A context-first, compounding writing system and plugin for developing ideas, drafting, revising, stress-testing, and atomizing content without losing the writer's authentic voice.
Built for Divings by Sangam and packaged for Claude Code, OpenAI Codex, Google Antigravity, Cursor, Windsurf, and open agentic coding environments.
Most AI writing workflows are amnesiac: you prompt an agent, correct its mistakes, polish the draft, and then start from zero on the next piece.
Compound Writing turns every writing session into a compounding asset:
- Context Before Craft: Your persistent identity, syntax rules, editorial constraints, and domain standards live in maintained files (
VOICE.md,STYLE.md,DOMAIN.md). - Scaffolding Over Synthetic Prose: Drafts default to statement headings, tensions, and facts. Sangam writes the sentences, protecting human voice and Substack authorship integrity.
- The Compounding Loop (
sw-save): When you make recurring line edits or reject an AI formulation, runningsw-savecodifies that learning permanently so you never give the same feedback twice. - Platform Atomization (
sw-atomize): The Substack essay is the authoritative source; LinkedIn carousels (Gamma) and Koe/Shipper captions are cleanly atomized from it.
Install directly via the Claude plugin system:
claude plugin add https://github.com/divingsbysangam/compound-writingOr clone the repository into your project:
git clone https://github.com/divingsbysangam/compound-writing.git .claude/plugins/compound-writingOnce installed, all 35 skills are available as slash commands:
/sw-scribe # Master router
/sw-interview # Mine real build memories and numbers
/sw-draft # Generate structural scaffolds
/sw-ai-check # Scrub AI residue, banned words, and em dashes
/sw-atomize # Turn Substack essay into LinkedIn caption & carousel
/sw-save # Save confirmed lessons into persistent context
The repository includes native .agents/skills/ and .agents/rules/:
# Add as a workspace customization or copy to global config
cp -r skills/* ~/.gemini/config/skills/In chat, simply type sw-interview, sw-draft, or ask: "Use sw-ai-check on this draft."
The repository includes root AGENTS.md and .agents/rules/ which Cursor and Windsurf load automatically.
Recognized via .codex-plugin/plugin.json.
Run sw-commands to list all workflows, or sw-help for decision routing.
sw-scribe— Front-door router: evaluates what you have and routes to the smallest useful workflow.sw-emergent— Composes custom multi-stage pipelines for unconventional essays.
sw-setup-project— Initializes a piece folder withVOICE.md,STYLE.md, anddrafts/.sw-onboarding— Calibrates or refreshes voice rules between Reflective and Technical modes.sw-save— The Compounding Engine: turns feedback or manual rewrite diffs into permanent updates inVOICE.mdorSTYLE.md.
sw-brainstorm— Surfaces raw ideas in Technical/Salesforce or Reflective mode.sw-interview— Interrogates the author to extract real build memories, physical scenes, and empirical numbers. Never hallucinates.sw-thesis— Formulates one falsifiable, arguable claim that a knowledgeable peer could dispute.sw-promise— Clarifies reader takeaway and runs the "So What?" chain to the bottom.sw-outline— Structures material into 4-part short-form or 6-beat long-form arcs.sw-hook— Generates cold openers (concrete memory, confession, or specific result). Bans throat-clearing.sw-transition— Builds causal bridges between sections.sw-analogy— Generates Matt Levine-style physical analogies for dry architecture.sw-simplify— Strips corporate jargon into plain English without losing technical precision.
sw-draft— Defaults to a scaffold, not finished prose (protects Pangram authorship detection).sw-bluf— Checks the First Paragraph Rule (core tension landed in first 2–3 sentences).sw-dev-edit— Checks structure, stakes, symptom-before-diagnosis, and self-implication.sw-line-edit— Polishes syntax, sentence length variance, and connective flow.sw-voice-check— Audits draft againstVOICE.md(mandatory contractions, five human signals).sw-ai-check— Strict anti-AI purge: zero em dashes, eliminate banned lexicon, remove structural overcompletion.sw-tracks— Removes scaffolding notes, bracketed assumptions, and process residue.sw-final-pass— Executes the 15-point pre-delivery verification checklist.
sw-objections— Surfaces skeptical developer and architect pushback.sw-asshole— Attacks weak or under-supported claims uncharitably.sw-hemingway— Enforces verbal economy, cuts filler (actually, sitting, by then), and mandates active voice.sw-hitchcock— Tests narrative tension and suspense across the beats.sw-reader— Audits first-time reader cognitive friction and unexplained jumps.sw-mom— Tests clarity for non-technical readers on reflective pieces.sw-sedaris— Identifies places for self-implication, humor, or a wry parenthetical aside.sw-sorkin— Audits dialogue snap and conversational momentum.sw-vonnegut— Enforces core storytelling fundamentals.sw-panel— Synthesizes multiple reviewer perspectives simultaneously.sw-debate— Pits conflicting editorial reviewers against each other.
sw-atomize— Transforms approved Substack essay into a Koe/Shipper LinkedIn caption (8–14 word bold claim, no closing question, PascalCase hashtags) + Gamma carousel slide headings.sw-sf-check— Cross-references technical and release claims against the official Salesforce Release Notes PDF and verifies dev org boundaries.
- Zero Em Dashes: Absolute ban. Use hyphens, commas, colons, semicolons, or periods.
- Contractions Always: "Don't", not "do not"; "couldn't", not "could not".
- Scaffolding by Default: Substack drafts default to outlines, beats, tensions, and required facts. Sangam writes the sentences.
- Nothing Invented: All scene details, metrics, and quotes are sourced strictly from memory via
sw-interview. - Salesforce Truth: The release notes PDF is the sole source of truth for release claims.
MIT © Sangam Gella