---
source_hash: 812261005e37815209783a8e8e189655e649bd3192da70dda8f3286382d4c1b5
---
# Build Your Team: Agent Context Pack

## Outcome

Define a workflow by its observable work and handoffs before assigning owners. This pack supports the public essay, Build Your Team: The Work Comes First. It is a portable method, not a scheduler, model ranking, or permission grant.

## Stage Contract

For each stage, identify its accepted input and revision, prerequisites, output artefact, next receiver, required judgement, acceptance check, owner, and approval boundary. Describe the evidence the receiver needs to accept that output. Share relevant agreed artefacts, not private chat transcripts or unrelated context.

Separate three acceptance claims:

1. Structure: The output exists, parses, and has the declared shape.
2. Relationship: It uses the accepted input and revision its contract requires.
3. Outcome: It serves the next receiver and the intended result.

A check proves its bounded claim. Valid structure alone cannot prove semantic correctness, timing, or fitness for the brief.

## Dependency Order

The illustrative video sequence is Script → Voice → Timed Transcript → Animation → Publish. It assumes that animation follows a reviewed voice recording and timings derived from that recording. It is not a universal video-production sequence.

A stage requires an accepted prerequisite, not merely an earlier box in a list. If a prerequisite is unusable, its dependent outputs are also unusable. Revision relationships matter separately from order: changing accepted audio can make a previously valid transcript and animation stale. Retain previous artefacts while reviewing a replacement. Parallelise work only when it has accepted inputs it can use independently.

## Ownership And Oversight

A deterministic script can perform explicit transformations and structural checks. A model can propose interpretations or outputs. A human defines the outcome, accepts editorial judgements, and authorises public release. Record which operations an owner may perform and which decisions need approval. Choosing an owner does not grant permissions.

Access comes from application controls and tool boundaries. Markdown, CSV, or database storage alone does not determine what a model can read or change.

## Intermediate Checks

The essay's lab simulates a Voice pronunciation defect or a Timed Transcript offset. A pronunciation listen can catch the former; a transcript timing comparison can catch the latter; an animation preview can inspect the relevant speech or sync issue. An unrelated earlier check must not be credited with detecting a defect it does not inspect. Final human review remains required.

When a check fails, hold acceptance at that stage, preserve the failed output for diagnosis, return a concrete failure to its owner, and review dependent work after the accepted input changes. The simulation models these selected defects only. It establishes no measured performance benefit and certifies no whole production.

## Start With One Stage

Use the [Stage And Owner Worksheet](https://aris-space.com/downloads/build-your-team/stage-owner-worksheet.md). Define one observable outcome, its deliverable, and receiver. Write one stage contract, then follow its output into the next stage. Check what becomes stale when an accepted input changes and who is allowed to accept or release the result.

Read the [full essay](https://aris-space.com/documents/ari-os/build-your-team) for its authored discussion and interactive examples.