkdir -p ~/.claude/skills
cp -r research-paper-analyzer ~/.claude/skills/

# Available in one project only
mkdir -p /path/to/project/.claude/skills
cp -r research-paper-analyzer /path/to/project/.claude/skills/

Windows PowerShell:

New-Item -ItemType Directory -Force "$HOME\.claude\skills" | Out-Null
Copy-Item -Recurse .\research-paper-analyzer "$HOME\.claude\skills\"

Claude Code watches these directories, so adding or editing a skill takes effect in the current session without a restart. The main exception is a top-level skills directory that did not exist when the session started; run /reload-skills in that case.

Cowork and other Claude surfaces

Consult the surface's own skill-installation documentation; the loading mechanism (folder path, upload, or catalog) varies by surface, but the SKILL.md files here are portable without modification. Claude Code's documentation also lists skills enabled on your claude.ai account as loading in Cowork, cloud, and signed-in terminal sessions, so a Claude.ai upload may already cover those.

Troubleshooting

• Skills are greyed out in Claude.ai: code execution or skills may be disabled by an organization owner, or code execution may be off in your own settings.
• Upload fails: the usual causes are a folder name that does not match the skill's name, a missing SKILL.md, invalid characters in the name or description, and a ZIP over the size limit.
• Claude Code does not see a new skill: run /reload-skills if ~/.claude/skills/ was created after the session started.
• The skill does not trigger on its own: name it explicitly in the request, or call it as a slash command in Claude Code.

Repository layout

.
├── README.md
├── LICENSE
├── darpa-test/
│   └── SKILL.md
├── humaner/
│   ├── SKILL.md
│   ├── references/
│   │   └── sources.md
│   └── scripts/
│       └── scan.py
├── humanertr/
│   ├── SKILL.md
│   ├── references/
│   │   ├── lexicon-tr.md
│   │   └── sources.md
│   └── scripts/
│       └── scan_tr.py
├── llm-council-advisor/
│   └── SKILL.md
└── research-paper-analyzer/
    └── SKILL.md

Adding a skill to this repository

1. Create <skill-name>/SKILL.md. The folder name must equal the name field.

2. Frontmatter: name is at most 64 characters, lowercase letters, digits, and hyphens, and must not contain the reserved words "anthropic" or "claude". description is non-empty, at most 1024 characters, free of angle brackets, and states both what the skill does and when to use it.

3. Use only the frontmatter keys that claude.ai and the API accept on upload: name, description, license, compatibility, metadata, and allowed-tools.

4. Keep SKILL.md under about 500 lines, and keep exactly one SKILL.md per skill folder; the skill-creator validator rejects extras because claude.ai and the API do on upload.

5. Validate before committing, from the skill-creator directory:

   python -m scripts.quick_validate path/to/skill-folder

6. Add the skill to the tables, the layout tree, and the provenance list in this README.

Built .skill packages are git-ignored; they are release artifacts, not source.

Provenance

• darpa-test is built directly from George Heilmeier's original nine-question catechism (public framework, DARPA, 1970s onward), independently phrased for this implementation.
• humaner turns the patterns catalogued on Wikipedia:Signs of AI writing (WikiProject AI Cleanup) into prescriptive rules, independently phrased; the flagged words and phrases are the page's own search terms, and its prose is not reproduced. The physics and LaTeX exemptions, the Turkish candidates (an extrapolation the page does not cover), the scanner script and the evidence notes in references/sources.md are additions made here.
• humanertr applies the functions behind humaner to Turkish. No Turkish counterpart of the Wikipedia page or corpus study of lexical markers was found, so the tiered lexicon is an extrapolation made here, as are the scanner and the register rules; references/sources.md lists the two Turkish studies consulted (Er, 2025; Karakaya Özyer, 2026), which report no lexical markers.
• llm-council-advisor implements a five-persona-plus-Chairman adversarial synthesis pipeline; the parallel-subagent dispatch described in some tool specifications is adapted here to a sequential single-context pipeline for environments without true parallel subagent execution.
• research-paper-analyzer adapts a series of prompts for analyzing research papers posted on Threads by Hamidullah Khan, PhD (@iamkhanphd; lead post). The prompt texts were reconstructed from the series and are not reproduced verbatim. The shared ledger, the paper-type evidence rubric, the evidence tiers, the pipeline ordering, and the verification checklist are additions made here.

License

MIT — see LICENSE.