Supporting
AI-assisted editorial workflow
A human-in-the-loop editorial system that turns engineering artifacts into weekly field reports through planning, retrieval, drafting, critique, verification, and publication.
Source of truth
Notion — idea metadata
Repo — article body
DEV.to — live post
Skill files — agent behavior
Humans remain in the loop for revision and final publish.
Publishing thoughtful engineering writing every week is less about pressing a publish button and more about keeping voice, evidence, and judgment intact while agents help with capture, drafting, and critique.
A longer style rule did not stabilize voice; score-first review produced QA feedback when the draft needed editorial judgment.
I designed and built the AI-assisted editorial workflows — planning, retrieval, drafting, critique, verification, and publication — as reusable role-based skills with explicit contracts, while developing production software and publishing the field-report series they produce.
Ideas live as Notion cards; article bodies live as repo markdown; DEV.to is the publish surface.
Operator skills cover the lifecycle Inbox → Candidate → Drafting → Published: capture/triage, weekly schedule (one Drafting card at a time), bounded context retrieval, draft generation, adversarial critique, and draft sync — with humans editing markdown and approving irreversible publish steps.
Lifecycle
Inbox → Candidate
Candidate → Drafting
Drafting → Published
One Drafting card at a time.
Editorial field-report pipeline
Capture → triage → schedule → refresh → context → draft → critique → sync → human publish. Full editorial skill chain.
Talk trackWalk the full operator skills: Capture grounds Notion Inbox cards; Triage promotes Candidates; Schedule picks one Drafting card; Refresh is optional freshness; Context gathers evidence before Draft; Critique analyzes before it scores; Sync only creates a DEV draft — humans own live publish.
Node detail
no node selected
Select a node
Eight operator skills and one public output
Each node owns one contract in the chain. Select one to read its summary and evidence; press it again, or click the canvas, to close.
Ownership is explicit: Notion wins for idea metadata, the repo wins for article body, DEV.to wins for the live post, and skill files win for agent behavior.
Cover-image review is an explicit gate after a post shipped without one. Cadence defaults to one thoughtful DEV post per week.
Contract: Humans remain in the loop for revision and final publish.
Explicit gates
Cover-image review
Human revision
Final publish
No editorial KPI is claimed on this page.
Retrieval, generation, and critique are separate stages rather than one prompt.
Critique analyzes before it scores. Voice training prefers a strong structural exemplar over an ever-growing style encyclopedia. MCP-backed Notion and GitHub integrations supply grounded context without letting agents silently rewrite the source of truth.
The system continuously produces a public weekly engineering field-report series on DEV — covering agents, LLMs, evaluation, analytics, and AI-assisted engineering practices from production software workflows — through structured context retrieval, critique, and evidence-based refinement rather than one-off chatbot drafting.
Artifacts and related writing
The published series this pipeline produces.
Field report — the voice-training decision in block 06.
Field report — the critique stage in block 04.
Field report — the reasoning behind the critique contract.
Public field reports linked above. Prefer those artifacts over invented editorial KPIs. · Governance built on the same role-based pattern — Renovate governance ladder →