> ## Documentation Index
> Fetch the complete documentation index at: https://docs.app.useharmonia.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Content strategy is rarely the bottleneck. Execution is.

> How Harmonia turns authorized startup sources into a governed content workflow, and what remains unverified.

Startups usually have useful things to say: a founder interview, product demonstration, customer lesson or launch note. The difficult work is turning that material into accurate, platform-specific content, coordinating review and proving that the approved result exists.

Harmonia organizes that work as a durable content operation. An operator supplies authorized material and a brief. The system preserves both separately, normalizes source evidence, proposes strategy, pauses for approval, prepares content, reviews it and presents exact effects for authorization. Execution ends with a receipt and independent verification rather than a model saying “done.”

## Judgment and authority

Seven named specialists divide the cognitive work: Nimi analyzes evidence, Ryan develops strategy, Temi plans approved outputs, Noni writes and produces, Dara reviews, Maya presents validated state, and Nova answers operator questions. Native Strands tasks run through AgentCore Runtime using Bedrock Claude and Nova models. Host-preloaded skills guide their methods; skills are not facts or permission.

The deterministic coordinator chooses the next permitted task. Models cannot invent source references, approve their own work, select an arbitrary next stage or declare that an effect completed. Host code validates typed results and persists accepted state in DynamoDB.

| Concern                                     | Authority                |
| ------------------------------------------- | ------------------------ |
| Interpret evidence and propose content      | Bounded specialists      |
| Validate identities, scope and provenance   | Deterministic host       |
| Authorize exact consequential effects       | Authenticated operator   |
| Claim and execute approved operations       | Deterministic worker     |
| Establish resulting bytes or provider state | Independent verification |

## One approval boundary

Dashboard, conversational studio and allow-listed Telegram routes share the same action contract. Natural language can locate a pending decision and explain it. It cannot manufacture approval. The operator sees the action, destination, payload, risk and cost; the recorded decision binds to the current canonical identity. Material changes invalidate that authority.

Strategy approval and effect approval are separate. Permission to develop an idea does not grant permission to publish every later representation of it. Scheduled work and retries use the same immutable operation and claim machinery as immediate execution.

## Delivery is not permission

SQS can deliver twice, and a worker can disappear after a provider accepts a request. A missing receipt does not prove that nothing happened. The worker reconstructs current DynamoDB state, checks the exact approved command and claims ownership before crossing the provider boundary.

```text theme={null}
source and operator brief
  → validated proposal
  → authenticated approval
  → immutable command and atomic claim
  → exact provider or artifact operation
  → receipt
  → independent readback
```

A completed claim suppresses duplicate execution. A live claim belongs to its current worker. An expired claim with an unknown effect requires reconciliation. This preserves uncertainty instead of turning it into permission to spend or publish again.

## Native media, durable artifacts

Amazon Transcribe supplies timed speech evidence. Nova Canvas creates images and Nova Reel supports native asynchronous video generation. ElevenLabs provides instrumental music with a thirty-second default. Generated media requires explicit prices, paid admission, media enablement and the applicable approval; the capabilities remain unavailable to normal selection until authenticated readiness is established.

ffmpeg and HyperFrames retain deterministic rendering paths for approved clips and compositions. S3 stores the source and artifact bytes. Content-pack verification reads constituent bytes independently and recomputes digests. Official publishing adapters remain subject to platform credentials, consent, policy and independent readback requirements.

## Proof is part of the workflow

Metadata-only telemetry connects stage, task, approval, claim, cost and verification records without copying prompts, transcripts or private reasoning into traces. AgentCore Memory can contribute scoped advisory context from eligible decisions and verified outcomes; it cannot replace source evidence or current workflow state.

| Evidence level         | What it establishes                             |
| ---------------------- | ----------------------------------------------- |
| Implemented            | A code path exists                              |
| Locally tested         | Specified contracts passed controlled checks    |
| Deployed               | A particular revision exists in an environment  |
| Authenticated success  | Real provider infrastructure completed the call |
| Independently verified | A separate read established the resulting state |

This AWS edition has no authenticated deployment or complete provider workflow proof yet. Local SDK-contract tests and native local-service integration tests do not establish Bedrock quality, generated-media availability, live publication or billed-cost accuracy. All paid calls remain disabled by default.

## Provenance and scope

This is a separate AWS/Strands adaptation of Harmonia's earlier Google-based implementation. It preserves product ideas and useful workflow code while replacing runtime and infrastructure contracts. It does not claim that the AWS adaptation was created during the original contest period, that it satisfies that contest's Google-specific requirements, or that earlier deployment evidence proves the new stack.

The original work drew product ideas from an earlier unsubmitted social-media automation experiment. That history remains relevant disclosure. Current readiness must be established with evidence for this checkout and its actual providers.

See [System architecture](/architecture/overview), [Agent platform](/agent-platform), [Models and costs](/models-cost-evaluation) and [Evaluation](/evaluation).