I built SharePlane because I was losing too much good work.
Not losing it in the literal sense. The conversations usually still existed. The research was somewhere in a chat history. The reasoning, model comparisons, evaluation results, discarded approaches, architectural decisions, and half-finished arguments had all been saved by whatever tool I happened to be using.
Technically, the work was there.
Operationally, it was gone.
When I needed it again, I often could not find the right session. When I found the session, I could not always locate the important part inside it. When I located the important part, I still had to reconstruct what had been decided, what evidence supported it, what remained unresolved, and what was supposed to happen next.
So I regenerated the work.
Sometimes the replacement was better. Sometimes it was not as good as the thinking I had already done. Either way, I was repeatedly paying for work that had already been completed.
That is what I mean by work evaporation.
Work evaporation occurs when useful intellectual work continues to exist somewhere, but can no longer be reliably found, understood, trusted, continued, or reused when it matters.
Stored is not the same as recoverable.
Recoverable is not the same as reusable.
And a transcript is not the same as an artifact.
That distinction became foundational to SharePlane.
This started as my failure
I want to be direct about responsibility.
AI did not create this problem for me. Nobody else did either. My personal organization system was not good enough.
I did a great deal of research and systems thinking, but I did not consistently convert that work into durable, structured outputs. I relied too heavily on memory, filenames, search, scattered notes, and the comforting fiction that I would remember where everything was later.
That worked poorly before AI.
AI made the failure impossible to ignore.
The reason was simple. It gave me more hands.
Ideas that previously would have taken weeks to research, test, write, and operationalize could suddenly move much faster. I could contrast one agent against another. I could challenge an answer with a different model. I could build evaluation harnesses, test competing interpretations, explore implementation paths, and refine the result through several rounds without requiring a team for every step.
That changed what I could realistically attempt.
It did not change how much I could remember.
Research in defined work settings has found that generative AI can increase individual productivity, although the effects vary substantially by worker and task. In one field study of 5,172 customer-support agents, access to an AI assistant increased issues resolved per hour by an average of 15 percent, with different effects across experience levels. (arxiv.org)
My claim is not that everyone becomes 15 percent more productive, or that customer support represents every form of knowledge work. It does not.
The relevant point is narrower: AI can materially increase the rate at which a person produces potentially valuable work.
When production rises, preservation becomes a scaling problem.
A weak personal knowledge system may lose a few notes when output is low. Increase the number of research threads, drafts, decisions, prompts, evaluations, evidence sets, agent trajectories, and implementation artifacts, and the same weakness starts destroying real value.
AI did not merely give me better answers.
It increased the volume of work that needed to survive.
Chat history is not a knowledge system
Most AI tools preserve conversations. That is useful.
It is also nowhere near sufficient.
A conversation is working context. It contains exploration, repetition, misunderstandings, dead ends, corrections, provisional language, abandoned structures, and conclusions that may have changed three times before the session ended.
That mess is not a defect. It is often what serious thinking looks like while it is happening.
The mistake is treating the record of that thinking as the finished knowledge product.
We do this because the transcript feels complete. Every word is sitting there, so nothing appears to have been lost. This is the digital equivalent of keeping every paper that has ever crossed a desk and declaring the filing problem solved.
The hard question is not whether the information exists.
The hard question is whether another person, another agent, or your future self can determine:
- what was concluded;
- why it was concluded;
- which evidence supports it;
- which alternatives were rejected;
- what remains uncertain;
- which version is authoritative;
- what may be reused;
- what must happen next.
When those answers remain trapped inside a long conversation, the work has not been preserved. It has merely been retained.
Research on computational reproducibility demonstrates the same broader distinction. CORE-Bench tested whether AI agents could reproduce scientific results using provided code and data. The best evaluated agent achieved only 21 percent accuracy on the hardest tasks. The materials existed, but existence alone did not make the work operationally reproducible. (arxiv.org)
Personal AI-assisted work is not scientific computing, and I am not pretending the benchmark directly measures chat retrieval. The architectural lesson still holds.
Possession of the components does not guarantee recoverability of the result.
The reasoning machinery matters
I do not accept everything an AI model tells me at face value.
Fluency is not evidence. Confidence is not correctness. A polished answer may still be built on a bad assumption, weak sourcing, hidden compression, or a complete misunderstanding of the actual problem.
So I contrast agents.
I test one interpretation against another. I ask models to challenge claims, identify missing evidence, expose contradictions, and evaluate proposed outputs. When the work is important enough, I build harnesses and explicit acceptance criteria around it.
That process is part of the intellectual asset.
Agent-evaluation researchers have identified major problems with current benchmarks, including excessive focus on accuracy, inadequate holdout sets, benchmark overfitting, conflation of different user needs, and poor standardization. They argue that these weaknesses impair reproducibility and make it harder to determine whether an agent is useful in real applications. (arxiv.org)
The conclusion is not that every conversation needs an elaborate laboratory protocol.
It is that the final answer often tells only part of the story.
Sometimes the useful artifact includes:
- the claim that survived testing;
- the evidence that changed the conclusion;
- the agents that disagreed;
- the criteria used to evaluate them;
- the approach that failed;
- the reason it failed;
- the unresolved risk that should remain visible.
That material should not be dumped indiscriminately into the public article. Readers do not need to crawl through every abandoned thought merely because storage is cheap and restraint apparently went out of fashion.
But the machinery should be preserved when it is necessary to trust, reproduce, challenge, or extend the work.
That is structured intellectual provenance.
The argument, the evidence, and the provenance
Reader-oriented meaning.
Support and constraints.
How the work became trustworthy.
A serious SharePlane artifact can expose three distinct layers.
The argument
This is the reader-oriented article, presentation, recommendation, design, or operating model.
It should be clear. It should have a governing thesis. It should respect the reader’s time. It should not force someone to reconstruct the argument from the debris field that produced it.
The evidence
This includes sources, observations, tests, comparisons, and findings that support or constrain the argument.
Evidence should perform a known job. It should prove, complicate, bound, or operationalize a claim. A large bibliography is not a substitute for reasoning.
The provenance
This records how the work developed.
It may include model contrasts, evaluation methods, decisions, rejected alternatives, transformations, authorities, hashes, dependencies, and unresolved questions.
Work on computational pipelines has similarly argued that capturing the information required to reconstruct processing is important for reproducibility and trust, especially as automated systems produce outputs that cannot all be inspected manually. (arxiv.org)
SharePlane applies that principle to AI-assisted intellectual work.
The public artifact should remain readable.
The evidence should remain inspectable.
The provenance should remain available.
Those are different jobs, and combining them into one undifferentiated transcript serves none of them particularly well.
Documentation is not cleanup anymore
Most people treat documentation as something that happens after the real work.
That model fails when the work is moving faster than a person can reliably classify and recover it.
Extraction must become part of execution.
A meaningful session is not complete merely because the thinking stopped or the model produced an answer. It is complete when the durable value has been identified and placed where it belongs.
That may include:
- approved public copy;
- atomic claims;
- supporting evidence;
- architectural decisions;
- evaluation results;
- reusable prompts;
- implementation constraints;
- unresolved risks;
- next actions;
- links to related work;
- the authority governing the current state.
This cannot rely entirely on someone remembering to tidy things up later.
Later is where knowledge goes to die politely.
The extraction has to become mechanical enough that the system does not depend on memory, energy, mood, or whether five other ideas arrived before the current one was properly booked.
The goal is not to preserve every token.
The goal is to preserve everything required to trust, continue, or reuse the work.
The transcript is evidence. The repository is authority.
This principle now governs how I think about SharePlane.
The transcript remains useful. It shows how an idea developed. It can preserve provenance, explain a decision, resolve a dispute, or support an audit.
But the transcript should not remain authoritative merely because it contains the original conversation.
Authority belongs in a governed system.
For me, that is a repository-backed structure where artifacts can be versioned, linked, validated, reviewed, hashed, and advanced through explicit stages.
The repository is not authoritative because Git possesses mystical knowledge-management properties. Left unattended, a repository can become an extremely precise landfill.
Authority comes from governance:
- named states;
- explicit ownership;
- traceable decisions;
- evidence binding;
- version control;
- validation;
- controlled transitions;
- durable relationships between artifacts.
That is the difference between storing files and operating a knowledge system.
The transcript supports the work.
The repository governs what the work has become.
What SharePlane is actually for
SharePlane publishes articles, but publishing is only the visible surface.
Its deeper purpose is to convert temporary AI-assisted cognition into governed, durable, inspectable, and reusable knowledge.
It exists because the productive capacity of an individual is changing faster than the systems most individuals use to manage what they produce.
AI can now help one person perform work that once required more time, more specialized support, or a small team. That does not make the individual an institution. It does mean that the individual may start producing institution-scale context without institution-scale memory, governance, or continuity.
That gap matters.
Without a system, more capacity can create more waste.
More research can mean more forgotten research.
More agents can mean more untraceable conclusions.
More output can mean more duplication.
More speed can mean losing important decisions faster.
SharePlane is my attempt to close that gap.
It takes the reasoning, evidence, decisions, evaluations, and artifacts produced during temporary sessions and turns them into work that can survive independently of those sessions.
That helps me avoid rebuilding what I have already built.
It also allows other people to inspect, challenge, adapt, and reuse the work. That matters to me. I get genuine satisfaction from helping people think more deeply and avoid repeating work that has already been done.
But usefulness to others depends on doing more than publishing conclusions.
The method has to survive too.
The design test
Can the important work be continued without reopening the original conversation?
The real test for SharePlane is not whether it creates an attractive page.
It is this:
At the end of a meaningful working session, can the important decisions, evidence, methods, artifacts, risks, and next actions be recovered and continued without reopening the original conversation?
If the answer is no, the workflow is incomplete.
The transcript may remain forever.
That does not mean the work survived.
AI gave us more hands.
Now we need systems capable of remembering what those hands produced.