Audience Agent OS v2 · by Noor Intelligence

A BUILD SYSTEM FOR CLAUDE CODE AND OTHER CODING AGENTS

Give your coding agent a build plan it can't bluff its way through.

Audience Agent OS is a master prompt, plugin and template set that steers an AI coding agent to build a real content app: audience research, drafts in your brand voice, a separate reviewer, exact-version approval, one genuine send, and a report that says "unknown" when it doesn't know.

Get the Plugin Pack · £79 Prompt Edition · £29 Pay in sats over Lightning

One-off payment · instant download after card checkout · free v2.x updates

Big agent builds tend to fail in the same three ways.

01 · BREADTH OVER DEPTH

Fifty screens, no working journey

The agent scaffolds every feature it was told about and finishes none of them end to end.

02 · MOCKS IN DISGUISE

"Done" means "the code exists"

Missing credentials become stubbed success. Tests mirror the code instead of trying to break it.

03 · TRUST IN THE MODEL

Permissions decided by a prompt

The model's output gets to approve, send or spend. One injected web page and it does.

HOW IT WORKS

Seven milestones. Each ends at a gate the agent has to prove it passed.

Every gate check is reported as PASS, FAIL or NOT_RUN with its evidence. Passing gates are crossed without waiting for you. The agent stops only for a failed hard check, a real blocker, or something that needs your approval, like a purchase or a live send.

M0

Recon and contract

Requirements table, integration sheet, defaults checked against your services

M1

Auth and tenancy

Organisations, workspaces, roles, isolation tested with real users

M2

Brand knowledge

Versioned facts labelled confirmed, sourced, hypothesis or unknown

M3

Drafts and review

Six roles, immutable versions, separate Reviewer, spend caps

M4

Approval and send

Approval tied to a hash, idempotent dispatch, crash test

M5

Calendar and reports

Real metrics only; missing data shown as unknown

M6

Full registry and hardening

All 50 roles, backup, restore drill, clean walkthrough

The end-of-turn report your agent sends at every gate

Milestone:  M4 — Approval and dispatch
Verified:   INV-4 dispatch auth · INV-12 webhook dedupe
Implemented, not verified:  reconciliation job
Gate checks:  INV-4 PASS · INV-5 PASS · INV-12 PASS · INV-13 NOT_RUN
Blockers:   none
Needed from owner:  approve one test send to your own inbox
Next task:  T-041 bundle-scan test for INV-13

Illustrative format, not a real build log.

THE GUARDRAILS

Fourteen invariants, each with a test that has to reject the bad case.

INV-1

Tenant isolation at two scopes

Organisation and workspace, tested with users in two organisations and with different workspace access in one.

INV-4

Send only what was approved

The content hash must match the approval, and the approver must still hold approval rights at send time.

INV-5

No blind resends

Intent and idempotency key are saved before the provider call; unknown outcomes are reconciled first.

INV-6

Prompt injection contained by code

A step that has read imported text gets read-only tools. It cannot send, approve, change recipients or spend.

INV-11

Bounded AI spend

Budget is reserved atomically in the database before every model call, so parallel workers can't overshoot.

INV-14

Missing is not zero

Reports show unknown for missing data and label results attributed, correlated or unknown.

Release is blocked until every hard check is PASS. A check that was never run blocks release just like one that failed. The built-in Reviewer also checks UK ad and affiliate disclosure, email consent and unsubscribe, and health or financial claims.

WHAT'S INSIDE

Two editions

IncludedPrompt EditionPlugin Pack
Master prompt v2 (.md), works with any coding agentYesYes
REQUIREMENTS, INTEGRATION, CHECKPOINT and DECISIONS templatesYesYes
Claude Code plugin, passes claude plugin validateNoYes
Commands: start (or resume), gate, rolesNoYes
Role registry: 50 roles across 10 departments, plus Director and Reviewer, in YAMLNoYes
Role JSON Schema and registry validator scriptNoYes
Free v2.x updates and email supportYesYes

PRICING

One payment. Use it on every build you run.

Prompt Edition

£29

For anyone using Cursor, Codex, Aider, Lovable or another agent. Paste the prompt, copy the templates, start M0.

  • Master prompt v2 (.md)
  • Four milestone templates
  • Single-user licence
Buy Prompt Edition · £29

Claude Plugin Pack

RECOMMENDED

£79

For Claude Code users who want the build to start, gate and resume with one command each.

  • Everything in Prompt Edition
  • Validated Claude Code plugin with three commands
  • 52-role YAML registry, schema and validator
  • Licence covers one organisation
Buy Plugin Pack · £79

Prefer to pay in sats?

Enter your email and choose an edition. You get a Lightning invoice for the right number of sats at the live BTC/GBP rate. Pay it from any Lightning wallet: your download opens as soon as the payment confirms, and the link is emailed to you too.

whiznurturing992525@getalby.com

STRAIGHT ANSWERS

What this is, and what it isn't

It is

  • A specification and working method your agent follows, written to be audited line by line
  • Built for a defined stack (Next.js, Supabase, Postgres jobs), with every default checked against what you already use
  • Yours to edit: the prompt, templates and registry are plain text

It isn't

  • Finished software or a code boilerplate. Your agent writes the code.
  • Proven by a completed public build yet. The prompt makes the agent report one honest outcome instead of claiming success.
  • Free to run. Model, hosting and email costs are yours; spend caps default to £20 per workspace per month.

FAQ

Before you buy

Which coding agents does it work with?

The prompt is plain Markdown, so any agent that can read files, run commands and write code can follow it. The Plugin Pack adds commands specific to Claude Code.

How do I install the plugin?

Unzip it and start Claude Code with claude --plugin-dir ./audience-agent-os, then run /audience-agent-os:start in your project. The README covers other agents.

Can I build apps for clients with it?

Yes. You own what you build. You can't resell or share the prompt, templates or registry themselves. For agency or redistribution licences, email info@baladanmeri.com.

Which integration does it send through?

One, which you choose. Email newsletter via Resend is a placeholder; the agent checks it against services you already have before building.

Can I get a refund?

If the files are faulty, incomplete or not as described, yes: I fix, replace or refund. Because the download is instant, the 14-day change-of-mind right ends once you download, which you agree to at checkout. Haven't downloaded yet? You can cancel within 14 days for a full refund. Read the full refund policy.

How do I get my download?

Card and Lightning payments both go straight to a private download page, and the same link is emailed to you so you can come back to it. No email after 10 minutes? Check spam, or write to info@baladanmeri.com. Lost the link? Email info@baladanmeri.com.