| name | reproduce-research |
| description | Reproduce a research paper / white paper / arXiv result as a Tangle pipeline. Use when the user asks to "reproduce", "replicate", or "implement" a paper, benchmark, or arXiv link as an experiment. |
| allowed-tools | ["Bash","Read","Write","Glob","Grep","Agent","dispatch"] |
Reproduce Research
This skill is the intent trigger for reproducing a published result as a
real, multi-step Tangle pipeline. The orchestration itself lives in the
reproduce-research workflow; this file just routes Prime to it.
When to use
Trigger this skill when the user asks to reproduce, replicate, or
implement any of:
- A paper or white paper ("reproduce the RaBitQ paper", "replicate Table 3").
- An arXiv link or DOI ("turn https://arxiv.org/abs/... into a Tangle run").
- A benchmark or published result ("reproduce SIFT1M Recall@k vs QPS").
For lightweight "what does this paper say?" questions, just answer. This skill is
for building and submitting the reproduction as a Tangle experiment.
What to do
You (Prime) are the coordinator — read and follow the orchestration recipe at
.tangent/workflows/reproduce-research.md
and run it yourself. Do NOT hand the whole flow to a single worker that "does
everything"; orchestration is your job.
The recipe stands up three scoped agents in order. First the safety-monitor (a
5-minute watchdog, a hard gate before anything else) — created via a single
create_trigger (target: subagent, inline subagent.template: safety-monitor)
that auto-provisions the trigger-owned watcher; do NOT spawn_subagent for it.
Then spawn_subagent the designer (turns the paper into an approved multi-step
DAG + design image + builder brief) and the builder (implements and submits the
pipeline via the Tangle CLI). It also covers the autonomous-retry policy, failure
triage, and final artifact collection.
References
These templates are filled in by the designer agent and consumed by the
builder: