synthesize

Synthesize Research

Force the researcher to write its final synthesis and conclude, even if it would otherwise continue exploring. If the user names a branch, add <branch>name</branch> to the launch prompt to conclude only that branch (output: .research/branches/<name>/synthesis.md). A main-line synthesis automatically compares all branches.

When to use

• Current results are good enough; lock them in
• The agent is stalling (eval flat/declining for several cycles)
• A deadline forces a conclusion regardless of completeness
• The user wants a checkpoint before pivoting to a related research line

Step 1: Verify .research/ exists

Use the Read tool on .research/state.md. If the file does not exist, tell the user there is no research to synthesize and suggest /research-agent:research to start one.

Step 2: Check whether already synthesized

Use the Glob tool for .research/synthesis.md. If it already exists, ask the user whether they want to overwrite (re-synthesize, which makes sense after /research-agent:resume reopened the concluded run on explicit confirmation) or stop.

Step 3: Launch the researcher with synthesis framing

Spawn the researcher agent (Agent tool, subagent_type research-agent:researcher) with this prompt:

<mode>synthesize</mode>

<working_directory>
{absolute path to current working directory}
</working_directory>

Synthesize your research and conclude. Do NOT run new attempts. Do
NOT start new cycles. The user has decided to lock in current
results.

Step 1: Read .research/state.md and .research/log.md in full to
recall every sub-problem, hypothesis, and attempt.

Step 2: Update .research/state.md with final statuses. Mark every
sub-problem as one of: resolved, abandoned, or unresolved. Mark
every hypothesis as one of: confirmed, refuted, inconclusive.

Step 3: Write .research/synthesis.md per the Termination Protocol
in your system prompt. Required sections: Goal, Outcome (one
paragraph), Key Findings, Failed Approaches, Open Questions,
Recommendations.