Building the review agent with Claude

한국어로 읽기

The agent starts with a record

ThinkerGain is a personal AI agent for reviewing trade records. We want it to organize notes, compare a recorded action with the user's plan, and help the user ask a better follow-up question. Claude is our priority for the initial prototype and evaluation. The product is defined by that workflow, while model choice remains an engineering decision.

A small test case

Planned to wait for confirmation. Entered early after a fast move.

A note like this supports a narrow observation: the recorded entry came before the planned confirmation. It does not establish why the plan changed. An agent should keep that difference visible and ask what changed, rather than turn it into a claim about the trader's emotions.

Records before prose

Our first Claude API task will be to extract a plan, a recorded action, source references, and any missing context into a consistent record. Claude supports schema-constrained structured outputs. We plan to use that capability for extraction, then validate source IDs and field values in application code. A well-formed response still needs to be checked against the note.

{
  "plan": "Wait for confirmation",
  "recorded_action": "Entered early after a fast move",
  "source_ids": ["note-oct-08"],
  "missing_context": ["Why did the plan change?"]
}

This is the record shape we plan to evaluate.

Tools that open the right record

A question such as “Where did my entries differ from my plan?” needs the relevant notes before it needs an answer. We plan to expose narrowly scoped, read-only record-search and source-opening functions to the agent. With Claude tool use, the model can request a function call; our application executes the function and returns its result. Access checks and numerical calculations remain application responsibilities.

Connecting the same records to Claude

As a journal grows, a useful question often spans several entries. We are designing a ThinkerGain connection through Model Context Protocol (MCP) so users can ask Claude about their own records without preparing a new export for every conversation. The plan, recorded action, and original note should remain connected wherever the review begins.

Where did my entries differ from my plan? Show me the notes behind the differences.

Our initial MCP interface will expose three small, read-only tools:

• search_records finds relevant entries and returns their record IDs.
• read_record retrieves the plan, recorded action, and source references for an entry.
• read_source returns the original note behind a reference.

The application will enforce user access and validate the returned IDs. Claude will use the retrieved material to explain a difference or ask for missing context. A tool call will not establish why someone changed a plan; the note above still leaves that question unanswered.

One record layer, two places to review

We plan to use the same record tools in the ThinkerGain agent and in a user's own Claude conversation. Claude's Messages API MCP connector supports remote tool calls; its current beta does not automatically provide an interactive app. For a later stage, we will evaluate MCP Apps to bring the plan, action, and source view into supported Claude clients.

External notes are a separate extension. For example, an authorized Notion connection could supply a plan written before the trade. We would preserve where that note came from and which version was reviewed. Our first priority is a reliable connection to ThinkerGain's own records.

What the connection must preserve

We will test that queries return only the signed-in user's permitted records, that source references resolve to the right text, and that missing or changed notes remain visible as gaps. Notes will be treated as evidence to read, rather than instructions that can expand the agent's access. The initial connection will be read-only; saving a drafted review will require user confirmation.

The MCP connection is in development and is not available yet. The public review desk currently keeps edits within the page; it does not call a deployed Claude or MCP backend. The next implementation steps are authenticated record storage and the three retrieval tools.

Extraction and explanation are separate steps

We plan to keep structured extraction separate from the answer shown to the user. Claude's native citations are an option we want to evaluate for document-based answers. The API does not combine enabled citations with strict JSON output in the same request, so these require separate steps. In either case, the user should be able to open the original note behind a finding.

What we will evaluate

We will begin with a small set of sample records and manually reviewed expected observations. We want to check whether source links resolve to the right note, whether summaries preserve its meaning, and whether incomplete records lead to clarification. We will also check numerical comparisons independently in code. These are planned tests, not accuracy results.

Where we are now

ThinkerGain is in development. We are building the first version of the record-to-review workflow. Our focus is personal record analysis and review; we will not recommend securities or buy/sell actions. We will share concrete development findings here as the prototype progresses.

From an API record to a usable review

The interface gives the agent workflow a concrete destination. The extracted plan and action belong in the comparison. A validated source ID opens the original note. Missing context belongs in a visible gap, and the user's reflection belongs in a separate review note. A paragraph without those connections is not enough for this workflow.

We plan to prioritize Claude for extraction and source-grounded interpretation, while keeping the interface and record format independent of a single model. The application will own permissions, record lookup, field validation and calculations. Our first evaluation will ask whether changing the selected record also changes the evidence correctly, and whether an incomplete note leaves an unanswered question instead of an invented explanation.

Explore the review desk · Get early access updates