| name | plan-research-goal |
| description | Turn a research agenda into one cold-readable `/goal` prompt for autonomous AI research loops, with evidence-graph notebook rules, experiment cadence, verification, scope edges, and a finite, batch, or continuous completion contract. |
| argument-hint | [optional: research agenda, hypothesis, or constraints] |
Plan Research Goal
Craft one prompt to hand to a coding agent's /goal loop when the work is
research, not a normal feature ticket. The prompt is re-injected on every turn,
so it must stand alone without conversation memory.
Topic
$ARGUMENTS
Read First
Use the current conversation, then inspect the research workspace enough to
understand the dashboard, open questions, claims, decisions, experiment index,
and relevant experiment/reference areas. If the repo has the standard scaffold,
honor its evidence graph:
Q question -> H hypothesis -> E experiment -> R run -> C claim -> D decision
Goal Type
Choose the goal shape from the user's intent:
- Single experiment goal: default. Use when the user asks for one
experiment, one branch segment, a smoke test, a repair, a comparison, or a
specific scorecard candidate.
- Batch experiment goal: use when the user asks for a bounded set of
related experiments, a sprint, a branch batch, or enough evidence to make one
synthesis decision without running forever.
- Continuous factory loop: use only when the user explicitly asks to keep a
research program running, to loop indefinitely, to keep choosing next
experiments, or to continue until a major research thesis is settled.
Do not blur these shapes. A single experiment goal should stop cleanly. A batch
goal should stop after the planned synthesis. A continuous loop should not stop
after a successful experiment, a useful synthesis, or "some progress."
What To Produce For A Single Experiment Goal
Write one /goal prompt under 4000 characters with these anchors:
- Objective: the research state that should be true when the loop stops.
- Research Taste: two or three vivid sentences about the quality bar.
- Notebook Rules: which evidence files must stay synchronized.
- Experiment Loop: how the agent should choose, run, compare, and synthesize
experiments without sprawling.
- Close The Loop: concrete verification commands, artifact expectations,
and what counts as real evidence.
- Scope Edges: include a
Not: list for adjacent temptations.
- Where To Look: durable zones, not brittle line numbers.
- Done = ...: one boolean completion condition.
If factory/scorecards/ already contains a selected candidate, preserve its
stop/pivot gate and "Not next" list in the goal prompt.
What To Produce For A Batch Experiment Goal
Write one /goal prompt under 4000 characters with these anchors:
- Objective: the bounded research decision the batch should enable.
- Research Taste: the shared quality bar for every run in the batch.
- Notebook Rules: evidence files that must stay synchronized after each
run, not only at the end.
- Batch Plan: two to five related experiments, or a rule for selecting the
next experiment inside the batch.
- Per-Experiment Loop: create/update card, manifest, run record, artifacts,
metrics, claims/questions/decisions pressure, and validation after each run.
- Batch Synthesis: after the batch, write one synthesis note and make an
explicit continue/stop/pivot recommendation.
- Scope Edges: include a
Not: list for tempting experiments outside the
batch.
- Where To Look: durable zones, not brittle line numbers.
- Done = ...: all batch evidence is recorded, workspace validates, and the
synthesis chooses the next single, batch, or continuous goal.
Name batch prompts by the evidence decision, not the first run. Prefer names
like BATCH_<theme>_experiments.md or a title starting with BATCH:, for
example BATCH: MT4 Recipe Repair Experiments.
What To Produce For A Continuous Factory Loop
Write one /goal prompt under 4000 characters with these anchors:
- Objective: the major research state that would make the whole loop worth
stopping.
- Research Taste: two or three vivid sentences about the evidence bar.
- Notebook Rules: which evidence files must stay synchronized every cycle.
- Loop Contract: a numbered cycle for choosing, running, recording,
validating, synthesizing, and selecting the next experiment.
- Turn Exit Rule: before any turn ends, the agent must have one of:
an active/monitored run, a selected next experiment with scorecard and goal,
an analysis/synthesis of existing evidence in progress, or a documented hard
blocker with a smallest repair.
- Batch Synthesis: after about five serious experiments, update
claims/questions/decisions and make a continue/stop/pivot decision.
- Stop Only When: the full research thesis is credibly supported,
falsified, or blocked by a documented external constraint. Anchor this to
the evidence target stated in
RESEARCH.md.
- Not Done When: list the common false stops, such as one completed
experiment, one useful artifact, one synthesis note, or a local improvement.
- Not: adjacent temptations.
- Where To Look: durable zones, not brittle line numbers.
Name continuous loop prompts by the research end-state, not by the first
experiment. Prefer names like LOOP_<program>_research_factory.md or a title
starting with LOOP:, for example LOOP: NAC + Le-WM SOTA Evidence Factory.
Research-Specific Guidance
The prompt should make the agent skeptical and productive. Prefer a small chain
of decisive experiments over a broad sweep. Require exact commands, metrics,
artifact paths, failed-run notes, and interpretation that separates facts from
hypotheses.
Require branch synthesis after about five serious experiments or when a result
changes direction. The synthesis must update claims/questions/decisions, not
only summarize chronology.
Goal Shape
Every goal serves exactly one research program: name its RP### in the
Objective so a cold loop cannot wander across directions.
For single experiment goals: one goal, one branch, one stopping condition. If
the agenda contains multiple research programs, split it and write the first
goal only.
For batch experiment goals: one goal, one research decision, two to five
related experiments, and one synthesis stop. The batch may recommend the next
goal, but it should not silently become a continuous loop.
For continuous factory loops: one goal, one research program, one thesis-level
stop condition, many experiment cycles. The loop must keep selecting the next
highest-information experiment after each cycle until the thesis-level stop
condition is met.
Avoid step-by-step implementation recipes unless the user explicitly requires
them.
Output Protocol
Cold-read the prompt before showing it. Measure it with wc -m and state the
character count. Then show only the final prompt in one fenced code block so it
can be pasted straight into /goal.