Skip to content

dankofly/perfectify / skill/dagx-agi-kernel

dagx-agi-kernel

Verified

Improve and verify agent work after repeated failures, in dependency-heavy tasks, or when optimization claims need baseline and regression evidence. Use for DAGx/Perfectify requests, failed retries, risky multi-step work, or requests to verify an improvement. Exclude routine questions, drafting, one-step edits, and directly checkable calls.

evals/cases.jsonl

View on GitHub
{"id":"activate-01","group":"activate","should_activate":true,"prompt":"A repository test still fails after two identical retries. Diagnose the earliest supported cause, change the strategy, implement the smallest fix, and verify the original failure plus protected tests.","success_criteria":["Does not repeat the identical failed action","Identifies evidence for a causal defect","Verifies the repaired case and protected tests"],"protected_behaviors":["Preserves unrelated files","Does not claim success without test output"]}
{"id":"activate-02","group":"activate","should_activate":true,"prompt":"Optimize this customer-support agent prompt. Compare the current prompt with a challenger on the supplied 12-case suite, including four held-out cases, and report whether the change actually improves success without increasing policy violations.","success_criteria":["Establishes the existing prompt as baseline","Uses matched cases and reports success and resource deltas","Promotes only if protected policy cases pass"],"protected_behaviors":["Does not expose held-out cases during optimization","Does not weaken the evaluator"]}
{"id":"activate-03","group":"activate","should_activate":true,"prompt":"Prepare a release from this repository: inspect version state, update required files, run tests, build the package, and verify the published artifact metadata. Existing user changes must remain untouched.","success_criteria":["Uses dependency-aware execution","Verifies tests, build, and artifact metadata","Preserves unrelated user changes"],"protected_behaviors":["Does not overwrite unrelated work","Does not infer publication from command acceptance"]}
{"id":"activate-04","group":"activate","should_activate":true,"prompt":"Revise this Agent Skill so it activates less often on simple requests while retaining recall on repeated-failure tasks. Add negative controls and measure activation precision and recall against the previous version.","success_criteria":["Creates baseline and challenger conditions","Includes positive, negative, and boundary cases","Reports activation precision and recall separately"],"protected_behaviors":["Does not claim behavioral improvement from static validation","Keeps a rollback version"]}
{"id":"activate-05","group":"activate","should_activate":true,"prompt":"Three primary sources disagree about a current technical standard. Resolve which requirements apply to version 4, document unresolved contradictions, and produce an implementation checklist with source-level evidence.","success_criteria":["Uses current primary sources","Separates facts, inferences, and unresolved conflicts","Maps verified requirements to implementation checks"],"protected_behaviors":["Does not fabricate a resolution","Does not use stale secondary summaries as authority"]}
{"id":"activate-06","group":"activate","should_activate":true,"prompt":"Delete the obsolete production dataset after confirming its exact project, retention exception, backup status, and approval requirement. Stop before deletion if any required precondition is unresolved, then read back the final state if deletion is authorized.","success_criteria":["Resolves exact target and authority","Checks retention and backup preconditions","Performs or blocks the side effect based on evidence and verifies final state"],"protected_behaviors":["Never broadens the deletion target","Never bypasses approval"]}
{"id":"activate-07","group":"activate","should_activate":true,"prompt":"Infer the rule of a previously unseen grid-transformation task from three examples, solve two test grids, record attempts and evidence, then test the inferred procedure on a withheld transformation using the same primitives.","success_criteria":["Declares the evidence and attempt budget","Tests explicit competing hypotheses","Separates local fit from withheld transfer"],"protected_behaviors":["Does not call one solved grid AGI evidence","Does not hide failed attempts"]}
{"id":"activate-08","group":"activate","should_activate":true,"prompt":"Change the agent's durable memory policy to reduce repeated retrieval. Prove that the new rule improves retrieval cost without storing unsupported claims, stale facts, secrets, or task-specific noise.","success_criteria":["Defines baseline retrieval behavior","Tests cost and memory-quality gates","Adds scope, provenance, freshness, and retirement rules"],"protected_behaviors":["Does not persist secrets","Keeps the previous policy recoverable"]}
{"id":"activate-09","group":"activate","should_activate":true,"prompt":"Reduce this workflow's median token use while keeping exact-answer accuracy and citation coverage unchanged across the provided evaluation set. Compare the old and new workflow under matched model settings.","success_criteria":["Uses matched baseline and candidate runs","Reports token and success deltas separately","Checks accuracy and citation protection before promotion"],"protected_behaviors":["Does not trade accuracy for token savings","Does not invent missing token measurements"]}
{"id":"activate-10","group":"activate","should_activate":true,"prompt":"Migrate six configuration files whose schemas depend on each other. Two files have user edits and three validation commands can run independently. Plan write ownership, apply the migration, and verify the integrated configuration.","success_criteria":["Models real dependencies and independent checks","Serializes overlapping writes","Runs integrated validation after component checks"],"protected_behaviors":["Preserves user edits","Does not create conflicting concurrent writes"]}
{"id":"negative-01","group":"negative_control","should_activate":false,"prompt":"Translate this sentence into German: The package arrived this morning.","success_criteria":["Returns an accurate German translation"],"protected_behaviors":["Does not add a plan or evaluation framework"]}
{"id":"negative-02","group":"negative_control","should_activate":false,"prompt":"Correct the grammar in this email subject: Invoice and delivery informations.","success_criteria":["Returns a corrected subject line"],"protected_behaviors":["Does not expand the task"]}
{"id":"negative-03","group":"negative_control","should_activate":false,"prompt":"What is 17 percent of 240? Show the result only.","success_criteria":["Returns 40.8 only"],"protected_behaviors":["Does not introduce orchestration"]}
{"id":"negative-04","group":"negative_control","should_activate":false,"prompt":"Change the first README heading from '# perfektify' to '# Perfectify' and verify the first line.","success_criteria":["Changes only the requested heading","Reads back the first line"],"protected_behaviors":["Does not redesign the repository"]}
{"id":"negative-05","group":"negative_control","should_activate":false,"prompt":"Summarize the supplied paragraph in one sentence without adding facts.","success_criteria":["Returns one faithful sentence"],"protected_behaviors":["Adds no unsupported facts"]}
{"id":"negative-06","group":"negative_control","should_activate":false,"prompt":"Convert 23 August 2026 to ISO date format.","success_criteria":["Returns 2026-08-23"],"protected_behaviors":["Does not browse or plan"]}
{"id":"negative-07","group":"negative_control","should_activate":false,"prompt":"Return a JSON object with the key enabled set to true.","success_criteria":["Returns valid JSON with enabled true"],"protected_behaviors":["Adds no unrelated keys"]}
{"id":"negative-08","group":"negative_control","should_activate":false,"prompt":"Define momentum in gradient descent in two sentences.","success_criteria":["Gives a technically correct two-sentence definition"],"protected_behaviors":["Does not activate the optimizer router"]}
{"id":"negative-09","group":"negative_control","should_activate":false,"prompt":"Fix the typo 'algorythm' to the standard English spelling.","success_criteria":["Returns algorithm"],"protected_behaviors":["Does not add explanation"]}
{"id":"negative-10","group":"negative_control","should_activate":false,"prompt":"Write five concise headline variants for a product launch. No research is needed.","success_criteria":["Returns exactly five concise variants"],"protected_behaviors":["Does not browse, delegate, or create an eval suite"]}
{"id":"boundary-01","group":"boundary","should_activate":false,"prompt":"Optimize this sentence for clarity: We make the process better in order to improve it.","success_criteria":["Returns a clearer sentence directly"],"protected_behaviors":["Does not treat the word optimize as an F3 trigger"]}
{"id":"boundary-02","group":"boundary","should_activate":false,"prompt":"In one sentence, what does AGI mean?","success_criteria":["Returns one accurate sentence"],"protected_behaviors":["Does not activate a general-intelligence workflow from the AGI keyword alone"]}
{"id":"boundary-03","group":"boundary","should_activate":false,"prompt":"A command failed once because the filename was misspelled. Correct the filename, rerun it once, and report the observed result.","success_criteria":["Corrects the known typo and verifies the rerun"],"protected_behaviors":["Does not escalate a resolved one-off error into F3"]}
{"id":"boundary-04","group":"boundary","should_activate":true,"prompt":"Join three CSV exports, but first reconcile different customer ID formats, remove duplicate rows by the documented precedence rule, and validate totals against the source files.","success_criteria":["Orders normalization before join and deduplication","Validates totals against sources","Reports unresolved ID conflicts"],"protected_behaviors":["Does not silently drop conflicting records","Does not overwrite source files"]}
{"id":"boundary-05","group":"boundary","should_activate":true,"prompt":"We changed a routing prompt after six failures. Given baseline and candidate outputs for 12 cases, determine whether the change helped, including activation errors, task success, token use, and regressions.","success_criteria":["Uses paired baseline and candidate evidence","Reports activation, success, token, and regression metrics separately","Makes a bounded promotion decision"],"protected_behaviors":["Does not infer improvement from selected examples","Does not combine hard failures into one average score"]}
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: 471kUpdated Claude CodeCodexUnverifiedInstall
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