| name | research-seed-expansion |
| description | Expand from one or more seed papers to find relevant prior work, follow-up work, and adjacent papers around a topic. Use when Codex needs to start from known papers instead of a fully bounded corpus, trace citations or other relation signals outward, maintain explicit provenance from each discovered paper back to the seeds, rank the resulting reading map, and stop based on a documented saturation argument rather than pretending to cover the whole literature. |
Research Seed Expansion
Overview
Expand outward from one or more seed papers without losing the trail back to why each discovered paper matters. Start by locking the seed set and expansion plan, build an expansion ledger, traverse relation signals in controlled hops, then report a ranked reading map with provenance and a clear saturation argument.
Do not pretend this workflow proves global coverage of a field. This skill is for seed-driven exploration, so the hard requirements are provenance, disciplined traversal, explicit boundaries, and an honest explanation of why the expansion can stop.
Workflow
- Write the expansion plan before traversing.
- Record the topic, seed papers, why these seeds were chosen, expansion goal, allowed relation types, hop budget, and stop rule.
- Decide whether you care more about prior work, follow-up work, adjacent methods, or a balanced map.
- Read seed-plan.md and use its template.
- Verify and lock the seed set.
- Normalize the seed metadata and make sure each seed is the paper you actually intend to expand from.
- Freeze the initial seed set before collecting neighbors.
- If you later add a new seed, mark it explicitly as a scope change instead of smuggling it into the graph.
- Build the expansion ledger before free-form reading.
- Create a ledger artifact in the task workspace. Use CSV or JSON for anything beyond a small expansion.
- Give each paper a stable row and capture how it was discovered.
- Record at least one provenance path from every non-seed paper back to a seed.
- Read expansion-ledger.md.
- Traverse outward in controlled passes.
- Perform
hop-0 on the seeds themselves to confirm what problem, method, and lineage they represent.
- Perform
hop-1 from the seeds across the allowed relation types.
- Perform
hop-2 or deeper only from high-signal nodes or when the user explicitly wants a broader map.
- Read traversal-and-ranking.md.
- Use automation for collection, not for judgment.
- Use scripts to fetch citation lists, normalize metadata, de-duplicate candidates, and cache abstracts or PDFs.
- Do not let scripts decide which papers are genuinely relevant, seminal, or central to the topic.
- Use sub-agents only after the ledger exists.
- Keep seed locking, relation-choice policy, de-duplication, merge, and final saturation judgment with the main agent.
- Assign sub-agents disjoint frontier slices or relation surfaces with fixed output fields.
- Require one structured result per assigned ledger row.
- Read subagent-protocol.md.
- Stop on saturation, not exhaustion.
- Decide whether the frontier is still producing genuinely new signal.
- Look for repeated references, converging lineages, stable top papers, and declining novelty in new candidates.
- Read saturation-and-reporting.md.
- Report the reading map with provenance and stop rationale.
- Return the seed set, expansion boundary, ledger summary, ranked results, lineage notes, and saturation argument.
- Do not claim that the full literature has been covered unless the user separately bounded the corpus.
Operating Rules
- Freeze the seed set early. Any later seed addition is a visible scope change.
- Every non-seed paper must have explicit provenance back to at least one seed.
- Prefer the smallest useful hop budget. Uncontrolled expansion turns into aimless graph walking.
- Use scripts to mechanize fetching and normalization, not to replace semantic reading.
- Distinguish
relevant, adjacent, and not worth further pursuit; do not flatten the graph into one undifferentiated list.
- If the expansion is still producing new high-signal branches, say it is not saturated yet.
Read These References When Needed
