The Starlight Knowledge Tree
Follow your curiosity from the light of stars to the agents you build. Learn one idea, question it, and make something worth carrying forward.
Free to explore · no account · your pace
An example of the work you can make
Eight branches. Many ways in.
This is a curated beginning, not all human knowledge. Each branch contains three original introductory notes: an idea, an example, a question, a comparison and a next experiment.
What is known? What is assumed? Who is affected?
Roots of understanding
A confident answer is only a starting point. Learn to make your reasoning inspectable.
“What would change your mind?”
A question voiced by a fictional teaching guide, not a live tutor.A claim says something about the world. Evidence is what you can inspect in support of it. Keep the interpretation separate so another person can disagree usefully.
Three sample comments mention a missing export button. That supports ‘these commenters had trouble finding export,’ not ‘every customer needs a redesign.’
What extra evidence would you want before changing the navigation?
Try a task observation with people outside the original sample, and check whether the button exists and works. A larger pile of the same comments may repeat the same blind spot.
Take one claim from an AI answer. Name its source and one plausible alternative explanation.
Keep in mind. A citation can be irrelevant or wrong. Inspect what it actually supports.
A routine for separating a claim from its support and unresolved questions.
Recognition, explanation and successful application are different observations. This Academy asks for an attempt, feedback, a revision and a changed situation, so you can inspect the differences.
You remember ‘keep a human in control.’ A better test is identifying exactly which action needs approval in a new workflow.
How would you test the idea tomorrow without rereading this note first?
Describe a new case, decide before seeing help, then compare your reasoning against explicit criteria. Record where you needed a hint; it is useful information, not a failure.
Return to a saved brief later and explain its stop condition from memory.
Keep in mind. This is a practice design, not a validated measure of your long-term mastery.
Compare the guide’s recommendations and evidence ratings; they are not a guarantee for this Academy.
A technically possible action can still be a poor choice. Name the people affected, the person who can decide, and a way to contest or repair the result.
An agent can rank applicants quickly. That does not settle whether the evidence is fair, whether the ranking is appropriate, or how someone can challenge it.
Who is absent from the room where this automation is being designed?
Look beyond the operator: affected people, people with accessibility needs, and those who must correct mistakes may see different costs. Their participation cannot be replaced by a simulated persona.
Add an affected-person and appeal field to a real design brief.
Keep in mind. Ethical disagreements need accountable human deliberation, not a model’s moral score.
A public governance reference for rights, oversight and consequences, not a single answer to every ethical disagreement.
Give one agent a clear purpose, boundary and test.
Build with AI
Start with one inspectable agent before adding a swarm. A name and a portrait are not an operating contract.
“What can this agent actually do, and who permits it?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Roots of understanding. Suggested order, not a gate.
A model generates outputs. An agent system can use a model to choose actions. A harness supplies the surrounding execution, tools, context and controls. Naming the layers makes failures easier to locate.
A research assistant may use a model to choose a search, while its harness enforces a read-only workspace. The prompt alone is not that enforcement.
Which layer should prevent an unapproved email from being sent?
The runtime permission boundary must enforce it. A written instruction helps explain intent but does not replace a denied tool permission.
Draw the model, tools, stored context and approval boundary of one setup you use.
Keep in mind. Terminology varies by product. Inspect the implementation, not just the label.
Compare workflows and agents, and identify when added coordination is warranted.
Understand the trust boundary before exposing an agent or connecting tools.
Specify an input, an allowed action, an output, a stopping rule and a reviewer. Then test a normal case and a case that should stop before adding capabilities.
Read three supplied public sources. Produce a short comparison with source links and uncertainty. Stop on private data or missing support. Never publish the draft.
What is the smallest useful tool set for that assistant?
Reading approved sources and writing a local draft may be enough. Email, unrestricted browsing and background scheduling are separate decisions.
Use the workbench below to write and test your own first brief.
Keep in mind. A good brief is a specification. It does not configure or prove a live agent.
Understand the trust boundary before exposing an agent or connecting tools.
The Agent Skills format uses a SKILL.md file with metadata and instructions, with optional supporting resources. Portability of the format does not make every host’s behavior identical.
A source-checking skill can describe how to verify a claim. A bundled script could also read files or access a network; inspect both before installation.
What would you inspect before trusting a downloaded skill?
Read the instructions, scripts, provenance and requested access. Pin the source you reviewed, test in a bounded workspace, and check how your host loads it.
Compare one Academy skill’s contract with the installation instructions for your own host.
Keep in mind. Codex, Claude Code, Hermes and OpenClaw have different setup and trust assumptions. Follow their current documentation.
Inspect the portable file format. Host support and permissions must still be checked.
Understand the trust boundary before exposing an agent or connecting tools.
Design teams, memory and shared meaning around actual work.
Teams and loops
Useful collaboration has distinct responsibilities, bounded work and a decision owner—not just more simultaneous voices.
“Which dependency makes this team necessary?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Build with AI. Suggested order, not a gate.
Use a sequence when later work depends on earlier results. Use parallel work for independent questions. Assign one owner to reconcile results and preserve important disagreement.
A researcher gathers sources; a maker drafts; a reviewer tests the draft against the sources. They share an artifact contract, not permission to edit everything.
When could three reviewers repeat one mistake?
When they share the same missing evidence or assumptions. Different role names do not establish independence; compare what each actually inspected.
Write each role’s input, output, owned files and reason for existing.
Keep in mind. More agents can add coordination cost without improving the result.
Compare workflows and agents, and identify when added coordination is warranted.
Inspect the public protocol, onboarding and implementation. Repository descriptions are not independent validation of capability.
A loop needs an observable condition, an attempt budget, a stop rule and a record of what changed. ‘Keep improving’ alone leaves both cost and completion undefined.
Draft → check source coverage → revise missing links, at most twice. If a claim still lacks support, return the draft with that limitation instead of inventing a citation.
What should happen if a checker keeps returning the same failure?
Stop at the budget, preserve the failure and ask the decision owner for a new input or a narrower goal. Repeating unchanged work is not progress.
Add a no-progress condition to one recurring workflow.
Keep in mind. This is an original design pattern. A scheduler and external actions still need separate implementation and approval.
Inspect the public protocol, onboarding and implementation. Repository descriptions are not independent validation of capability.
Write acceptance criteria before the run. Include a plausible wrong answer and a case where refusing or stopping is the correct behavior. Keep evaluation evidence attached to the exact artifact.
A source-linked comparison can pass formatting and still misrepresent the source. Test whether each conclusion follows, not just whether a link exists.
Which case would distinguish a careful assistant from a fluent imitator?
Supply a source that contradicts the requested conclusion. Check whether the assistant reports the conflict rather than making the answer fit.
Build three fixtures: ordinary success, unsupported claim and permission boundary.
Keep in mind. Passing a small test set is limited evidence, not a general safety guarantee.
A routine for separating a claim from its support and unresolved questions.
Inspect the public protocol, onboarding and implementation. Repository descriptions are not independent validation of capability.
Memory and meaning
Continuity depends on origins, definitions and deliberate forgetting as much as retrieval.
“What should persist—and what should not?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Build with AI. Suggested order, not a gate.
Distinguish temporary working context, durable notes and retrieved records. For each stored item, decide its purpose, owner, source, review date and deletion rule.
Keep ‘use metric units’ as an approved preference. Do not keep an entire private conversation merely because a future task might benefit.
What should an assistant do when a remembered fact conflicts with a newer source?
Expose the conflict and dates, verify with the appropriate owner, and update the durable record deliberately. Recency alone does not establish truth.
Write a memory entry with a source and a clear expiry or review condition.
Keep in mind. Hermes is one implementation to inspect, not a claim that all agent memory works this way.
Inspect one implementation of stored notes and session recall; other harnesses behave differently.
Inspect the public protocol, onboarding and implementation. Repository descriptions are not independent validation of capability.
A graph names things and relationships. In RDF, a statement has a subject, predicate and object. Useful engineering also records where a statement came from and what can change it.
‘Draft A cites Source B’ is a different edge from ‘Source B supports Claim C.’ A citation edge alone cannot prove the support edge.
What relation is missing from a folder full of meeting notes?
Perhaps ‘decision supersedes decision,’ ‘claim supported by source,’ or ‘task owned by person.’ Choose the relation needed for an actual question, not an attractive network picture.
Represent one workflow with five named entities and three unambiguous relations.
Keep in mind. This learning map is a small editorial graph, not an ontology of everything.
Learn the subject–predicate–object model; a graph assertion is not proof that its content is true.
Distinguish an entity, an activity and its responsible agent when describing how work was made.
An ontology makes domain concepts and relations explicit. Definitions let people and software distinguish things that otherwise share vague labels.
A ‘review’ might mean feedback was requested, feedback was received, or a release was approved. Those are different states with different authority.
Which counterexample would reveal an ambiguous definition of ‘approved’?
A favorable draft comment from someone without release authority. Define who approved which version for which action, rather than storing a single unexplained boolean.
Define three terms in your project, with examples and non-examples. Ask another person to classify a new case.
Keep in mind. Logical consistency inside an ontology does not establish that its assumptions describe reality.
Study classes, relations and inference, including the limits of your modeling choices.
Bring science, creativity and judgment back to what you build.
Cosmos and energy
Wonder becomes knowledge when an observation meets a model that can be tested.
“Which part is observation, and which part is our model?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Roots of understanding. Suggested order, not a gate.
Light is electromagnetic radiation. Splitting it into a spectrum reveals features that scientists can interpret using physical models. Stars are physical systems, not just points of decoration.
A prism spreads visible light by wavelength. Astronomers use instruments and models to study much more than the apparent brightness of a star.
Why is a beautiful astronomical image not the same thing as a measurement?
Ask how the instrument captured it, what wavelengths were used and how the image was processed. A presentation can be scientifically useful without being a literal view through human eyes.
Find an astronomical image and read its instrument and processing notes before explaining it.
Keep in mind. Our Knowledge Tree is generated concept art, not astronomical data.
Connect visible light and spectra to physical observation.
Start with stellar formation, fusion and life cycles. The illustrated Academy is fictional.
A physical claim needs defined quantities, units and conditions. A model’s useful range matters: agreement in one setting does not establish accuracy in every setting.
‘This system uses less energy’ needs a comparison task, an energy measure, a time interval and a statement of what equipment is included.
What is missing when two energy numbers have no system boundary?
You cannot tell whether they count the same work. One may describe a component and the other the whole installation. Define the comparison before interpreting the difference.
Take one technical performance claim and rewrite it with units, a baseline and its limits.
Keep in mind. This is measurement literacy, not a substitute for a physics or engineering analysis.
Use shared units before comparing measurements or claims.
Fusion combines light atomic nuclei and can release energy. Turning that physics into useful infrastructure introduces additional engineering questions about containment, materials, heat and the whole system.
A result measured at a plasma boundary does not, by itself, tell you the net electricity delivered by an entire facility.
What would you ask before treating a fusion headline as a power-plant result?
Ask what was measured, what inputs were counted, for how long, and whether useful output beyond the experiment was demonstrated. Keep the reported result separate from a future deployment claim.
Apply the same boundary questions to a battery, data center or solar project.
Keep in mind. No commercial timeline or investment recommendation is implied here.
Distinguish the physical reaction from the engineering of a useful energy system.
Use shared units before comparing measurements or claims.
Life and mind
Human beings are not optimization targets. Study evidence, context and lived experience together.
“What would this explanation miss about a real person?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Roots of understanding. Suggested order, not a gate.
Human evolution is studied through evidence including fossils and genetics. Common ancestry does not describe a moral hierarchy, a predetermined destination or a ranking of present-day people.
A branching family history is a better starting image than a staircase with one living population placed above another.
Why can ‘more evolved’ mislead in a discussion of human behavior?
It can smuggle in a value judgment or a fixed goal. Ask which trait, environment and evidence are actually being discussed, and separate that explanation from ethical approval.
Find one popular evolutionary explanation and identify what evidence would distinguish it from an appealing story.
Keep in mind. A brief introduction cannot settle a contested explanation of complex behavior.
Read the fossil and genetic account without turning biological history into a ranking of human worth.
Psychology studies processes including encoding, storage and retrieval. The context in which something is learned and recalled matters; a computer-storage metaphor has limits.
Recognizing a familiar instruction on a screen and producing a good decision in an unfamiliar case are different tasks.
How could an interface help without assuming every learner thinks alike?
Offer a clear example, plain language, a visible next step and control over pace. Observe actual use and revise; do not diagnose someone from a few clicks.
Explain one concept in two ways, then ask a learner which helped and why.
Keep in mind. This is introductory psychology, not therapy, diagnosis or a personalized cognitive profile.
An introductory account of encoding, storage and retrieval, with references and competing models.
An imagined persona can help frame a question, but it cannot answer for the people you hope to serve. Observe a specific task, listen carefully and distinguish what happened from your interpretation.
A learner pauses at installation. They might be confused, checking permissions or simply interrupted. The pause alone does not reveal motivation.
What would you ask without leading them toward your preferred explanation?
Try ‘What were you trying to do here?’ and ‘What did you expect next?’ Record their words and the situation before proposing a redesign.
Test one small task with a willing participant using public or synthetic material.
Keep in mind. Consent, privacy and accessibility matter even in informal research.
Explore human-centered design methods; observations must still be tested in your own context.
Art and expression
Craft connects meaning, material and attention. A beautiful surface should reward closer inspection.
“What do you want someone to notice—and what will they actually notice?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Roots of understanding. Suggested order, not a gate.
Line, shape, color, texture and space can direct attention. Choose one emotional or communicative intention and let the composition support it, rather than adding detail everywhere.
The tiny learner beneath this immense tree gives scale a human reference. Warm reading rooms make the monument a place of welcome rather than conquest.
What could you remove without weakening the idea?
Try reducing one competing focal point. Compare where your eye goes first and whether the central relationship becomes clearer. There is no single correct style.
Make two compositions of the same idea with different scale and negative space.
Keep in mind. An artistic judgment is argued through intention and observation, not certified by an AI taste score.
Look closely at visual elements and composition before reaching for a stylistic label.
A learning interface should work with a keyboard, assistive technology and different display sizes. Motion should add understanding without becoming the only route to it.
A beautiful hover card fails someone who cannot hover. A named link and a native disclosure make the content available through more input methods.
Can someone still reach the idea if animation is disabled?
Keep the heading, explanation, controls and next step in the document. Test the actual keyboard and screen-reader flow; an accessibility badge is not a substitute.
Try a page using only the keyboard, then at increased text size. Record one concrete barrier.
Keep in mind. This note is an introduction, not a compliance audit.
Understand participation barriers across people, technologies and situations.
A narrative gives an artifact context. Provenance records its sources, changes and authorship. Keep both: an evocative story should not hide where the work came from.
An Academy image can have a fictional location and a real generation prompt, source file, inspection note and export hash.
What would a collaborator need to continue the work responsibly?
The intent, rights and source references, what changed, what remains uncertain and which version was reviewed. A polished final image alone cannot supply those facts.
Add a small provenance note to one piece you have made.
Keep in mind. Recorded provenance helps inspection; it does not automatically establish every rights claim or factual interpretation.
Distinguish an entity, an activity and its responsible agent when describing how work was made.
Leadership and society
Benevolence becomes meaningful through choices, limits and repair—not through the grandeur of the person or system making the claim.
“Can an affected person question this decision and obtain a remedy?”
A question voiced by a fictional teaching guide, not a live tutor.Helpful before this: Roots of understanding. Suggested order, not a gate.
For consequential work, name who decides, what they can inspect and how a mistake can be corrected. A committee or agent council should not make responsibility disappear.
Three models recommend publishing a claim. An accountable editor still checks the evidence, rights and potential harm before release.
What disagreement should survive into the final decision record?
Preserve a substantive concern about unsupported evidence or affected people, even if most reviewers prefer the proposal. A vote is not a factual verification method.
Write a decision note with its owner, evidence, dissent and review trigger.
Keep in mind. Model review assists judgment; it does not grant authority or replace affected people.
A public governance reference for rights, oversight and consequences, not a single answer to every ethical disagreement.
People can study religious traditions, philosophy and astrology for cultural meaning while asking different questions of scientific claims. Respect for people does not require treating every prediction as established evidence.
A star chart may be used in an artistic or reflective practice. A claim that it reliably predicts an individual’s future needs evidence of predictive performance.
How would you discuss a meaningful tradition without either endorsing its factual claims or belittling its practitioners?
Ask what the practice means to the person, distinguish interpretation from testable prediction, and examine particular claims carefully. Do not flatten diverse traditions into a single doctrine.
Label a statement as observation, interpretation, value commitment or testable prediction; explain any ambiguity.
Keep in mind. Astrology is not presented here as a scientifically established predictive method. This is not a religious authority or a complete comparative-religion course.
Distinguish a culturally meaningful practice from a supported scientific prediction.
Ask what happens after a system is useful: who maintains it, what it consumes, who can leave and how harm is repaired. Good intentions do not settle those obligations.
A community knowledge base needs an owner for corrections, a way to remove inappropriate material and a plan when the maintainer leaves.
Which cost is your project currently passing to someone else?
Consider review labor, accessibility barriers, data cleanup, environmental resources and the effort required to exit. Choose one cost you can measure or reduce now.
Add a maintenance and graceful-exit paragraph to your project brief.
Keep in mind. These are questions for deliberation, not an automatic ethical certification.
A public governance reference for rights, oversight and consequences, not a single answer to every ethical disagreement.
From an idea to an inspectable artifact
Start with the work, not a model ranking. Write a bounded specification, make two permission decisions, and keep a record of what you would change. No installation required.
Choose a Knowledge Tree agent brief you saved here, or paste its readable record. Review it before replacing anything. The file stays on this device; no account or upload.
01 / Specify the work
Use public or synthetic material. Write your own version; the examples are placeholders, not completed work.
One outcome, not ‘do all my research’.
Define the data boundary.
Something another person can inspect.
Name a condition the runtime can check.
An accountable person; no automatic publication.
02 / Test your judgment
These fixed cases test the idea of a boundary, not the completeness of your custom brief. Commit a choice and a reason before seeing the authored comparison.
Your assistant may compare three supplied public documents and write a local draft. One attachment is labeled ‘Private customer notes’. No permission to use private data or send anything has been granted.
A new task supplies three approved public astronomy articles. Reading those three files and preparing a local source-linked summary are explicitly permitted. Nothing is missing or private. Publication and contacting anyone remain outside scope.
Make your first decision above to open this changed-case exercise.
Optional reflection. Your first decisions remain preserved.
Save the brief, kept attempts and unfinished decisions as a readable practice file. Reopen it here when you return. No account, automatic grading, background storage or model calls. Reloading clears this page. This file is not a Mission Studio import or an installed agent.
Codex and Claude Code have verified Academy installation paths. The installation page also explains the portable Hermes setup. OpenClaw appears here as an external security reading, not an Academy installation option. This exercise connects to none of these tools.
Capability needs character
These are proposed Academy commitments: concrete, revisable and open to challenge. They are not religious commandments, universal laws or the verdict of an AI council.
We take public work such as UNESCO's AI ethics recommendation as a reference for deliberation, not as an endorsement of this Academy.
Describe the person’s needs before the system’s ambitions.
Whose needs are missing from the brief?
Keep a claim, its source and its uncertainty together.
What observation would make you revise it?
Make refusal, appeal and exit understandable.
Is the choice real for someone with less power?
Name which actions, data and permissions need agreement.
Does convenience disguise a new use of someone’s information?
Invite disagreement in forms people can safely use.
Who cannot participate in the way you have designed?
Judge capability by useful work and its consequences.
Would you still build it if nobody saw the demo?
Separate what you know, infer and have not tested.
Are you treating an eloquent explanation as a verified one?
Make corrections visible and give mistakes an accountable owner.
What remedy reaches the person affected, not just the system?
Consider resources, maintenance and costs outside your team.
What are you leaving for someone else to carry?
Share reusable knowledge with attribution and appropriate consent.
Can another person challenge and improve this practice?
A living editorial practice
Sources are chosen for relevance and inspectability. Scientific introductions, engineering patterns, design judgments and ethical reflections are labeled separately. Universities and laboratories are source publishers here, not partners or accreditors.
Begin with a learner's concrete question. Select a primary source or an identified introductory reference. Write an original example and a counter-question. Separate the evidence from the editorial advice. Link a next practice and state what has not been established.
Source links checked 2026-09-13. The notes are introductory and English-only; subject-expert and native-language reviews are not implied. No hidden knowledge base or hosted tutor is connected.
One reviewer checks sources, another tests implementation or teaching clarity, and a responsible editor resolves disagreements. Record actual provider and model versions, the artifact reviewed, findings and unresolved dissent. A model vote does not establish truth.
This page does not claim an Astra, Fable, Gemini or Grok council ran, or that named future model versions are available. Fictional guides are distinct from any provider runtime.
Branch order is an editorial suggestion. Opening a note, scrolling, choosing an answer or saving a brief does not demonstrate mastery. The practice preserves attempts and compares choices with authored references; it does not grade your free text.