How we build SchedulePost: an AI Orchestra, not a chatbox

The short version

Most AI social tools are a single chat box wired to a scheduling queue. You type a request, a model returns a block of text, and you paste it into a calendar. That works for a one-off caption. It falls apart the moment you want research-backed, on-brand, platform-native content that actually publishes reliably — week after week.

SchedulePost is built differently on purpose. Instead of one model doing everything in one pass, we run an AI Orchestra: a small team of specialised agents that each do one job well and hand work to the next. They run on your own AI key — Google Gemini or Anthropic — and the approved posts are handed to a background publisher we treat as real infrastructure. This article is the honest tour of how and why we built it that way.

Why one chat box is not enough

A single prompt asks one model to do five jobs at once: come up with an angle, find evidence, write for each platform, judge its own work, and remember your voice. Models are good at any one of those in isolation. Asking for all five in one breath is where quality quietly drops.

• No separation of concerns. The same pass that invents an idea also grades it, so weak ideas rarely get caught — the model is marking its own homework.
• Research gets lost. If sources are not kept attached to the draft, there is nothing to check a claim against later. Confident, unsupported sentences slip through.
• One caption everywhere. A single output cannot be genuinely native to LinkedIn, an X thread, and a Mastodon note at the same time.
• No durable handoff. Chat output ends in a text box. Getting it scheduled and reliably published is left to you and a browser tab.

Breaking the work into stages fixes each of these. We go deeper on the reasoning in the AI Orchestra explained; here is the architecture itself.

The pipeline, stage by stage

A campaign moves through the Orchestra as a series of handoffs. Each agent receives the previous agent's output, does one focused job, and passes it on. Nothing reaches your calendar until it has been through the whole chain and you have approved it.

1. Brief. You give the goal, audience, and topic. This is the brief every downstream agent works from, so the whole run stays pointed at one outcome instead of drifting.
2. Brainstorm. An agent proposes several distinct angles for the brief. You pick the ones worth pursuing before a word of final copy exists.
3. Source-aware research. A research agent gathers supporting material and keeps the sources attached to the work, so claims can be checked later rather than trusted blindly.
4. Platform-native writing. A writer produces independent drafts for each connected network — a LinkedIn post, an X or Bluesky thread, a Mastodon note — shaped to each platform's norms, not one caption reused.
5. Critic and review. A reviewer reads the drafts against the research and the brief, and can send weak or unsupported work back to be revised — or reject it outright.
6. Schedule. Approved posts go to the publisher, spread across good times informed by your analytics.

What each agent actually does

It helps to see the stages mapped to plain jobs. Each row is one responsibility we deliberately kept separate so it can be done well and checked.

Why a critic is a separate agent

The review pass is the part most single-prompt tools skip, and it is the one we are proudest of. A dedicated critic reads the drafts with fresh eyes — against the research and the original brief — and has the authority to send work back or reject it. An author defending its own first draft is a weak reviewer; a separate critic with permission to say no is a real quality gate.

Because the research sources stay attached from the research stage onward, the critic can actually check whether a claim is supported instead of just judging whether it sounds confident. That is the difference between a tool that polishes prose and one that protects your credibility.

Why BYOK is a first-class decision

Bring-your-own-key is not a billing detail we tacked on — it shaped the product. You connect your own Gemini or Anthropic key, your provider bills the tokens directly with no markup from us, and we charge only for the workflow around them. Keys are validated when you connect them and encrypted at rest.

• Transparency. You see exactly what the model usage costs, billed by your provider at their published rates, not folded into an opaque credit bucket.
• Model choice. You decide whether to run Gemini or Anthropic, and you can change your mind without changing tools.
• You own the AI relationship. The account, the limits, and the data terms are yours directly with the provider.
• Honest economics. We make money when the workflow is worth paying for — not by reselling tokens at a margin.

We make the full case in BYOK AI vs bundled credits, and you can price a real campaign on the AI cost calculator. Most people are surprised how little platform-native drafting costs once a per-token markup is removed.

Publishing as durable infrastructure

Getting good content written is only half the job. The other half is making sure "scheduled" actually means "handled" — not "a tab has to stay open and hope." So the publisher is built like infrastructure, not like a front-end timer.

• A background worker claims due posts safely, so the same post is never published twice.
• Transient failures are retried instead of silently dropped when a network hiccups.
• Interrupted jobs are recovered, so a restart mid-run does not lose your queue.
• Per-platform results are recorded, so you can see exactly what landed where.
• Terminal failures raise an alert, so you can fix and retry only what actually failed.

The full design rationale lives in publishing as infrastructure. The short version: reliability is a feature, and we built for it rather than assuming the happy path.

The reusable voice profile

An orchestra that writes nothing like you is just noise. A reusable voice profile — written once and editable — describes how you sound: sentence length, humour, words you avoid. Every writing pass starts from it, so drafts arrive already in your register and your job is editing rather than rewriting from a blank page. It is also what keeps tone consistent as you scale across platforms.

The analytics learning loop

The pipeline does not just run forward and forget. Analytics by network, time, and angle feed two things: best-time recommendations that the scheduler uses, and the brief for your next campaign. Each run is informed by the last instead of starting cold, which is how output gets better over weeks rather than just staying steady. See turn analytics into your next campaign for the loop in detail.

What you still do

We are deliberate about where the human stays in the loop. The Orchestra gets you most of the way — the angles, the research, the platform-native drafts, a critic pass — but you approve before anything schedules. The AI gets you roughly 90% there; your taste and your knowledge of your audience are the last 10%, and we built the approval step so that last 10% is fast.

Why we built it this way

Every design choice here points at the same goal: produce content you would be happy to put your name on, prove it where it matters, publish it reliably, and learn from it — all on AI costs you control. A chat box can draft a caption. An orchestra with research, a critic, a durable publisher, and a learning loop is a system you can hand a week of work and trust to run. That is the difference we build for.