Improve Skill Quality

Turn a failing or unconvincing evaluation into a targeted fix. The single most common mistake in this repo is rewriting skill prose in response to a verdict whose real cause was the eval, the fixtures, or the harness. Classify first, then fix.

When to Use

• An evaluation verdict is a regression, underpowered, or "no credible improvement".
• A skill wins in the isolated arm but not in the plugin arm, or is reported "not activated".
• /evaluate reports "Evaluation ran but produced no results".
• A skill scores well but costs too much (tokens, turns, wall time, plugin menu budget).
• Deciding whether to strengthen or retire a persistently weak skill.

When Not to Use

• Creating a new skill from scratch — use create-skill.
• Creating a new eval.yaml from scratch — use create-skill-test.
• Changing the harness itself (eng/skill-validator, eng/vally-adapter, evaluation*.yml).

Inputs

| Input | Required | Description | |-------|----------|-------------| | Verdict evidence | Yes | The /evaluate PR comment, or results.json from the run artifacts | | Losing trial transcripts | Yes for content fixes | Baseline vs. skilled output plus the judge's stated reason | | Stimulus-vote W/T/L and repeated-run W/T/L | Yes | Separates cross-task evidence from reliability | | Activation status per arm | Yes | Isolated and plugin activation are different failures |

Workflow

Step 1: Get the evidence before forming a hypothesis

Read InvestigatingResults.md for how to download artifacts and read results.json. Extract, per failing stimulus:

• authoritative stimulus-vote W/T/L and separate repeated-run W/T/L
• activation status in the isolated and plugin arms, separately
• the judge's verbatim reason on each losing trial
• whether any trial errored, timed out, or produced empty output

Do not change skill content until you can quote a losing trial and the judge's reason for it. For the other cause classes the evidence is different: harness failures are diagnosed from the job log and the spec, and power problems from the trial record — neither has a losing trial to quote, and demanding one is what sends people rewriting prose instead.

Step 2: Classify the failure

Work down this table and stop at the first row that matches. Rows are ordered by how often the symptom has been misdiagnosed as a skill-content problem — the fixture row is first because a fixture failure also presents as a setup or reliability failure and gets misfiled as one.

| Symptom | Real cause class | Go to | |---------|------------------|-------| | A fixture does not build, is untracked by git, breaks for the wrong reason, or contradicts itself | Fixture | Step 4 | | No results.json, "produced no results", or the spec never loaded | Harness / spec-load | Step 3 | | Trials errored, timed out, or returned empty output | Reliability | Step 3 | | Trajectories unmatched, a trial errored, or the summary disagrees — verdict reported inconclusive | Reliability (not power) | Step 3 | | Positive record (e.g. 16W/8T/1L), comparison conclusive, verdict still not a pass | Statistical power | Step 5 | | Skilled arm equals baseline arm by construction | Eval design | Step 6 | | Activated and lost on quality, judge names a concrete defect | Skill content | Step 7 | | Activated in isolation, not in plugin | Activation / routing | Step 8 | | Not activated in either arm | Frontmatter description | Step 8 | | Wins but costs far more than baseline | Scope and cost | Step 7 |

A verdict is only a measured result when the comparison was conclusive: adapt.mjs requires zero errored trials, zero unmatched trajectories, and an agreeing summary before it will report a pass or a regression. Confirm that before reading a record as a power problem.

Step 3: Rule out harness and reliability causes

See references/eval-triage.md for the full catalogue. The recurring ones:

• A spec declaring both config: and defaults: is rejected by vally, the job still exits 0, and

the PR comment blames "transient infrastructure". Merge them into one defaults: block.

• An errored trial is not automatically a fixture problem — judge-side auth and session.idle

failures look identical from the verdict and need harness fixes, not SDK pins.

• expect_tools: [bash] on an advisory question forces a restore or build and turns an answer into

a timeout with no quality gain.

• Genuine code-generation stimuli need roughly 360s; a timeout yields empty output, which fails

every grader and hides the real quality signal.

• Unmatched trajectories, an errored trial, or a summary that disagrees make the comparison

inconclusive: the remaining matched trials are biased, so the record is not a measured null and must not be read as a power or content problem.

Step 4: Verify the fixtures before touching the skill

Run python eng/eval-quality/check_eval_quality.py — it blocks eleven defect classes that can cost a real result here. Then confirm by hand:

• every fixture behaves as its stimulus assumes — a fixture meant to be healthy builds, and one

meant to be broken fails for the exact reason the stimulus is about and no other;

• every referenced fixture is in the git index (git ls-files), not merely on disk — .gitignore

has silently swallowed committed coverage fixtures;

• a fixture never states the same fact in two places that disagree — a Cobertura report whose

declared line-rate, summary totals and <line> elements differ is the canonical case — or the two arms legitimately read different truths.

Step 5: Check whether the eval could ever have passed

The gate has two independent bars, and confusing them is the usual misdiagnosis:

1. Distinct stimuli ≥ 5. Below that the verdict is reported

underpowered — never a pass, never a regression.

1. The sign test must reach p ≤ 0.05 over the discordant (non-tie) stimulus votes. Ties are not

discarded silently; they hold the discordant count down.

| discordant stimulus votes | records that pass | p | |---:|---|---:| | ≤ 4 | none, however good the skill | ≥ 0.0625 | | 5–7 | zero losses only (5W/0L) | 0.031 | | 8 | one loss survivable (7W/1L) | 0.035 |

So at exactly 5 stimuli a single tie is fatal — it leaves 4 discordant. At 6 stimuli one tie is survivable (5W/1T/0L); at 7, up to two are (5W/2T/0L). A loss is not.

So a positive record with a failing verdict is a power problem, not a content problem. Fix it by adding discriminating stimuli. Raising runs measures reliability for the same task and cannot clear the floor.

Step 6: Check whether the two arms differ at all

An eval that compares the skill against itself measures judge noise:

• A dormancy guard (expect_activation: false) must not also set constraints.reject_skills.

That makes the skilled arm skill-free, so the activation contract cannot observe a hijack. Schema version 4 retains the identical-arm comparison for diagnostics but excludes it from preference inference; unexpected isolated activation still blocks a pass.

• A skill with disable-model-invocation: true is absent from the model-facing skilled arm, so its

direct eval compares two identical arms regardless of whether graders inspect activation or answer content. Cover it through consumer outcomes instead; for example, filter-syntax is covered by run-tests and mtp-hot-reload.

• A grader whose config is missing its required key enforces nothing, so the stimulus has one

fewer assertion than it appears to.

Step 7: Fix skill content against the losing trial

Only now change the skill. Apply the patterns in references/writing-for-baseline-delta.md; the ones that most often flip a loss:

• Replace reference prose the model already knows with decisions it would otherwise get wrong.
• Add stop-conditions so a strong skill does not over-apply — but do not over-correct into

answering more narrowly than the baseline did.

• Scale output structure to input size; a dashboard for an 8-test suite loses to a direct answer.
• Require truthful validation reporting; claiming "Build succeeded" after a failed restore is an

automatic loss.

• Verify load-bearing API claims by compiling or probing, not by reading source.
• For cost regressions, gate rare or expensive paths behind references/ reads and size any

orchestration to the user's scope.

Step 8: Fix activation

Activation failures are frontmatter and routing failures, not body failures. See references/eval-triage.md. Summary:

| Failure | Fix | |---------|-----| | Not activated in any arm | Put the user's own words in description: symptoms, error codes, artifact names, quoted requests | | A sibling skill wins the prompt | Claim the exact ambiguous words in description, and add matching exclusions on both siblings | | Model answers with no skill at all | Raise the stakes in the description, de-crowd the plugin menu, verify with the plugin arm | | Boundary excludes real scenarios | Re-read every "do not use for" clause against every eval prompt and real workflow phase | | Description at the 1,024-char ceiling | Cut restated body content, not trigger phrases; check the plugin menu budget too |

Step 9: Re-validate

bash
dotnet run --project eng/skill-validator/src/SkillValidator.csproj -- check --plugin ./plugins/<plugin>
python eng/eval-quality/check_eval_quality.py
./eng/run-skill-evals.sh <plugin> <skill>

Then request the official run by submitting a PR review containing /evaluate (Files changed → Review changes), which binds the run to the reviewed commit. Before declaring a regression on the result, confirm the skill payload actually changed — reruns on byte-identical content have shifted 7W/2T/2L to 4W/5T/2L.

Validation

• ] For a content fix, a losing trial and the judge's stated reason are quoted in the PR description.
• ] The failure was classified before any content was edited.
• ] check_eval_quality.py and skill-validator check both pass.
• ] Distinct-stimulus count clears the power bar for the target effect and observed tie rate.
• ] Isolated and plugin activation are both reported.
• ] The PR body records root cause, fix, and validation so the lesson is reusable.

Common Pitfalls

| Pitfall | Solution | |---------|----------| | Rewriting skill prose in response to an underpowered verdict | Underpowered means too few distinct stimuli; add discriminating stimuli instead | | Adding defaults: runs: to a spec that already has config: | Merge into a single defaults: block; vally rejects specs with both | | Padding runs to clear the stimulus floor | Repeats measure reliability for one task; add stimuli | | Treating an errored trial as fixture nondeterminism | Read the stderr first; judge-side auth failures need harness fixes | | Fixing a "wrong" answer that the fixture actually made wrong | Check fixture self-consistency before blaming the response | | Strengthening a skill nobody uses and nothing passes | Weak eval signal plus thin telemetry is a valid retirement case | | Landing a fix without re-running | Verify the invoked payload contains the fix; judge noise is real |

References

• references/writing-for-baseline-delta.md — content patterns that beat the unskilled model
• references/eval-triage.md — symptom, cause and fix catalogue with PR citations
• eng/eval-quality/README.md — the eleven structural gate checks and why each exists
• eng/vally-adapter/InvestigatingResults.md — downloading artifacts and reading results.json. This is the current guide; the similarly-named eng/skill-validator/src/docs/InvestigatingResults.md documents the retired skill-validator evaluate schema and does not describe today's results.