Day 0 source-closed attempt
“The agent chooses a tool and uses it.”
Missing connection
How the tool result is checked
socratink
Learning by reconstruction
Build the idea.
See what holds.
Close the source before answering. Socratink preserves your explanation, then compares what you can rebuild when you return.
RFCs, API docs, papers, lectures, or notes · Continue as a guest
Use it before a design review, your first PR in an unfamiliar codebase, or an on-call rotation.
One connection · two source-closed attempts
Your own explanation, compared.
Day 3 source-closed attempt
“The agent uses a tool, checks the result against the goal, then continues, retries, or stops.”
Rebuilt on Day 3
Tool use → verification → decision
- Exact words preserved
- Source hidden first
- Compared after spacing
How a source-closed explanation becomes evidence
First, the source is available as context. Then it closes. The prompt asks who checks the tool result before the next step. The learner writes that an agent gets a goal, chooses a tool, and uses it to act. That answer shows tool use was recalled, but it omits the verifier that the prompt asked for. Updating the plan was not asked in this attempt, so it remains untested.
The source closes before your answer starts.
Your own explanation is what counts. Socratink records where the connection breaks, opens study around that break, then decides what should return.
Bring your source.
Use material you already care about: a paper, a chapter, a lecture, or your own notes.
Explain it closed.
Rebuild one connection before the source or AI opens. That first answer becomes the evidence baseline.
Return after spacing.
Study only what was missing, then rebuild the connection later without the source or AI.
What the record can honestly say
You viewed it.
Reading and AI explanations can help, but they do not show what you can produce.
You rebuilt it.
An explanation made from memory becomes evidence of what changed.
See what you can explain without help.
Your map changes only when you reconstruct. Reading and AI explanations stay context, never evidence.