vetra LABVersion 1.0 · 14 September 2026
AI-assisted editorial resource prepared during site development. Not generated by the deployed Atlas, Kite or Echo workflow. These are proposed practices, not a tested reliability guarantee.
Purpose
Give the next person or model enough evidence to continue a small research task without guessing what happened. Start with one question and one inspectable output.
Before starting
- Write the research question in one sentence.
- Define the output: a checklist, comparison, short report, or reproducible example.
- Set the allowed sources and actions. State what is out of scope.
- Define the stop condition and who may approve publication.
Atlas / research handoff
- Assign each source an ID; record its title, URL and retrieval date.
- For each factual claim, point to the relevant source section.
- Separate observations, interpretations and unverified assumptions.
- Record missing evidence and conflicting findings.
- Pass a concise evidence summary rather than an unstructured conversation.
Kite / artifact handoff
- Produce the actual document, not only a plan to write it.
- Keep source IDs next to claims that depend on them.
- Label original recommendations as recommendations.
- Record the artifact version, changes and unresolved questions.
- Describe only actions actually performed; a proposed test is not a passed test.
Echo / review handoff
- Open the cited sources and compare them to the claims.
- Check that the output answers the question and can be used as delivered.
- Identify contradictions, unsupported assertions and missing limitations.
- Record each check as passed, failed or not checked, with a short explanation.
- Return ready for human review or needs revision; do not silently treat failed checks as approval.
Copyable handoff template
Task:
Owner:
Artifact / version:
Status: draft | needs revision | ready for human review
Sources: ID | title | URL | retrieved date | relevant section
Findings: claim | source IDs | uncertainty
Actions actually performed:
Recommendations:
Review: check | passed / failed / not checked | evidence
Open questions:
Next action:
Background reading
Anthropic distinguishes predefined workflows from systems where models choose their own next actions. Its prompt-chaining pattern passes one stage's output into the next and can include intermediate checks. This supports describing VETRA's implementation as a bounded workflow, not an autonomous society.
Anthropic: Building effective agents ↗
MCP provides a protocol for exchanging context and exposing capabilities between clients and servers. It does not itself decide how models collaborate. A shared protocol therefore does not substitute for a task definition, review process or artifact history.
Limitations
This kit is an editorial proposal. It has not been benchmarked against agent workflows. Source correctness, model quality and human review remain separate concerns. Do not include secrets or personal information in a publicly shared handoff.