The Craft of AI Collaboration

Siyao Lan begins an AI collaboration with a question that needs to be made clear. He establishes why it deserves attention, which facts are supported, and what evidence would change the decision. For a paper, he asks whether its central contribution is a method, theory, or perspective. For engineering, he turns the goal into inputs, dependencies, and deliverable behavior. The scale of the question can change while responsibility for direction, standards, and tradeoffs remains with the researcher.

That presence appears in specific feedback. A missing condition sends the work back to its assumptions; a formulaic chapter sends it back to character actions and the reader’s experience; an unsuitable interface prompts a change in project structure. AI proposals enter the discussion, and human judgment determines what deserves to continue. The collaboration has a goal that can be revised and an author who keeps asking questions.

Siyao’s AI collaboration spans Codex and Claude Code. Reading, ideation, implementation, review, and expression are organized as clear responsibilities across platforms, accumulating collaboration at a thousand-agent scale. Within this practice, retained local Codex records from April to October 2026 contain 1,681 native subagent contexts with recorded activity. One context reads material, another implements a module, and a fresh context reviews the output under a different responsibility; the same model can take on different roles.

Within a single task, Siyao can direct and coordinate hundreds of agents around one goal, organizing reading, exploration, implementation, and review. Efficiency comes from clear task boundaries and dependencies: independent work advances in parallel, separate roles check key judgments, and artifacts and findings converge around the shared goal. Scale becomes coordinated action while the researcher retains direction and judgment.

At this scale, memory and handoffs become everyday work. The next participant needs to know where the question came from, which attempts failed, and which decisions the current version retains. Siyao gradually organizes collaboration into bounded tasks and records that let each contribution connect to earlier work. The number counts deduplicated contexts accumulated over time. It describes the scale of collaboration, while the resulting work is checked on its own terms.

He divides reading, ideation, implementation, writing, figures, and checking into responsibilities that can be owned. Independent tasks proceed in parallel; tasks depending on the same evidence are handed off in order. A producer delivers an artifact, and a reviewer returns to the original material to judge it. This division brings different perspectives to a question and sends a repair back to the affected location.

A harness is the coordination process that holds these handoffs: inputs, tasks, original drafts, versions, and reviews remain available, including where to resume after an interruption. A loop carries an unresolved question into the next round. A counterexample may change an assumption, a review may call for another comparison, and a failure may justify stopping a direction. Siyao designs the relationships between these responsibilities so the next action has an understandable reason.

The relationships between tasks determine how a complex project can be understood. Siyao organizes work through modules and dependencies: which observation supports a claim, which interfaces a change affects, and which missing evidence blocks a decision. Graphs make these connections visible, helping structure reuse, checks, and revision. Workbenches and workflows are gradually connecting these research objects while retaining their sources and versions.

Automated experiments follow the same relationships. Establish the observation a hypothesis requires, choose a suitable reference, hold the necessary conditions consistent, and ask whether the result distinguishes the explanations. In bounded research loops, a program can run successfully while the experiment remains insufficient to advance a scientific claim. Retaining that outcome lets the next round reconsider the question, experiment, and decision together. Human judgment continues to enter the process.

Through repeated reading, writing, and review, he organizes recurring requirements into skills: which material a task needs, how its expression should be structured, what must be checked, and when another role should take over. Rules are revised through actual tasks, and relevant parts are loaded when needed. Experience gains a reusable entry point while leaving room for fresh judgment.

This work includes his own designs and the study, adaptation, and integration of existing tools and research methods. Siyao checks the actual entry point and version, retaining provenance, reasons for changes, and task feedback. When InkTide was initially given a display website, he redirected it into a project for native tools and prioritized Codex implementation. The decision places collaboration close to its real use: the entry point itself is part of the design.

Collaboration takes shape in tools and creative work that can be opened, used, and inspected. Ariadne develops new research directions into proposals that retain hypotheses, pilots, and unresolved questions; InkTide connects materials, original drafts, revisions, and checks. Each tool has its own goals and means of validation, and its value develops through actual use and repeated improvement.

Engineering roles face execution and use. Sentinel focuses on the relationship between the final action and its execution context, with an open-source developer preview and inspectable software evidence delivered. InkTide connects author materials, original drafts, revisions, and checks. Through these tasks, Siyao develops his collaboration practice, bringing ideas into systems and retaining behavior and records that can be reviewed.

CIVILIZATION brings the collaboration into another form of creation. Characters need concrete desires and actions; chapters need varied scenes, causality, and rhythm; a long-form plan continues to change through writing. The music written for the novel, The Door That Stays, has its own lyrics, emotion, and listening experience. AI helps organize material and develop expression while the author continues to decide which people and feelings the story should reach.

Siyao hopes the collaboration produces a deliverable and a clearer understanding of why an assumption holds, which evidence changed the decision, and where an elegant expression lost a necessary condition. That learning calls for questions, revisions, and personal choices. A new understanding can return to the next round and help frame a question worth answering.

Unmet goals also remain part of the work. For InkTide to exceed a stage-matched strong baseline, it has to face specific comparisons and failures, then test the proposed changes afresh. This sustained judgment and revision is the craft expressed by “The Craft of AI Collaboration”: a person sets direction, AI extends the work, and evidence helps both see the next step. Maturity develops through concrete choices and continues in the researcher’s understanding and the author’s eye.