Skip to content

ValentinNikolaev/agent-skills / agent-plugins/skills/review-retrospective

review-retrospective

Verified

Analyze completed code or prose review runs, verified feedback, upstream misses, timing, and limit usage. Propose project lessons, skill improvements, or model changes without silently applying them.

FileKindSize
SKILL.mdskill md6.7 KB
agents/openai.yamlconfig301 B
references/evaluation.mdmarkdown2.9 KB
references/model-policy.mdmarkdown4.1 KB
references/run-contract.mdmarkdown6.0 KB
scripts/review_learning.pyscript17.0 KB

references/evaluation.md

View on GitHub
# Validate learning and model changes

For completed humanizer/detector runs, also use
[the prose evaluation gates](../../ai-detector/references/prose-retrospective.md#validate-a-lesson).
Preserved meaning/voice and reader benefit take priority over service metrics;
keep literal external observations separate and compare the exact copies.

Separate package/JSON checks from behavioral effectiveness. A helper accepting
an assessment does not prove that a model finds the defect, and an accepted bot
comment is not a labeled ground truth. Resolve or retain disputed evidence.

For a skill change, use a frozen original case, a near-neighbor clean control,
and a small held-out set covering other mechanisms. Keep evaluation-only answers
and retrospective diagnoses away from the reviewing agent. Compare old/new skill
versions with the same available inputs, model, settings, depth, and tools.
An upstream stage gets only information available before implementation.

For a model change, hold the skill and task conditions fixed. Use comparable
scope, risk, evidence access, checks, and tooling. Separate model latency from
CI waits, tool delays, warm/cold context, and repeated reads. Repeat a noisy
comparison only when it can change the decision within the authorized budget.

Assess verified findings, false positives, severity/verdict correctness, known
misses, and follow-up rounds alongside actual tokens, elapsed time, and comparable
usage snapshots. Measure recall only on labeled cases; unseen production misses
are unknown. Zero findings does not prove quality. A cheaper result that skipped
required checks is reduced coverage, not a demonstrated efficiency gain.

Preserve the original report and evidence, comparison settings, exact outputs,
evaluation limits, and total cost of the experiment. Do not claim a universal
speedup or saving from one incident. A trial can justify a limited experiment
rather than a project default. Protect confirmed critical cases and clean controls
before broadening a policy; state any observed regression.

For an application gap, first check whether a smaller entrypoint, a relevant
reference trigger, a deterministic validation gate, or a clearer handoff fixes
execution. Adding another universal checklist item needs separate justification.
For a project lesson, include its scope, sources, invalidation conditions, and
superseding decision. Never promote a local convention into a general skill rule.

For a PR proposal give the owning skill, concrete failure, minimal canonical
patch, sanitized failing/clean pair, behavioral comparison, and expected context
and maintenance cost. Keep private paths, organizational policies, and real
customer data out of the public patch. Existing canonical generation, validation,
versioning, and independent-review rules remain authoritative. Record a rejected
or ineffective candidate without appending a rule merely to complete the cycle.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall