agentsclimarketplace

Research harness work

Skill maxwell2732/claudecode-research-harness-workflow/skills/research-harness-work

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.From its SKILL.md

Install
npx -y skills add maxwell2732/claudecode-research-harness-workflow --skill research-harness-work

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

8.1 KB, as published. Nobody here has run it

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

InputAction
/research-harness-work 1.1Execute task 1.1 from analysis_plan.md
/research-harness-work 2Execute all tasks in stage 2 sequentially
/research-harness-work allExecute all cc:todo tasks sequentially
/research-harness-work --no-commitExecute 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:

  1. Read analysis_plan.md. If it does not exist, stop and tell the user to run /research-harness-plan first.
  2. 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.
  3. Confirm the analysis-ready dataset specified in analysis_plan.md exists under data/processed/.
  4. 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

Write the script at the path specified in the task definition.

Script requirements:

  • Language: use whichever language the study uses (R, Stata, Python). If not specified, use R.
  • Header comment: study name, task ID, date, analyst (Claude Code), identification strategy from study_spec.md §2
  • Project-relative paths only
  • Set random seed if any random operation is used
  • Open log file at the start; close at the end
  • Log: task ID, date, dataset loaded, N rows loaded, any sample restrictions applied with N dropped and N remaining, all model results including coefficients and standard errors, N in estimation sample

R template for regression task:

# Study: <study name>
# Task: <task-id> — <task description>
# Identification: <tag from study_spec.md>
# Date: YYYY-MM-DD
# Analyst: Claude Code

library(here)

log_file <- here("analysis", "<stage>", "logs", "<task>.log")
sink(log_file, append = FALSE, split = TRUE)

cat("Task:", "<task-id>", "\n")
cat("Started:", format(Sys.time()), "\n\n")

# Load
df <- read.csv(here("data", "processed", "analysis_ready.csv"))
cat("Loaded:", nrow(df), "rows\n")

# Sample restriction (if task-specific)
n_before <- nrow(df)
df <- df[<restriction>, ]
cat("After restriction:", nrow(df), "rows (dropped:", n_before - nrow(df), ")\n")

# Model
fit <- lm(<formula>, data = df)
cat("\n=== Results ===\n")
print(summary(fit))

# Save output
write.csv(broom::tidy(fit), here("output", "tables", "<task>.csv"), row.names = FALSE)
cat("\nOutput written to: output/tables/<task>.csv\n")
cat("Finished:", format(Sys.time()), "\n")
sink()

Do not hard-code numerical results into the script. The script computes them; the log records them.

Step 3 — Run the script

Run the script. Capture the exit code and the full log output.

  • Exit code 0: proceed to Step 4
  • Exit code non-zero (attempt 1): read the log, identify the error, fix the script, re-run
  • Exit code non-zero (attempt 2): read the log again, fix the specific error, re-run
  • Exit code non-zero (attempt 3): stop. Do not invent output. Write the error to analysis_plan.md under the task. Mark the task cc:blocked. Tell the user what failed and why.

Do not continue to the next step if the script did not exit with code 0.

Step 4 — Verify evidence

After a successful run, verify:

  1. Log file exists at the path specified in the task definition
  2. Output file (table or figure) exists at the path specified
  3. The log contains the numerical results that will be reported (coefficients, standard errors, sample size)

If any verification fails: do not mark the task cc:done. Report what is missing.

Step 5 — Update analysis_plan.md

Add an evidence row to the Evidence Ledger in analysis_plan.md:

TaskScriptLogOutputVerifiedDate
1.1analysis/descriptive/scripts/table1.Ranalysis/descriptive/logs/table1.logoutput/tables/table1.csvYESYYYY-MM-DD

Mark the task status cc:done.

Step 6 — Report completion

Print a task completion summary:

Task 1.1 — Complete

Script:  analysis/descriptive/scripts/table1.R
Log:     analysis/descriptive/logs/table1.log   [exists: YES]
Output:  output/tables/table1.csv               [exists: YES]
Status:  cc:done

Key results from log:
  N = <value from log>
  [other key stats from log]

Next task: 1.2 or run /research-harness-work 1.2

Do not paraphrase or interpret the results. Quote them from the log.

Forbidden Actions

  • Do not report any number that is not present in the log file
  • Do not mark a task cc:done unless the log file and output file both exist
  • Do not continue after a script failure without documenting the error
  • Do not change analysis_plan.md task scope without recording the change under the task
  • Do not use absolute file paths in scripts
  • Do not execute a task that is cc:blocked without user instruction
  • Do not run tasks that were not in the approved analysis_plan.md without user instruction

Evidence Requirements

  • Script file exists at the documented path
  • Log file exists at the documented path
  • Output file exists at the documented path
  • Evidence row added to analysis_plan.md
  • Task status updated to cc:done in analysis_plan.md

Completion Criteria (per task)

  • Task definition read from approved analysis_plan.md
  • No unresolved questions blocking the task
  • Script written with header comment, project-relative paths, log file
  • Script ran with exit code 0
  • Log file exists
  • Output file exists
  • Key results visible in log
  • Evidence row added to analysis_plan.md
  • Task status: cc:done

Escalation (task failure)

If a task fails after 3 attempts:

  1. Write to analysis_plan.md under the task:
    • What error occurred (quoted from log)
    • What was tried
    • What is needed to resolve it
  2. Mark task cc:blocked
  3. Tell the user: the data or design may not support this task. They must decide whether to revise the plan.

Do not invent output, use estimated values, or continue as if the task completed.

Handoff to Stage 6

After all approved tasks are complete:

All requested tasks are done. Review analysis_plan.md to confirm all expected tasks are cc:done.

When ready, run /research-harness-review to check identification, numerical accuracy, and causal claims before release.

Keep looking

Skills are one crate of 614,496. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.