Help us improve
Share bugs, ideas, or general feedback.
From research
Executes non-interactive end-to-end research pipeline: auto-configures program.md with defaults, runs judge+refine loop up to 3 iterations, then runs campaign from single goal command.
npx claudepluginhub borda/ai-rig --plugin researchHow this skill is triggered — by the user, by Claude, or both
Slash command
/research:sweepThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
<objective>
7-step setup wizard that produces a complete, ready-to-run research.md without executing the research loop. Walks the user through goal, metric, search space, constraints, evaluator design, and baseline measurement, then writes the file. TRIGGER when: user wants to set up a research project; user wants to plan before running the loop; user says "plan my research"; user has a goal but no research.md; user invokes /autoresearch:plan. DO NOT TRIGGER when: research.md already exists and the user wants to run the loop; user wants a one-shot answer; user wants to debug, not optimize.
Orchestrates full research pipeline from Brainstorming to Reporting via Planning, Implementation, Testing & Visualization phases with user checkpoints. Configurable for physics, AI/ML, statistics, math domains, depth, and agent personas.
Generates program.md for autonomous AI research experiments (Karpathy's autoresearch). Interviews user on codebase, metrics, constraints; explores code; tailors agent instructions from template.
Share bugs, ideas, or general feedback.
Non-interactive end-to-end research pipeline: auto-plan → judge gate → run. Single command from goal to result. Accepts a goal string and passes through all run/colab/team flags.
NOT for: interactive planning (use /research:plan); methodology review only (use /research:judge); running an already-approved plan (use /research:run).
Triggered by sweep "goal" [--flags]. Non-interactive end-to-end pipeline: auto-plan → judge gate → run.
Task tracking: create tasks for S1–S5 at start.
Extract <goal> — the first positional argument (quoted or unquoted string describing the optimization target).
Extract flags and their values:
--colab[=HW] — passed through to plan (Config.compute) and run; if =HW present, extract colab_hw--compute=local|colab|docker — passed through--team — passed through to run--codex — passed through to run--researcher — passed through to run; combine with --architect for dual-agent SOTA + architectural hypothesis pipeline (--journal and --hypothesis are not available in sweep mode)--architect — passed through to run; enables architectural hypothesis pass via solution-architect--skip-validation — passed to judge step (S3) to skip local metric/guard validation--out <path> — optional: write program.md to this path instead of project rootIf <goal> is missing or empty, stop:
⚠ sweep requires a goal prompt.
Usage: /research:sweep "goal description" [--flags]
Run plan mode steps P-P2 and P-P3 from plugins/research/skills/plan/SKILL.md (P-P0 profiling flow skipped — <goal> is always a text string; P-P1 scope guard skipped — goal was provided explicitly) with these behavioral overrides:
sweep: auto-config → but do NOT wait for user confirmation.--colab[=HW] or --compute=colab was passed, write compute: colab (and colab_hw: <HW> if provided) into the Config block.<--out path> if provided; otherwise to program.md at project root.
<path>.bak (overwrite any existing .bak), then proceed — no interactive confirmation in sweep mode.Print on completion:
sweep: plan → <output path> ✓
Initialize REFINE_ITER = 0, MAX_REFINE = 3.
Repeat up to MAX_REFINE times:
Increment REFINE_ITER. Run judge mode (J1–J6 from plugins/research/skills/judge/SKILL.md) against the program file.
--skip-validation if the user provided it; otherwise include validation (J4).JUDGE_REPORT).Print: sweep: judge iteration `REFINE_ITER`/`MAX_REFINE` → `VERDICT`
If APPROVED — exit loop with outcome approved.
If BLOCKED — exit loop with outcome blocked. Do not attempt to fix — BLOCKED means a fundamental design flaw that requires human redesign.
If NEEDS-REVISION:
REFINE_ITER < MAX_REFINE:
JUDGE_REPORT. Extract the ### Required Changes section.N_FIXES.sweep: applied N_FIXES fix(es) to <program path> — re-judgingREFINE_ITER == MAX_REFINE — exit loop with outcome unresolved.Safety net: the
.bakfile created in S2 is the undo path — edits applied during the loop modifyprogram.mdin place.
| Outcome | Action |
|---|---|
approved | Print sweep: plan approved (REFINE_ITER/MAX_REFINE iteration(s)) ✓ → proceed to S5 |
blocked | Print sweep: judge → BLOCKED ✗; show all critical findings from the report; print follow-up hint; stop |
unresolved | Print sweep: judge unresolved after MAX_REFINE iterations ✗; show remaining Required Changes from the last report; ask user with options: (a) proceed to run anyway (b) fix manually then re-run (c) abort — if a, proceed to S5; if b or c, print follow-up hint and stop |
Follow-up hint (blocked or unresolved):
Fix the issues above in <program path>, then:
/research:judge <program path> ← re-validate
/research:run <program path> ← run when approved
/research:sweep "revised goal" [flags] ← re-sweep from scratch
Run Default Mode (R1–R7 from plugins/research/skills/run/SKILL.md) against the program file from S2, passing through all flags:
--colab[=HW] / --compute--team--codex--researcher / --architect (combine both for dual-agent pipeline)Note:
--journaland--hypothesisare not available in sweep mode (see S1).
--teamand interactivity: when--teamis passed, sweep becomes semi-interactive — run mode's Phase B presents a user confirmation gate before Phase C begins. The gate cannot be bypassed from sweep context; sweep will pause and wait for user input at that point. This is expected behavior.
On completion, the standard R6 terminal summary is printed. Additionally, prepend:
sweep: complete — plan → judge → run pipeline finished