| name | pi-progress-synthesis |
| description | Create or revise PI-, collaborator-, or lab-facing scientific progress updates, slide narratives, and research briefs that connect goals, evidence, caveats, and decisions. Use when an agent must synthesize experiments for a scientific audience, prepare a lab meeting or progress review, or turn results into a decision-focused story. |
PI Progress Synthesis
Establish The Audience Question
- Identify the scientific decision, ambiguity, or reframing the audience can help with.
- Assume the audience is scientifically sophisticated but does not remember project-local jargon, run names, or benchmark contracts.
- Define the goal, data or model setup, constraints, and what would count as progress before presenting results.
- Keep the update science-facing. Omit scheduler details, agent operations, debugging chronology, and other process breadcrumbs unless operations are the subject.
Build The Story In Causal Order
Use this sequence unless the evidence requires a different one:
- Problem: why the project exists and which failure matters.
- Baseline: the nearest fair reference or control.
- Mechanism: what changed and why it might address the failure.
- Evidence: quantitative readouts plus representative positive and negative examples.
- Interpretation: what the evidence supports and what remains ambiguous.
- Decision: what should continue, stop, or be tested next.
Put the motivating failure before the proposed solution. When a simpler explanation such as more data, more capacity, or longer training is plausible, name it and explain what evidence separates it from the proposed mechanism.
Make Claims Proportional To Evidence
- Separate observation from interpretation.
- State why the selected baseline is the nearest fair comparison.
- Report denominators, split or source boundaries, and important cohort differences.
- Include negative and broken paths when they changed the decision.
- Name the strongest confound and one observation that would weaken the current interpretation.
- Do not imply that one failed extension falsifies an unchanged successful parent method.
- Prefer “suggests” or “falsifies this mechanism” over a stronger claim unless the evidence supports it.
Put Evidence Next To The Claim
- Show representative visual evidence inline when the claim is spatial, temporal, structural, or qualitative.
- Include both a current positive example and a consequential failure example when sample quality affects the decision.
- State clearly when a positive example is an upper bound, diagnostic bypass, or manually assisted result rather than the deployable path.
- Describe what each figure actually demonstrates; do not substitute the expected theoretical failure for the visible artifact.
- Put a short interpretation beside or below every result figure: what changed, what improved or worsened, and which decision it affects.
- Include artifact paths only when they help the audience inspect the evidence. Omit logs, manifests, configs, and run roots unless audit provenance was requested.
Choose The Right Medium
- Preserve the user’s requested format. If no format is specified, use a short slide deck for visually driven updates and a concise prose brief for primarily conceptual decisions.
- Use the
quarto-presentations skill when creating or revising a Quarto deck.
- Prefer one claim per slide, speaker notes for talk track, and plots over dense metric tables.
- Use tables only when exact mappings or comparisons are clearer than prose or a figure.
- Avoid meta headings about the intended audience or document-making process.
- Translate acronyms and project-local benchmark names into plain language before using them.
Invite Scientific Reframing
- Describe the observed problem, constraints, evidence, and remaining ambiguity before recommending a path.
- Do not present a narrow menu of agent-generated options as though it exhausts the scientific possibilities.
- End with a small set of direct questions, decisions, or falsifiers whose answers would change the plan.
Final Check
Before delivery, verify that the update:
- explains why the work matters before describing the method;
- connects every major claim to evidence;
- distinguishes measured results from interpretation;
- includes the nearest baseline, caveats, and negative evidence;
- makes the decision impact explicit;
- removes low-information process detail; and
- leaves the audience with one memorable status and one clear scientific question.