Atlas

Carry the context all the way through.

A feature rarely arrives as a complete writing brief. Its behavior may be described in a work item, clarified in a conversation, revised in a specification, and already mentioned in several articles. The difficult part is connecting those pieces and keeping the release decision current.

Atlas follows a feature from intake through research, drafting, affected-document planning, staging, and handoff. A release-wide relationship pass identifies shared announcements and related features before the individual feature runs begin. Each run leaves a saved dossier and an explicit next action.

The public repository makes the mechanics inspectable and includes an offline demonstration. Explore Atlas on GitHub →

Give every decision somewhere to live.

The feature dossier connects shipping scope, exclusions, product facts, dates, affected articles, and open questions to their sources. It records which sources were reached and which were unavailable. Extracted facts, inferences, and ambiguous claims remain distinct.

The agent reads and interprets the source material. Python validates and saves the resulting record, merges later findings without duplicating existing entries, and renders a readable Markdown companion. Conflicting positions stay visible until someone resolves them.

Find the articles the feature actually touches.

Atlas combines repository search with document-relationship evidence to build a deduplicated set of affected articles. It considers title and filename matches, platform relevance, article type, and agreement between discovery sources before preparing an edit plan.

The planner separates a verified change in an owned article from a plausible connection that needs a writer’s judgment. It also avoids treating a widely linked overview page as a strong edit candidate simply because it connects to many topics.

The planner produces routes and batches. Content editing belongs to the agent’s execution stage. That separation lets the proposed footprint be tested and inspected before an article changes.

AI for interpretation. Code for repeatable mechanics.

Atlas uses AI during the workflow. The agent follows an orchestration skill to collect context, interpret source material, draft content, and assess review signals. Supporting scripts handle the operations that should behave consistently for the same inputs.

Supporting utilities handle related-feature grouping, release moves, transcript parsing, identifier resolution, and repeat-notification checks. Source collection depends on the host’s actual access; inaccessible material remains an explicit gap. Harvested text is treated as untrusted source content.

Make the state explain the next step.

• Check the document before trusting the status field.Lifecycle resolution checks published and staged content alongside tracking signals. A stale checkbox should not trigger a second draft of something that has already shipped.
• Keep approval attached to the current situation.The agent’s review instructions account for later pauses and withdrawn approvals. An old sign-off is not enough to move a held feature forward.
• Make reruns accumulate useful context.Dossier merges, prior-feature notes, and assignment records preserve findings and avoid repeating the same bookkeeping. A reassignment can be distinguished from an unchanged assignment.
• Give uncertainty a destination.Conflicts, unresolved owners, and weakly supported article connections become explicit review or handoff items, with the reason retained.

Refined by the cases that should stop.

The original implementation combined deterministic tests with supervised, read-only research and simulated end-to-end runs. Those runs exercised already-published work, held features, conflicting release signals, unavailable evidence, and affected-document discovery.

They led to concrete fixes: refreshing changed feature details on reruns, reconciling duplicate source labels, correcting document-path handling, and narrowing proposed edits to stronger evidence. A historical approval replay also checked the transition from draft to ready-to-stage against the recorded workflow.

Atlas remained a tested prototype with production writes disabled. The deterministic core is directly testable; the agent’s judgments depend on its host, instructions, and available evidence. Unattended production publishing and a standalone background service were not established.

Run the workflow mechanics. Inspect the result.

The public edition brings the reusable core, orchestration instructions, tests, and fictional examples into a standalone repository. Employer-specific configurations, internal records, and private telemetry queries are excluded.

Its offline demonstration follows a fictional feature through saved research, lifecycle resolution, document planning, and handoff. The resulting dossier, plan, and run report let you inspect how the pieces connect. The sample release draft is fixture content; this demonstration makes no model calls or network requests.

python scripts/demo.py --output demo-output
python tests/run_all.py

View Atlas source and quickstart →

Atlas extends the documentation work described in Release and content tools. It also connects to a recurring concern in Personal Canon: preserve the context behind a decision so the next session can use it.