Four Claude Code skills for writing, research, investigation, and distribution. Each one runs a time-boxed, multi-agent pipeline that produces a sourced, cited, natural-sounding artifact. No API keys required for any of them (except whatever you already have for the backends they query).

note: AI-text detectors evolve daily. Entire paragraphs sometimes get flagged because of a phrase currently over-used by various models. Finding 2-3 fitting synonyms is usually enough to break the pattern (and no, this paragraph was not written by this product).

What's in the box

Command	Purpose
/selfwrite	Rewrite prose through an iteration loop until it reads natural and passes AI detectors
/selfresearch	Time-boxed academic research pipeline that returns a cited literature review
/selfinvestigate	Thesis-driven investigative pipeline across FEC, SEC, court, and news records
/selfpost	Browser-driven Twitter/X posting via the Claude for Chrome extension

Install

mkdir -p ~/.claude/skills
cp selfwrite.md selfresearch.md selfinvestigate.md selfpost.md ~/.claude/skills/

Requires Claude Code CLI. Verify with /selfwrite, /selfresearch, /selfinvestigate, or /selfpost in any project.

Backend reference cards (sources/*.md) are read at runtime by /selfresearch and /selfinvestigate. Either invoke the skills from this repo or copy sources/ into whichever project you're working in.

Node helpers for /selfpost:

npm install      # deterministic char counting, queue CLI, preflight
npm test         # runs the validators vitest suite

The skill runs without them but falls back to less-reliable Claude-interpreted logic.

Usage

/selfwrite (prose polish)

/selfwrite "task description" <duration>

Duration is Nm or Nh, minimum 10m. Examples:

/selfwrite "tighten this opinion column on housing policy" 20m
/selfwrite "rewrite the README as a summary of this analysis" 45m
/selfwrite "upgrade writing.md to NYT journalist quality" 6h

Selfwrite picks up your audience, purpose, register, and tone at intake, then iterates THINK → DRAFT → REVIEW → REVISE → SCORE → REFLECT until convergence. Every version is kept. Revisions that damage more than they improve revert automatically. Simple-rewrite mode polishes what's there; deep-rewrite mode runs a research tree to surface missing evidence.

Task	Budget
Quick edits (tighten a column, polish a README)	15-30m
Full rewrites (reports, memos, skill files)	45m-2h
Deep research rewrites (add evidence, counterarguments)	1-6h

/selfresearch (cited academic research)

/selfresearch "research question" <duration>

Duration is Nm or Nh, minimum 15m. Examples:

/selfresearch "known failure modes of RLHF" 30m
/selfresearch "effects of SNAP benefits on food insecurity" 1h
/selfresearch "EU AI Act compliance pathways for US firms" 2h

Queries Semantic Scholar, OpenAlex, arXiv, and (optionally) web. Runs PLAN → ITERATE → SYNTHESIZE → VERIFY → SUMMARIZE. Every claim carries a four-tier tag (SRC, SYN, INF, UNV) anchored to a specific extracted quote. The verifier confirms each tag before report.md is released.

Output	Budget
Evidence brief (1000-2000 words)	20-30m
Focused report (2000-4000 words)	30-60m
Literature review (3000-6000 words)	1-2h
Exhaustive survey	2h+

/selfinvestigate (thesis-driven investigation)

/selfinvestigate "thesis" <duration> [--stance=prove|disprove|investigate]

Duration is Nm or Nh, minimum 30m. Default stance: investigate (neutral). Examples:

/selfinvestigate "Donor networks shifted to funding Trump by 2020" 2h
/selfinvestigate "Company X's compliance failures were known at the board level" 90m --stance=prove
/selfinvestigate "The regulatory-capture narrative at Agency Y is overstated" 3h --stance=disprove

Queries FEC, OpenSecrets, SEC EDGAR, CourtListener, Wayback Machine, plus academic and web sources. Runs SCOPE → QUESTION WEB → RESEARCH → CONNECT → WRITE. A mandatory thesis-assessment user gate fires before writing, so prove can't produce a one-sided brief without the honest evidence moment.

Output	Budget
Quick investigative brief	45-60m
Standard investigation (3000-5000 words)	1-2h
Deep investigation (5000-8000 words)	2-4h
Book-chapter scope	4h+

Stance quickguide: investigate for a neutral read, prove when you have strong priors and want efficient consolidation of supporting evidence, disprove to stress-test the thesis by hunting counter-evidence.

/selfpost (Twitter/X posting)

/selfpost new "topic" [tweet|thread]
/selfpost list
/selfpost run [--unattended]
/selfpost show <id>
/selfpost cancel <id>

Examples:

/selfpost new "why small models win at tight tasks"
/selfpost new "the 2026 state of personal AI tooling" thread
/selfpost list
/selfpost run

new drafts a tweet or thread and saves it to queue/twitter/<id>.md with status: draft. You review the file, flip to status: ready, then /selfpost run posts flagged items via your logged-in Chrome browser. Up to 5 posts per run, 30-120s jitter between, 10-per-24-hours rolling cap. Session expiry, CAPTCHA, and rate limits all stop the run and mark the item failed with the error.

Requirements:

• Claude for Chrome extension installed and connected
• A Chrome profile logged into x.com
• Node.js 20+ (for the helpers)
• No Twitter API key, no OAuth, no scheduler

Two posting tiers:

1. Tier 1 (default): Interactive Chrome MCP. Two gates per tweet: an in-chat 'send' confirm and the Chrome extension's own Post-button approval. Use when you're at the keyboard.
2. Tier 2 (opt-in via --unattended): Playwright on a persistent profile. No gates. Use for scheduled or batch posting when you've manually logged the profile in once. Requires npx playwright install chromium chromium-headless-shell.

Shared patterns

All four skills write to runs/<skill>_<timestamp>/ with versioned outputs, a structured log, and a distilled skill.md. Install the distillate to carry learnings into future runs:

cp runs/<run-id>/skill.md ~/.claude/skills/<domain>.md

Tips that apply to all four:

• Answer intake questions concretely. "Skeptical CFO reading a one-pager" yields precise output; "general audience" yields vague output.
• Name a publication voice at /selfwrite intake (Economist, Reuters, NYT News Analysis). Loads a lexicon that constrains word choice and sentence rhythm.
• Pick /selfinvestigate stance deliberately. Neutral is the default; prove and disprove change how the reflector and wave-search weight evidence, not whether counter-queries run.
• Use the full time budget. The skills are engineered to keep going until the deadline. Longer budgets produce deeper output.

Requirements

• Claude Code CLI
• For /selfpost: Claude for Chrome extension, Chrome profile logged into x.com, Node.js 20+