| name | llm-as-verifier |
| description | Use the local llm-as-a-verifier tool to collect and rank evidence-aware coding-agent trajectories. Trigger when a user asks to use llm-as-a verifier, verify a coding-agent run, compare candidate runs, capture objective command/diff/final-answer evidence, or gate completion on a verifier result using the tool at C:\dev\Desktop-Projects\llm-as-a-verifier. |
LLM as Verifier
When this skill is triggered, use the actual llm-as-a-verifier CLI from C:\dev\Desktop-Projects\llm-as-a-verifier. Do not simulate, paraphrase, or manually recreate the verifier. The job is to collect evidence with verifier-collect, run verify-trajectories, inspect the JSON result, and let that result influence whether the task is complete.
Use llm-as-a-verifier as an evidence logger and deterministic verifier for coding-agent work. It is not a replacement for tests or review; it checks whether final success claims are supported by captured command output, diffs, and other trajectory evidence.
Fast Path
First, check whether the CLI is already available:
verifier-collect --help
verify-trajectories --help
Install once if those commands are missing:
python -m pip install --user -e C:\dev\Desktop-Projects\llm-as-a-verifier
Then confirm the CLI:
verifier-collect --help
verify-trajectories --help
evaluate-verifier --help
Initialize from the repo being worked on:
verifier-collect init `
--task-id <short-task-id> `
--task-description "<one-sentence task>" `
--candidate-id <candidate-id> `
--profile node `
--out .verifier\trajectory.json
Choose the closest profile: default, python, node, webapp, security-sensitive, or browser-agent.
Collect Evidence
Use the real command wrapper for meaningful verification commands:
verifier-collect run --file .verifier\trajectory.json -- npm test
verifier-collect run --file .verifier\trajectory.json -- npm run build
Capture the patch:
verifier-collect diff --file .verifier\trajectory.json
Record the final claim exactly enough for the verifier to check:
verifier-collect final --file .verifier\trajectory.json "Implemented the requested change. npm test and npm run build pass."
Run the verifier:
verify-trajectories .verifier\trajectory.json --out .verifier\verification_result.json
Read .verifier\verification_result.json before finalizing. Use the ranked candidate, score, confidence, positive and negative evidence, unsupported claims, missing evidence, risk factors, uncertainty reasons, and semantic certificate to decide whether more work is needed.
If the CLI command fails because it is not installed, install it from the local path above and retry. If it still fails, report the failure and do not claim that LLM-as-a-verifier was run.
Multiple Candidates
For competing runs, collect each candidate under the same task with a distinct candidate_id. Then run:
verify-trajectories .verifier\trajectory.json --out .verifier\verification_result.json
Prefer the top-ranked candidate only when its evidence supports the task. If scores are close or confidence is low, inspect pairwise comparisons and missing evidence before choosing.
Result Policy
Treat verification_result.json as a gate, not a rubber stamp:
- If the verifier reports contradicted final claims, nonzero command exits, missing tests, suspiciously small diffs, or unsupported success claims, fix or gather better evidence before finalizing.
- If the verifier passes but important manual risks remain, mention those risks in the final answer.
- If the verifier cannot run, report that clearly and fall back to normal verification commands.
- Do not claim success solely because the final assistant message says the work is done.
Repo Config
verifier-collect init creates .verifier\verifier.toml. Commit it only when stable repo-specific verifier policy is useful.
When editing TOML on Windows, quote filesystem paths with single quotes:
tool_path = 'C:\dev\Desktop-Projects\llm-as-a-verifier'
Use double quotes only when backslashes are escaped:
tool_path = "C:\\dev\\Desktop-Projects\\llm-as-a-verifier"
Final Answer
When this skill is used, include a short verifier status:
Verifier: ran verify-trajectories; top candidate <id>, score <score>, confidence <confidence>.
If verification was blocked, state the blocker and the strongest evidence that was still checked.