| name | research-harness-work |
| description | RES: Execute approved analysis tasks from analysis_plan.md. Writes scripts, runs them, saves logs and outputs, updates task status. Trigger: run analysis, execute task, run regression, produce table, produce figure. Do NOT load for: planning, review, release, setup, audit, cleaning. |
| description-en | RES: Execute approved analysis tasks from analysis_plan.md. Writes scripts, runs them, saves logs and outputs, updates task status. Trigger: run analysis, execute task, run regression, produce table, produce figure. Do NOT load for: planning, review, release, setup, audit, cleaning. |
| kind | workflow |
| purpose | Execute approved analysis tasks with reproducible scripts and mandatory log verification |
| trigger | run analysis, execute task, run regression, produce table, produce figure, /research-harness-work |
| shape | workflow |
| role | executor |
| pair | research-harness-review |
| owner | research-harness-core |
| since | 2026-05-29 |
| allowed-tools | ["Read","Write","Edit","Bash","Glob","Grep"] |
| argument-hint | [TASK-ID] [all] [--no-commit] |
| user-invocable | true |
| effort | high |
Research Harness Work
Execute approved analysis tasks from analysis_plan.md. For each task, write a reproducible script, run it, verify the log exists, and update the evidence ledger.
This skill runs after /research-harness-plan (human-approved) and before /research-harness-review.
Quick Reference
| Input | Action |
|---|
/research-harness-work 1.1 | Execute task 1.1 from analysis_plan.md |
/research-harness-work 2 | Execute all tasks in stage 2 sequentially |
/research-harness-work all | Execute all cc:todo tasks sequentially |
/research-harness-work --no-commit | Execute without committing; leave outputs for review |
Default behavior with no argument: ask the user which task or stage to run. Do not auto-select.
Pre-flight Checks
Before executing any task:
- Read
analysis_plan.md. If it does not exist, stop and tell the user to run /research-harness-plan first.
- Confirm the task ID exists in
analysis_plan.md and its status is cc:todo or cc:wip. Do not re-execute cc:done tasks without explicit user instruction.
- Confirm the analysis-ready dataset specified in
analysis_plan.md exists under data/processed/.
- If the task has
cc:blocked status: report the blocking reason and stop. Do not work around a blocked task.
Procedure
Step 1 — Read the task definition
From analysis_plan.md, read for the specified task:
- Script path
- Log path
- Output path
- DoD (definition of done)
- Assumptions
- Unresolved questions
- Identification strength label
If the task has unresolved questions: stop and list them. Ask the user to resolve them before the task runs.
Mark the task cc:wip in analysis_plan.md.
Step 2 — Write the analysis script
