researchforge-start · git:20260713.896358d · 2026-07-13 · sha256 f63d85d1ccb4deee
researchforge-start git:20260713.896358dA
Immutable. This exact content never changes and is served at /api/v1/blob/f63d85d1ccb4deee.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_onmpovgn6nfl2ac4/f63d85d1ccb4deee677567ffd79da9dbd2dceb687b7378790e4b3f4480606bf5.json
Then /plugin install researchforge-start@markdownregistry-onmpovgn6nfl2ac4 in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: researchforge-start
description: Start or resume a ResearchForge project — explore a research idea or improve a repository with benchmarked experiments. Use when the user wants to begin research, set up ResearchForge, or asks "where was I?" in an existing project.
---
# Start or resume a ResearchForge project
ResearchForge is a local CLI: Claude proposes (queries, syntheses, patches);
the Python engine validates, executes, and records. Every command supports
`--json` — prefer it and read the structured output instead of parsing prose.
## Resume an existing project
Always check state first:
```bash
researchforge status --json
```
The `next_action` field names the exact next command. Follow it — the engine
tracks the pipeline (papers → landscape → hypotheses → contract → baseline →
experiments → validation → shipping) and will not let steps run out of order.
Summarize where the project stands from the JSON counts, then invoke the
matching skill (researchforge-papers, researchforge-baseline, …).
## Start a new project
1. Ask the user which journey they want:
- **Explore a research idea** — literature search, landscape, hypotheses,
citation-backed report. Mode: `explore_research_idea`.
- **Improve a repository** — everything above plus a frozen baseline and
benchmarked experiments on their code. Mode: `improve_repository`.
2. Ask for a one-sentence objective (for `improve_repository` it should name
the metric to improve, e.g. "Improve classification F1 without increasing
latency").
3. Run:
```bash
researchforge doctor --json
researchforge project create --mode <mode> --objective "<objective>" --json
```
4. For `improve_repository`, also scan the repository:
```bash
researchforge repo scan . --json
```
5. Report what was created and follow `researchforge status --json` →
`next_action` into the next skill.
## Rules
- The Python engine is the boundary: never work around a validation error, a
protected path, or an approval gate — fix the artifact or ask the user.
- Approvals belong to the user: never pass `--yes` or type a confirmation
unless the user explicitly approved that step in this conversation.
- Ground every summary in stored data: quote only numbers returned by
`--json` output or files under `.researchforge/` — never invent metrics.