talont-org/prepare-research-pr

Prepare Research PR

Reads a completed experiment report and plan, synthesizes a directional recommendation, selects 1-2 exp-lens lenses, writes a context file per lens, and writes a PR prep file for downstream use by compose-research-pr. Does NOT invoke any exp-lens skills or create a PR.

Arguments

/autoskillit:prepare-research-pr {report_path} {experiment_plan_path} {worktree_path} {base_branch}

• report_path — Absolute path to research/{slug}/README.md (post-rename by commit_research_artifacts)
• experiment_plan_path — Absolute path to the experiment plan (from ${{ context.experiment_plan }})
• worktree_path — Worktree root directory
• base_branch — PR target branch

Critical Constraints

NEVER:

• Invoke exp-lens skills — they are run in separate sessions by the recipe orchestrator
• Create files outside {{AUTOSKILLIT_TEMP}}/prepare-research-pr/ (relative to the current working directory)
• Fail silently — always emit all three output tokens as your final output
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use Agent subagents (not slash commands) for reading and synthesis
• Emit prep_path, selected_lenses, and lens_context_paths as your final output
• Write all output paths as absolute paths

Lens Selection Table

| experiment_type | Primary Lens | Secondary Lens (optional) | |---|---|---| | benchmark | exp-lens-fair-comparison | exp-lens-estimand-clarity | | configuration_study | exp-lens-iterative-learning | exp-lens-sensitivity-robustness | | causal_inference | exp-lens-causal-assumptions | exp-lens-estimand-clarity | | robustness_audit | exp-lens-severity-testing | exp-lens-validity-threats | | exploratory | exp-lens-estimand-clarity | exp-lens-exploratory-confirmatory |

Experiment Status Badges

• CONCLUSIVE_POSITIVE — hypothesis supported; results meet success criteria
• CONCLUSIVE_NEGATIVE — hypothesis refuted; results clearly below threshold
• INCONCLUSIVE — results ambiguous; insufficient signal or mixed evidence
• FAILED — experiment could not be completed; results invalid

---

Workflow

Step 0: Parse arguments and create temp dir

Parse positional args:

• arg[1] = report_path
• arg[2] = experiment_plan_path
• arg[3] = worktree_path
• arg[4] = base_branch

Derive feature_branch:

FEATURE_BRANCH=$(git -C "{worktree_path}" rev-parse --abbrev-ref HEAD)

Create temp directory:

mkdir -p {{AUTOSKILLIT_TEMP}}/prepare-research-pr/

Generate a timestamp ts (format: YYYY-MM-DD_HHMMSS) for unique file naming.

Step 1: Read report via Agent subagent

Spawn an Explore subagent to read {report_path} and extract:

• Executive summary
• Conclusions
• Recommendations
• Results section (findings table, key metrics)
• Experiment status badge (CONCLUSIVE_POSITIVE / CONCLUSIVE_NEGATIVE / INCONCLUSIVE / FAILED)
• Title from the first H1 heading or title: field

Store extracted content as report_content.

Step 2: Read experiment plan via Agent subagent

Spawn an Explore subagent to read {experiment_plan_path} and extract:

• experiment_type (benchmark / configuration_study / causal_inference / robustness_audit / exploratory)
• H0 (null hypothesis)
• H1 (alternative hypothesis)
• Independent Variables (IV) table
• Dependent Variables / Metrics (DV) table
• Controlled Variables list
• Success Criteria
• Methodology section

Store extracted content as plan_content.

Step 3: Synthesize recommendation via sonnet subagent

Spawn a sonnet subagent with the report's Conclusions, Recommendations, and experiment status. Produce a 1-3 sentence directional recommendation (inverted pyramid — most important finding first). Store as recommendation.

Step 4: Select lenses via sonnet subagent

Spawn a sonnet subagent with experiment_type and the plan's IV/DV structure. Select 1-2 exp-lens slugs using the Lens Selection Table above.

Output: a list of 1-2 slugs, e.g. ["fair-comparison", "estimand-clarity"]. Store as selected_lens_slugs.

Step 5: Write one context file per lens

For each slug in selected_lens_slugs, write a context file to:

{{AUTOSKILLIT_TEMP}}/prepare-research-pr/exp_lens_context_{slug}_{ts}.md

The context file must contain (enough for the lens to build its diagram without reading the entire CWD):

# Experiment Context: {slug} Lens

## Hypotheses

| | Hypothesis |
|---|---|
| H0 | {null hypothesis} |
| H1 | {alternative hypothesis} |

## Independent Variables

| Variable | Values / Levels | Role |
|----------|-----------------|------|
{rows from plan's IV table}

## Dependent Variables / Metrics

| Metric | Unit | Threshold |
|--------|------|-----------|
{rows from plan's DV table}

## Controlled Variables

{list from plan}

## Success Criteria

{from plan}

## Experiment Type

{experiment_type}

Record the absolute paths in lens_context_paths list (same order as slugs).

Step 6: Build results summary

From the report's Results section, construct:

• Findings table (Finding / Result / Confidence)
• Experiment status badge
• Key metrics (if standardized metrics referenced in the report, include inline)

Store as results_summary (plain text, will be embedded in prep file).

Step 7: Write PR prep file

Write to {{AUTOSKILLIT_TEMP}}/prepare-research-pr/pr_prep_{ts}.md:

# PR Prep: {title from report}

## Metadata

- report_path: {absolute path}
- experiment_plan_path: {absolute path}
- feature_branch: {branch}
- base_branch: {branch}
- experiment_type: {type}
- status_badge: {CONCLUSIVE_POSITIVE | CONCLUSIVE_NEGATIVE | INCONCLUSIVE | FAILED}

## Recommendation

{recommendation from Step 3}

## Results Summary

{results_summary from Step 6}

## Hypothesis Table

| | Hypothesis |
|---|---|
| H0 | {H0} |
| H1 | {H1} |

## Metrics Table

| Metric | Unit | Threshold |
|--------|------|-----------|
{DV rows}

## Methodology

{condensed from plan's Methodology section}

## Selected Lenses

{comma-separated slugs}

## Lens Context Paths

{comma-separated absolute paths}

Store the absolute path as prep_file_path.

---

Output

Emit these tokens as literal plain text (no markdown formatting on the token names) as your final output:

prep_path = /absolute/path/{{AUTOSKILLIT_TEMP}}/prepare-research-pr/pr_prep_{ts}.md
selected_lenses = fair-comparison,estimand-clarity
lens_context_paths = /abs/ctx_fair-comparison_{ts}.md,/abs/ctx_estimand-clarity_{ts}.md

Where:

• selected_lenses is a comma-separated list of lens slugs (no spaces)
• lens_context_paths is a comma-separated list of absolute context file paths in the same order as selected_lenses