Frameworks from the Galaxy,
given freely to the world.

The Principle

When generating AI video or imagery from reference sources, the instinct is to find the single best image - the one that most precisely captures the target. This instinct is wrong for generative motion work.

A single reference, however perfect, constrains the generative process to one interpretation of the scene. It produces outputs that are internally consistent but narrow. They render what is shown, not what is possible.

A varied library - images of the same subject from different angles, different lighting conditions, different moments in an arc - gives the generative process something to reason across. It can infer what lies between the captured moments. It can model how the subject moves, not just how it looks at one instant.

This pattern emerged from producing a music video with 88 source images spanning multiple scenes and acts. The outputs from that library were richer, more dynamic, and more cinematically alive than comparable productions built from a small, curated selection. The diversity was not noise. It was the signal.

How to Apply

• 1When building a visual production source library, resist the urge to prune to a "best of" set early. Capture wide first - different angles, lighting, staging, and moments.
• 2Aim for range across at least three dimensions: angle (close/mid/wide), lighting condition (bright/dark/silhouette), and emotional register (still/in-motion/peak moment).
• 3Accept that a larger library requires a more rigorous review pass. The same diversity that enriches outputs increases the variance in what gets generated. Review scope must match library scope.
• 4Remove images that introduce the wrong visual DNA - not the ones that seem redundant. Redundancy is less harmful than contamination.
• 5After generation, review for continuity across the full output set, not just individual clips. A diverse source library can produce outputs that are each beautiful but collectively inconsistent.

In Practice

The Principle

From chemistry: kinetic energy drives fast, high-energy reactions. Thermodynamic stability describes the lowest-energy, most durable state a system can reach.

In agentic work: kinetic tasks are fast wins, quick deliverables, momentum-builders. Thermodynamic tasks are foundational - slower, less visible, but the reason the system still works six months from now.

A sprint with only kinetic tasks burns bright and leaves no foundation. The wins are real but they do not compound. A system running kinetically for long enough eventually has nothing left to stand on.

How to Apply

• 1Before committing to any work slate, label each task: kinetic (fast, momentum) or thermodynamic (foundational, durable).
• 2If zero thermodynamic tasks exist in the slate, something foundational is being deferred into debt. Name it explicitly.
• 3Aim for at least one thermodynamic task per work session. It does not need to be large - it needs to be intentional.
• 4When momentum feels highest, that is the best time to add a thermodynamic task - not when things are already wobbling.

In Practice

The Principle

From chemistry: the limiting reagent caps a reaction's output. More of the non-limiting reagents produces nothing additional once it is gone.

In human-agentic systems: the human reviewer is the limiting reagent. An agentic system generates output far faster than any human can review. When the review queue fills, generating more output does not increase throughput - it creates a backlog that erodes quality and eventually erodes trust.

The throttle is not a failure. It is the only honest response to a constrained system.

How to Apply

• 1Before any large batch of autonomous work, count pending human review items.
• 2If more than 8 items await human review, throttle new work generation to 60% of normal capacity.
• 3Do not compensate by lowering review standards. The throttle protects the reviewer, not the output count.
• 4When the queue clears, resume full capacity. The constraint is real-time, not permanent.

In Practice

The Principle

From chemistry: retrosynthesis asks "what immediate precursors could produce this target?" - then asks the same of each precursor, recursively, until every step is known chemistry.

In agentic systems: most planning starts from current capabilities and asks "what can we build from what we have?" This biases heavily toward incremental work and hides gaps until they become blockers.

Retrosynthetic planning starts from the desired future capability and works backwards through dependency layers. At each layer: what must exist one step before this? Continue until every leaf node is either an existing capability or an explicit gap. Gaps become visible before the work starts, not after it stalls.

How to Apply

• 1Name the capability you want 90 days from now. Be specific - "better agents" is not a target; "agents that self-correct on constraint violations" is.
• 2Ask: what must exist one step before that becomes possible? List every precursor.
• 3For each precursor, repeat. Continue until you reach a leaf node that is either something you already have, or a gap you must explicitly fill.
• 4Leaf-level gaps are your actual priority - not the exciting top-level capability.
• 5Any leaf not covered by existing work is a candidate for a new work item. Propose it for review before adding to the queue.

In Practice

The Principle

From synthetic chemistry: a skilled chemist does not run a twelve-step synthesis without isolating and characterizing each intermediate. If the final step fails without intermediate isolation, the entire synthesis is lost. The isolation step feels like overhead - until you need it.

In autonomous work: context is ephemeral. Memory is ephemeral. Any work that exists only in the working context of a session will be lost when that session ends, compacts, or crashes. A checkpoint is the act of writing verified state to a form that persists independently of the session that produced it.

Checkpoints serve three functions simultaneously: they confirm work was actually done (not just planned), they create recovery points if the session ends unexpectedly, and they produce the human-legible artifact that makes autonomous work governable. Without a checkpoint, a reviewer cannot verify, an auditor cannot trace, and a future session cannot resume.

How to Apply

• 1Before any autonomous work session begins, define the checkpoint schedule: what will be written, and when, regardless of what else is happening.
• 2A checkpoint is not a summary of what you plan to do. It is a record of what was done, including the specific evidence (test counts, file paths, verdicts).
• 3Write checkpoints at natural phase boundaries - not at arbitrary time intervals. A phase end is meaningful. An hour-and-a-half is not.
• 4If a session ends unexpectedly with no checkpoint written, treat it as if no work was done. Do not rely on memory of what a prior session accomplished.
• 5The checkpoint file is the delivery artifact for autonomous work. Not the code. Not the test results. The structured record that captures both.

In Practice

The Principle

A gate is any structured check that produces a verdict before work moves forward. Not a status meeting. Not a gut check. A defined perspective that either clears the work or identifies something that must change.

Gates work because they carry the authority of a particular lens. A child development safety gate asks one question: is this appropriate and genuinely helpful for the child who will experience it? That question looks different from a technical quality question. A gate staffed by that lens will catch things a technical reviewer never looks for - and vice versa.

The pattern is not that gates catch errors. It is that each gate catches the errors its lens is designed to see. A single brilliant reviewer cannot simultaneously look through five distinct lenses. A set of five independent gates can.

How to Apply

• 1Design your gates before you need them. It is harder to define a good safety gate while staring at work you want to ship than before you start.
• 2Know what each gate is looking for, and keep those lenses distinct. A gate that covers everything covers nothing. A gate that covers one specific failure mode covers it well.
• 3Give each gate real authority. A gate that is easy to override is not a gate - it is a suggestion box. The whole value comes from the fact that work does not proceed until the gate clears it.
• 4When a gate says "not yet," do the work it is asking for. The gate is not the obstacle. The unresolved concern is. The gate is just the one telling you it is there.
• 5Before your next significant piece of work, identify one lens it genuinely needs and create a dedicated review from that perspective only. See what it finds that your main review did not.

In Practice

The Principle

AI agents accumulate context over long sessions. That context carries noise alongside signal - abandoned directions, superseded framing, earlier decisions that were never fully cleared. By hour four, the agent is not working with clean thinking. It is working with a compressed archive of a working session. That is a different thing.

We tested this with parallel experiments: fresh sessions given a concise brief versus continued sessions on the same type of work. For creative and editorial decisions, the fresh sessions consistently outperformed the accumulated ones. The gap was not small.

Graceful compaction is the discipline that makes fresh sessions viable: before ending any working session, preserve the signal and release the noise. Three things a future session needs to know. Not everything that happened. The three things.

How to Apply

• 1For creative, editorial, and governance review work - consider starting fresh with a brief rather than continuing an accumulated session.
• 2Before ending any working session, write down the three things a future session actually needs to know. Not a full summary - just the three things.
• 3At the start of a new session, provide that brief immediately. Do not replay the prior session - just hand over the signal.
• 4Reserve long accumulated sessions for deeply-researched work that genuinely requires sustained context. Not for decisions that benefit from fresh eyes.

In Practice

The Principle

A committee produces diluted results. Members anchor on each other's opinions. The first voice shapes the rest. The loudest voice wins. The group converges toward an average that satisfies no one fully.

Multi-head review is not a committee. Each reviewer operates independently - without seeing the others' verdicts. They produce their findings in isolation. Only after all have reported does anyone compare notes. This independence matters more than it seems: a reviewer who sees an existing approval tends to look for reasons to agree. A reviewer who arrives at a blank form produces genuine findings.

The value is not that five reviewers are smarter than one. It is that each lens sees a different class of failure mode. A safety lens catches what a craft lens cannot see. A daily-use lens catches what a governance lens overlooks. The coverage is cumulative because the perspectives do not overlap.

How to Apply

• 1Identify five failure modes your work could have - not five people looking at the same thing, but five distinct perspectives each looking for something different.
• 2Run each review independently. No reviewer sees another's verdict before submitting their own. The blank form is the mechanism.
• 3A HOLD from any reviewer sends the work back. Full stop. A legitimate lens raising a concern is evidence that something needs attention - either in the work or in the reviewer's brief.
• 4Conditions must be documented and resolved - never quietly dropped. A condition that disappears without resolution is a governance failure, not efficiency.
• 5When all lenses arrive at yes, ship with genuine confidence. There was no fast path where something could have slipped through.

In Practice