agentsclimarketplace
Sign in

Verify subagent output

Skill jcdavis131/cursor-agent-skills/skills/verify-subagent-output

After parallel subagents (SME reviewers, research agents, task agents) return, spot-check their output against the named contract/standard before synthesizing or marking the task done. Use whenever you dispatched subagents and they've returned artifacts — especially before treating their output as ground truth in a synthesis or a "done" report.From its SKILL.md

Install
npx -y skills add jcdavis131/cursor-agent-skills --skill verify-subagent-output

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

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.7 KB, 572 tokens by cl100k_base, as published. Nobody here has run it

Verify Subagent Output

Subagent output is the most trusted-yet-unverified thing in an agent loop. You dispatched it; it returned prose; you're tempted to trust it. Don't — spot-check first.

Why this matters

  • SMEs drift from the contract — they skip a section, add an unrequested one, or pad.
  • Subagents can claim evidence they didn't actually gather ("verified against the spec" without opening the spec).
  • A synthesis built on unverified subagent output inherits every gap.
  • A "done" report built on it misleads the user.

The spot-check

You don't re-do the subagent's work. You check conformance + substance:

  1. Conformance to the contract. Each returned file has the agreed sections, in the agreed shape, at the agreed length. Missing section = drift.
  2. Conformance to the named standard. If output is supposed to be "OKF-conformant" or "ADRs" or "spec-0008-shaped", check the shape against that standard by name.
  3. Substance, not just shape. At least one non-trivial claim per section that a careful reader would accept as evidence the SME actually did the work — not filler. "Security was reviewed" is filler; "authn boundary at app/auth.py:42 is missing a workspace check" is substance.
  4. A sample, not all. For 5 SMEs, spot-check 1–2 in depth and skim the rest. Full re-verification defeats the fan-out.

The report

State the result as a one-line verdict that names the standard:

Bundle is fully OKF-conformant with five substantive reviews.

Or, if drift found:

OKF bundle: 4/5 conformant; security review missing the Risk section — sent back for completion.

When to send back vs. fix yourself

  • Send back (re-dispatch the subagent with the specific gap) when the gap is large or in the SME's domain.
  • Fix yourself when the gap is a missing boilerplate section you can write in one line.

When NOT to verify

  • The subagent's output is purely advisory and won't feed a synthesis or a "done" claim.
  • The cost of the spot-check exceeds the cost of being wrong (rare — usually the spot-check is cheap).

Pair with

  • sme-fanout — verification is the step between "all SMEs returned" and "synthesize".
  • validate-gate — same ethos: cheap gate before declaring done.
  • follow-procedure — the named standard you check against is the procedure.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 326,834. 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.