100xPrism

Stop vibe coding. Ship production-grade software.

One source of truth. 68 modules generate native config for Claude Code · Cursor · Codex · Pi. Quality gates run on every commit.

---

Install

npm (any platform — macOS, Linux, Windows):

npm install -g 100xprism && 100xprism install

curl (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/rajitsaha/100xprism/main/get.sh | bash

Either way, 100xprism install clones the toolkit to ~/100xprism and provisions your AI tools. The npm package is a thin launcher — the modules, hooks, and plugins live in that clone, which 100xprism update keeps current. Before installing, 100xPrism cleans legacy startup hooks, stale command links, and any old owned dashboard process. Install does not edit ~/.zshrc, ~/.bashrc, or ~/.bash_profile. If you want optional aliases for the current terminal only, run source ~/100xprism/shell/aliases.sh.

Windows: plugin sync works, but native Windows module emit is being reworked (#54). For full module support today, install under WSL with either method above.

Set up a project:

cd your-project && 100xprism init

Keep up to date:

100xprism update                    # pull latest, then add/update/remove skills + plugins
100xprism update --plugins-only     # refresh plugins only (repo already current)
100xprism update --no-slim          # skip the one-time always-on index slim
100xprism slim --dry-run            # preview what slimming would change
100xprism uninstall                 # stop dashboard + remove legacy shell-startup entries/symlinks
npm install -g 100xprism@latest     # (optional) upgrade the launcher itself

install and update are fully reconciling, not append-only — every run:

• adds newly shipped skills, slash commands, and curated plugins,
• updates changed ones in place, and
• removes skills, slash-command aliases, and 100xprism-managed plugins that were deleted or merged upstream.

Your own hand-authored skills/commands and any plugins you enabled yourself are never touched. See docs/USAGE.md for details.

Cloned to a custom path? The default install lives at ~/100xprism. If you cloned elsewhere, update your shell + Claude Code config — see Custom install location.

---

The pipeline

/understand → /context → /issue → /spec → /fix → /commit
                                                    ↓
              /techdebt ← /gate → /grill → /pr → /push → /release

Every /commit and /push runs a 5-point gate — tests, security, build, Docker, cloud. Nothing ships without passing.

---

What you get

68 modules	28 slash commands + 40 auto-trigger skills — see full reference below
2 Claude Code core plugins	GitHub + security guidance by default; other integrations are profile recommendations or manual opt-ins
7 database engines	Postgres, Cloud SQL, Snowflake, Databricks, Athena, Presto, Oracle — one /db interface
27 SaaS CLIs	/connect installs + authenticates GitHub, AWS, Stripe, Supabase, and more from .env
4 project templates	node-fullstack · node-frontend · python-api · docker-compose
CI/Release pipelines	Drop-in GitHub Actions for lint + real-DB tests + E2E + semantic-release

---

Optional vendor skill packs

Some skills only matter to some projects, so they don't ship by default. /pack installs them on request instead:

100xprism pack               # list every pack, install state, and anything detected here
100xprism pack detect        # only what matches the current project
100xprism pack install databricks
100xprism pack uninstall databricks

Packs are ownership-tracked per install — 100xprism pack uninstall only removes what it added, never a plugin or command you enabled yourself. Where the vendor ships a CLI (e.g. databricks aitools install), that's preferred; otherwise 100xPrism installs directly per platform. First pack: Databricks (30+ skills — Unity Catalog, Asset Bundles, Lakeflow, model serving, vector search).

---

Token, delivery & value economics

For a while, the goal was singular: make spec-driven, agentic development as autonomous as possible — let the agent plan, build, gate, and ship with less and less human steering.

That part is mostly working. The harder, more important question is the one that comes next: what is all this autonomy costing, what did it help deliver, and what was that delivery worth? Every agent run spends real money in tokens. The dashboard measures cost and observable engineering delivery separately; business or human value remains explicitly unmeasured until an outcome source is connected.

This is a first, deliberately humble attempt to make the measurement chain visible to everyone, because watching it isn't one person's job. It's everybody's responsibility.

100xprism tokens                  # local dashboard
100xprism tokens --json           # fast, versioned provider-counter report
100xprism tokens --json --tool codex
100xprism audit --json             # standing-context estimate + skills/plugins/hooks
100xprism value                    # delivery economics

• 100xprism tokens — one offline, machine-wide dashboard at a single URL: the input/output/cache split, a startup-bloat meter, an estimated code-vs-files-read-vs-logs-vs-chat composition, and $ cost — by project, model, session, and skill. Claude Code and Codex provide exact local token counters; Cursor agent-transcript JSONL and Antigravity local artifacts contribute project/session/activity coverage but are never assigned invented token cost because their local formats expose no counters. Cursor chats, state.vscdb, and legacy transcript .txt are outside the collector's scope. It auto-refreshes every 30 seconds once started. Start it explicitly with 100xprism tokens, 100xprism dashboard, or 100xprism install --dashboard; shell startup never starts it.
• 100x-value — the dashboard joins exact local token counters to observable delivery evidence. It shows every directory that consumed tokens plus agentic projects discovered machine-wide via marker files. Git supplies commits, deduplicated merged PRs, releases, files, insertions, and deletions; non-repos use an explicitly labeled filesystem-mtime estimate. Directories from unsupported tools show — cost, never a misleading $0. Delivery unit costs use attributed spend only, show attribution coverage, and are never presented as business ROI.
• Budgets, provenance, recommendations & GitHub PR insights — optional daily/weekly limits drive dashboard and shell alerts. A provenance strip shows pricing coverage, outcome-join coverage, source counts, and date range. Recommendations are ranked by estimated opportunity and pair the observed evidence with an action that preserves AI-native autonomy. Optional GitHub CLI integration can fetch PR metadata for locally checked-out remotes plus configured users/repos.

Paying only for the skills a repo uses

An installed module's description may be re-sent on every turn; its body is loaded only when invoked by tools that support progressive disclosure. Fresh installs therefore default to the 12 must-have workflows plus one generated 100x-resolver. Specialist bodies stay on disk and are loaded through /100x <workflow> (or the closest native skill invocation) only when needed.

100xprism optimize              # enforce must-only for user scope + this repo
100xprism optimize --all-projects
100xprism optimize --skills=profile  # widen to detected project profiles
100xprism optimize --skills=all      # restore every module
100xprism slim ...                   # compatibility alias

Current deterministic description-footprint estimates (characters ÷ 4): Claude falls from ~4,816 tokens in all mode to ~574 in must; Cursor falls from ~1,942 to ~261. Codex and Pi use the same must-first selection. CI enforces committed must-mode budgets, and 100xprism audit separately inventories instruction files, aliases, plugins, and hooks. These are standing-context estimates, not provider-billed usage.

Full guide: docs/token-optimization.md.

---

What it actually changed

Two open-source products built with this toolkit. Everything in the Measured column is read straight from git history and reproducible with 100xprism value — nothing is modelled.

Measured from git	100xprism	agentbreeder
First commit → first public release	6 days	32 days
Releases shipped	26	17
Median gap between releases	1 day	2 days
Merged PRs	70	171
Commits	204	772
Lines added	78,447	541,370
Peak commits in a single day	39	61
Active engineering days	29	53

Combined: 43 releases · 255 merged PRs · 650K lines · 94 active engineering days.

The same toolkit drove two further products that are not open source — a draft-content generation SaaS and an AI-native real-estate deal-analysis SaaS — adding 2,219 commits · 970 merged PRs · 1.13M lines over 171 active days.

Across all five: 3,282 commits · 1,225 merged PRs · 1.78M lines added · 5,846 files, by one person.

The comparison, stated as an estimate

There is no counterfactual — nobody built these twice. So this is arithmetic on a stated assumption, not a measurement, and you should substitute your own numbers:

Assume a conventional team merges ~2 feature-sized PRs per engineer-week. 1,225 merged PRs ÷ 2 ≈ 612 engineer-weeks ≈ 11.8 engineer-years, delivered in 265 active days. At a $85/hr contract rate that block of work prices at roughly $2.1M.

Three honest caveats. PR size varies enormously, so PR count is a coarse unit. A solo builder skips coordination overhead a team pays, but also carries no review redundancy. And the figure moves a long way if you assume 1 or 4 PRs per engineer-week — which is exactly why the assumption is written down rather than buried in a headline number.

What is not claimed: that any of this is business value. Revenue, retention, and customer outcomes are unmeasured here, and the dashboard deliberately refuses to invent a score for them.

Where the speed comes from

Lever	Effect
/spec → /orchestrate → /fix	A vague request becomes an implementation-ready spec before any code is written
/gate + the gate-on-commit hook	Broken work cannot reach main — enforcement is a hook, not a reminder
/issue	An observation becomes a root-caused, actionable GitHub issue in one step
/release	26 releases at a 1-day median cadence, because shipping is one command
40 auto-trigger skills	Marketing, SEO, pricing, and CRO work happens in-repo instead of waiting on a contractor

---

Slash commands

The following 28 slash commands are available. Run them inside Claude Code. In Codex, use the generated repo skill by name instead, for example $gate, $commit, or /skills.

Lifecycle

Command	What it does
/branch	Create a conventional feature branch (feat/, fix/, chore/)
/commit	Gate → stage → conventional commit
/grill	Adversarial code review before opening a PR
/pr	Gate → push branch → create PR
/push	Gate → push → monitor CI → verify production health
/release patch|minor|major	Semantic versioning + publish to PyPI/npm/Docker Hub
/launch	Full deploy pipeline in one command

Quality

Command	What it does
/gate	Mandatory 5-point quality gate (tests, security, build, Docker, cloud)
/test	All test layers (unit, integration, E2E) — loops until 95% coverage
/lint	Auto-detect and fix all lint errors (ESLint, TypeScript, ruff)
/security	Vulnerability + secrets scan, auto-fix where possible
/cloud-security	GCP IAM, networking, PII, and compliance scan
/eval	Run module evals — check triggers and output quality

Engineering

Command	What it does
/spec	Turn a vague request into an implementation-ready spec
/fix	Autonomous bug fixer — CI failures, docker logs, Slack pastes
/orchestrate	Plan-first methodology for complex multi-step tasks
/techdebt	Dead code, duplication, stale TODOs
/context	7-day git + GitHub activity dump — orient before coding
/update-claude	Write a CLAUDE.md rule after any correction

Data & Infrastructure

Command	What it does
/db	Query any of 7 database engines from one interface
/query	Plain-English analytics — describe what you want, get SQL
/connect	Install + auth 27 SaaS CLIs from .env

Documentation & Architecture

Command	What it does
/docs	Detect code changes and update documentation
/issue	Investigate a bug and create a detailed GitHub issue
/architect	Architectural Q&A and decision matrices
/enterprise-design	Full technical blueprint — IA, API, data model, stack

Auto-trigger skills (40)

These modules activate automatically when you describe a relevant task — no slash command needed.

Category	Modules
Marketing copy	copywriting, copy-editing, cold-email, email-sequence, ad-creative, social-content
SEO	seo-audit, ai-seo, programmatic-seo, schema-markup, site-architecture
CRO & conversion	page-cro, signup-flow-cro, onboarding-cro, form-cro, popup-cro, paywall-upgrade-cro
Growth & strategy	content-strategy, marketing-ideas, marketing-psychology, launch-strategy, referral-program, churn-prevention, free-tool-strategy, ab-test-setup, analytics-tracking, pricing-strategy
Sales	sales-enablement, competitor-alternatives, paid-ads, revops, product-marketing-context
Design & UX	visual-system-architect, interaction-engineer, figma-translator, motion-designer, data-viz, a11y-auditor
Engineering	subagents, terminal-setup

After 100xprism slim, most of these load through the generated 100x-resolver catalog instead of sitting in your always-on context — same capability, a fraction of the standing token cost.

---

How it works in your tool

Tool	Generated artifact	Auto-trigger?
Claude Code	~/.claude/skills/<slug>/ + slash command aliases	Yes — per description
Cursor	.cursor/rules/<slug>.mdc (one file per module)	Yes — per description
Codex	AGENTS.md + .agents/skills/<slug>/ + .codex/hooks.json	Yes — repo skills

Every supported tool loads module bodies on demand rather than inlining them, so the always-on context stays small. tier: core marks the modules Cursor keeps resident (alwaysApply: true) — since v3.1 that is gate alone, because gate is the only one that must fire unprompted; commit, push, and branch are explicitly invoked and their gate enforcement is guaranteed by the gate-on-commit hook, not by prompt residency. Everything else is fetched when its description matches. Claude Code plugins remain Claude-specific; use Codex /plugins for Codex-native plugins.

Removed in v3.0.0: Windsurf, Copilot, Gemini, and Antigravity. Those adapters emitted a single concatenated file (~60K chars) that sat in context on every turn — the opposite of progressive disclosure.

This deletes files in your projects. 100xprism update removes the .windsurfrules, GEMINI.md, ANTIGRAVITY.md, and .github/copilot-instructions.md it previously generated from every project in ~/.100xprism/tracked-projects; 100xprism init does the same for the project it runs in.

Only files carrying the Generated by 100xprism header are touched, so a hand-written file of the same name is left alone. If you edited a generated file, it is still removed — but every removal is copied to ~/.100xprism/removed-artifacts/<timestamp>/ first, and is only removed once that copy succeeds. These files are also normally committed, so deletions show up in git status for you to review.

---

Deprecations & removals

What recent releases took away, and what to do about it. Full detail in the changelog.

Version	Removed / changed	Migration
3.1	commit, push, branch dropped from tier: core — they no longer sit resident in Cursor's context	None. They are still slash commands; gate enforcement now comes from the gate-on-commit hook rather than prompt residency
3.1	Specialist modules leave the always-on skill index on your first update	Automatic and announced once. Undo with 100xprism slim --skills=all; skip with 100xprism update --no-slim
3.1	Project config moves out of CLAUDE.md into .claude/100xprism.yml	None required — modules read the new file first and fall back to CLAUDE.md, so existing repos keep working
3.1	New CLAUDE.md scaffolds use a router table instead of commented config blocks	Applies to newly scaffolded projects only; existing files are never rewritten
3.0.1	Nine modules re-tiered core → on-demand (Cursor context down 61%)	None. No content changed, no effect on Claude Code, Codex, or slash commands
3.0.0	Windsurf, Copilot, Gemini, Antigravity adapters deleted	Automatic. Generated files are backed up to ~/.100xprism/removed-artifacts/<timestamp>/ before removal. Use Claude Code, Cursor, or Codex
3.0.0	emit-concat subcommand, render_concat, render_index_only	Internal API — use emit-cursor / emit-codex / emit-claude-code
3.0.0	Instruction-file discovery narrowed to CLAUDE.md AGENTS.md .cursorrules	Move any config out of .windsurfrules / GEMINI.md / copilot-instructions.md
2.4.9	11 orphaned module reference files	None — they were unreachable from any SKILL.md
2.4.9	Dashboard and update-check no longer auto-start from shell startup	Start explicitly: 100xprism tokens, or 100xprism install --dashboard
2.4.x	systems-architect → merged into enterprise-design; conversion-copy → merged into copywriting	Use the surviving module; update prunes the old skill and its alias

Nothing is removed without a backup. Every deletion inside your repositories is copied to ~/.100xprism/removed-artifacts/<timestamp>/<full-project-path>/ first and removed only if that copy succeeded — and only if the file carries the Generated by 100xprism header in its first 10 lines.

---

Common CI traps it fixes

npm install 404 inside Docker · useState(false) opacity-0 breaking Playwright · integration tests silently excluded from the gate. Full breakdown →

---

More

• Full usage guide — daily patterns, multi-project setup, CI templates, project config, troubleshooting
• Architecture — why modules replaced workflows + skills
• Token usage & optimization — audit your plugin/skill footprint and monitor token spend with the local dashboard
• Changelog · Roadmap · Issues

---

Built by Rajit Saha · 23 years building data and platform systems at scale

If this saves you time, star the repo.