prompt-test
Skill1 starsChanged 4 months agoMITView on GitHub
Smoke-test a prompt change by running the reviewer with the OLD and NEW prompt against the same target PR(s) and capturing a before/after comparison. Required evidence for any non-trivial prompts/default.md change.
- Reads credentials
2 other projects use a version of this.
Adopted by
Most recently taken up by
- teachskillofskills-ai/DigitalMarketingPro-techshutheir own version, first seen Oct 11, 2026
- indranilbanerjee/digital-marketing-protheir own version, first seen Feb 12, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- indranilbanerjee/digital-marketing-pro★ 934same name
- teachskillofskills-ai/DigitalMarketingPro-techshu★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository13 other files it gives its agents.
- AGENTS.mdAGENTS.md
- add-providerSkill
- ai-diff-reviewerSkill
- dailybotSkill
- deepworkplanSkill
- herdr-peersSkill
- releaseSkill
- ai-diff-reviewer-address-reviewSkill
- ai-diff-reviewer-apply-reviewSkill
- ai-diff-reviewer-generate-extensionSkill
- ai-diff-reviewer-open-prSkill
- ai-diff-reviewer-setupSkill
- ai-diff-reviewerSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 3 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 23
Does NOT modify the prompt for the user.
- Very specific: line 24
Does NOT post comments on the test PR — tests/eval/run_eval.py runs the review loop in-process with the GitHub submission path stubbed out, …
- Distinctive: line 42
If the prompt isn't actually changed in the working tree, ask the user whether they intended to test a different file.
- Very specific: line 46
Run everything through the in-tree harness tests/eval/run_eval.py (stdlib, offline — it fetches the PR context and runs the review loop …
- Distinctive: line 98
After running each PR, aggregate:
- Very specific: line 126
Anthropic rate limits: Add time.sleep(5) between PRs if you're testing >5 in a row.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Skill: Prompt Test
Objective
Provide before/after evidence that a change to prompts/default.md (or a custom prompt file) actually improves review quality. The skill runs the reviewer twice against the same PR — once with the OLD prompt, once with the NEW — and produces a comparison the user can paste into a PR description.
Without this evidence, prompt changes are opinion, not engineering.
Non-goals
- Does NOT modify the prompt for the user. The user owns the prompt edit; this skill evaluates it.
- Does NOT post comments on the test PR —
tests/eval/run_eval.pyruns the review loop in-process with the GitHub submission path stubbed out, so the target PR's conversation is never touched.
Inputs
target_prs— list of PR numbers to test against. For substantive changes use 3–5 covering different change types (feature, bugfix, refactor, docs). For targeted changes one PR is enough.repo—<owner>/<name>of the target repo. Defaults to the repo the skill is invoked from.
Pre-flight
# API keys present?
test -n "$ANTHROPIC_API_KEY" || { echo "Set ANTHROPIC_API_KEY"; exit 1; }
test -n "$GITHUB_TOKEN" || { echo "Set GITHUB_TOKEN"; exit 1; }
# On a branch with the prompt change?
git diff main...HEAD -- prompts/default.md | head -1If the prompt isn't actually changed in the working tree, ask the user whether they intended to test a different file.
Steps
Run everything through the in-tree harness tests/eval/run_eval.py (stdlib,
offline — it fetches the PR context and runs the review loop without posting
anything; never run scripts/reviewer.py against a live PR to smoke-test a
prompt). Pick target PRs from tests/eval/corpus.json whenever possible so
the comparison is scored against labelled expectations, not eyeballed.
1. Check out each target PR at its head
HEAD_SHA=$(gh api "repos/${REPO}/pulls/${N}" --jq .head.sha)
git worktree add "/tmp/aiprr-pr-${N}" "$HEAD_SHA"2. Run with the OLD prompt
git show main:prompts/default.md > /tmp/prompt-old.md # or the previous tag
python3 tests/eval/run_eval.py run --repo "$REPO" --pr "$N" --worktree "/tmp/aiprr-pr-${N}" \
--provider openai --api-base "$AZURE_OPENAI_BASE_URL" --model "$AZURE_OPENAI_MODEL_DAILY" \
--api-key-env AZURE_OPENAI_API_KEY --prompt /tmp/prompt-old.md --out "results/old-${N}.json"Any runner works (--provider anthropic --api-key-env ANTHROPIC_API_KEY for
the default backend; --provider grok --model grok-4.5 for the Grok CLI when
it is installed). Use the same runner and model for OLD and NEW.
3. Run with the NEW prompt
python3 tests/eval/run_eval.py run ... --prompt prompts/default.md --out "results/new-${N}.json"4. Compare
python3 tests/eval/run_eval.py score results/old-*.json results/new-*.jsonThe table gives, per run: must-find recall, false positives (labelled
must_not_flag), unlabelled findings (inspect them — they are either new true
positives to add to the corpus or noise), severity match, summary present,
suggestion blocks, turns, tokens, cost, wall clock. Add a qualitative bullet
list from the JSON bodies:
- "NEW flags X that OLD missed." (better)
- "NEW misses Y that OLD caught." (regression)
- "Severity of Z shifted from
info→warning." (calibration change)
Ship only when recall is ≥ and false positives ≤ on every model you could run; record blocked models honestly (no key, quota) instead of assuming.
Aggregation across multiple PRs
After running each PR, aggregate:
## Prompt smoke-test summary
Tested against N PRs covering <types>.
### Net findings
- **New catches old missed:** <list with PR links>
- **New misses old caught:** <list with PR links>
- **Calibration shifts:** <summary>
### Verdict
<one sentence: ship / refine / abandon>The summary belongs in the PR description for the prompt change.
Output
Print the summary to stdout. Optionally write it to tmp/prompt-test-<branch>.md if the user wants to attach a file.
Common failure modes
- Comparing across different runners or models. OLD and NEW must share runner + model; the runner's own tool loop changes the result more than most prompt edits.
- Trusting finding counts. A prompt that doubles false positives "finds more"; the corpus score is the metric.
- Anthropic rate limits. Add
time.sleep(5)between PRs if you're testing >5 in a row. - Target PR closed. PRs in closed/merged state still work for read-only review; skip if the diff is gone (rare).
- Provider cost. Each run is one full review at the configured model. For 5 PRs × 2 runs = 10 reviews. Estimate cost up front; if that's a problem use a cheaper model via
AIPRR_MODELfor the smoke test.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

