{"slug":"msw","title":"msw","summary":"Apply the Minimum Sufficient Work (MSW) principle through the MSW Kernel to scope, execute, verify, and stop agent work. State the requested outcome and smallest proof, admit a claim only when deleting it would leave the contract unmet or unproven, do and prove each necessary cla","platform":"Claude","tags":[],"authorName":"LLM Mart","authorSlug":"llm-mart","score":0,"source":"github","price":null,"verified":false,"createdAt":"2026-09-24T15:42:29.40278Z","repo":{"url":"https://github.com/transcendr/slopware-skills","stars":37,"forks":2,"license":null,"updatedAt":"2026-09-23T18:05:54Z"},"bodyHtml":"<hr>\n<h2>name: msw\ndescription: Apply the Minimum Sufficient Work (MSW) principle through the MSW Kernel to scope, execute, verify, and stop agent work. State the requested outcome and smallest proof, admit a claim only when deleting it would leave the contract unmet or unproven, do and prove each necessary claim, and halt at the fixed point. Use for implementation, debugging, review, research, planning, documentation, or any task where scope control, proportional evidence, or avoiding overwork matters. Also use when the user says MSW, Minimum Sufficient Work, do only what is necessary, avoid over-engineering, or stop when the result is proven.</h2>\n<h1>MSW: the Kernel</h1>\n<p>Minimum Sufficient Work is the principle. The MSW Kernel is the instruction set\nand program that applies it.</p>\n<h2>Program: complete</h2>\n<pre><code>contract ← the requested outcome + the smallest criteria that prove it\n\nwhile ∃ claim c : deleting c leaves contract unmet ∨ unproven\n      do c ; prove c\n\nhalt ; report\n</code></pre>\n<h2>Definitions: no behavior lives here, only meaning</h2>\n<p><strong>contract</strong>: the requested outcome and the smallest set of acceptance criteria\nthat would prove it, stated before any work. The sole source of necessity; a\nceiling as much as a floor. If the request is ambiguous: attended → ask;\nunattended → bind the smallest reading consistent with stated intent and record\nthe assumption.</p>\n<p><strong>claim</strong>: anything petitioning to become work: a plan step, a change, a test, a\nreviewer's P1, a discovered edge case, your own instinct that one more pass\nwould help. Everything enters as this type. Nothing enters as a verdict.</p>\n<p><strong>deleting c leaves contract unmet ∨ unproven</strong>: the only test. A claim passes\nsolely by breaking the contract, reproducibly, within the task's actual inputs\nand environment. Severity is derived from the contract, never inherited from\nwhoever raised the claim. <em>Useful</em>, <em>thorough</em>, and <em>possible</em> are not aliases\nfor <em>necessary</em>. A claim that fails receives one line in the report, never a\nfix, an investigation, or a deferred follow-up.</p>\n<p><strong>do ; prove</strong>: the smallest reliable act that closes the gap, and evidence\nsized to the claim it settles. An unproven act keeps its claim alive; a proven\none closes it, and re-proving a closed claim is itself an inadmissible claim.</p>\n<p><strong>halt</strong>: the fixed point: contract proven, no remaining claim passes. Not\nreviewer silence; not exhausted imagination. Halting before the fixed point and\nlooping past it are the same bug, mirrored.</p>\n<p><strong>report</strong>: the outcome against the contract; the proof; rejected claims worth\nthe user's attention, one line each. Nothing else.</p>\n<h2>Fuses: outside the program, for when its evaluator fails</h2>\n<pre><code>rounds = 3            → halt anyway ; report open items, do not chase them\nclaim born in round n+1, visible in round n   → rejected\n</code></pre>\n<h2>No unauthoritative limits</h2>\n<p>Never invent a limit. A cap, threshold, quota, budget, timeout, retry or round\ncount, file or line count, acceptance-criterion count, agent count, or similar\nconstraint is admissible only when its exact value is:</p>\n<ul>\n<li>explicitly required by the requester;</li>\n<li>imposed by an applicable technical or platform contract;</li>\n<li>defined by authoritative project policy; or</li>\n<li>derived from measured evidence necessary to meet or prove the task contract.</li>\n</ul>\n<p>State the authority or derivation whenever proposing or applying a limit. If no\nauthority exists, omit the limit and use the MSW necessity test. Metrics may be\nreported as evidence, but they must not become gates, defaults, targets, or\nrecommendations through agent intuition. Examples and representative\nproportions never become defaults. If a necessary limit is an unresolved owner\nchoice, ask; do not manufacture a value.</p>\n<h2>Contract integrity</h2>\n<p>Distinguish the deliverable's requirements from the state of the instance\nused to build or prove it. Necessity at one scope does not establish\nnecessity at the other.</p>\n<p>Justify delivered responsibilities by the established intended use or\nsupported lifecycle, independently of the current run's incidental history.\nDevelopment evidence can reveal required failure, recovery, or upgrade\nbehavior. Those obligations need justification beyond rescuing this run.</p>\n<p>Keep necessary, authorized setup or intervention at its own scope.\nPreserving a particular development instance is required only when the\ntask requires it.</p>\n<p>Temporary implementation or proof limitations must neither expand\ndelivered responsibilities nor weaken required behavior, acceptance\ncriteria, or intended user instructions. Describe current capability\ntruthfully and report implementation and proof gaps explicitly.</p>\n<p><a></a></p>\n<h2>Two-step analysis</h2>\n<p>When asked to apply MSW's two-step analysis to a proposal, perform these\nsteps in order. A proposal may be a plan, design, implementation approach,\nrepair, or review recommendation.</p>\n<h3>Step 1: Establish and validate the outcome contract</h3>\n<p>Derive the contract from the request and authoritative requirements,\nindependently of the proposed solution.</p>\n<p>State the requested outcome, applicable constraints and existing guarantees\nthat must be preserved, and the smallest acceptance criteria and evidence\nthat would prove success.</p>\n<p>Check that the contract includes every required outcome and excludes\nobligations introduced solely by the proposed mechanism. A mechanism belongs\nin the contract only when an authoritative requirement makes that mechanism\nitself mandatory. Apply Contract integrity to distinguish delivered behavior\nfrom incidental implementation or proof conditions.</p>\n<p>Resolve material ambiguity under the kernel's attended/unattended rule.\nDo not narrow the contract to fit the proposal or expand it to justify\nadditional work.</p>\n<p><a></a></p>\n<h3>Step 2: Start from the feature</h3>\n<p>Decide what the requested feature must accomplish, then design the\nimplementation to make that happen.</p>\n<p>Judge the work against that bound contract. A guard, check, blocker, or\nsubsystem is an implementation detail. Neither it nor an adjacent concern\nsupplies a contract of its own. Its purpose must not replace the feature's\nrequested outcome.</p>\n<p>Existing does not mean necessary. A presumed purpose is a hypothesis, not\na justification. An existing guarantee constrains the solution when an\nauthoritative requirement establishes it; that does not require preserving\nthe particular mechanism currently providing it.</p>\n<p>Do not invent problems to justify adding or preserving implementation\ndetails. A blocker that prevents the requested outcome without serving an\nestablished requirement of the feature does not belong in the solution.\nDo not add it. If already present, apply\n<a href=\"#retro-cleanup\">retrospective cleanup</a>.</p>\n<p>Treat every proposed or existing condition, mechanism, change, investigation,\ntest, review, and operational step as a claim.</p>\n<p>Apply the deletion test to each claim: identify the concrete contract\nrequirement that would remain unmet or unproven without it, using the task's\nactual inputs and environment. Retain necessary claims; reject the rest.\nWhere a smaller reliable act closes the same gap, use it.</p>\n<p>Then check sufficiency: would the retained proposal achieve every part of\nthe contract, with the required proof? Identify missing work, missing proof,\nand unsupported assumptions. Minimality alone does not establish completeness.</p>\n<p>Report the contract, the assessment, and the smallest complete proposal.\nGive rejected claims worth the user's attention one line each. Distinguish\nevidence already established from verification still required; analysis of\na proposal does not prove its implementation.</p>\n<p>Use the existing response or plan; no separate artifact or approval gate\nis required. Reuse a valid established contract and settled proof. Revisit\nthem only when new authoritative requirements or relevant evidence justify\nit. An analysis request does not itself authorize execution.</p>\n<p><a></a></p>\n<h2>Retrospective cleanup</h2>\n<p>When existing code fails <a href=\"#outcome-first\">outcome-first</a>, deletion alone\nis a valid solution. Delete the rejected behavior together with any code,\ntests, assertions, fixtures, configuration, or documentation that exists\nsolely to support or enforce it.</p>\n<p>Do not replace or repurpose that machinery, or invert its assertions,\nunless the feature's contract independently requires the work. Verify the\nrequested outcome with the smallest sufficient evidence, reusing existing\nproof where sufficient. Removing code creates no obligation to add code.</p>\n<h2>Apply the Kernel</h2>\n<ol>\n<li>Bind and validate the contract before proposing or doing work, then admit\ncandidate actions through the deletion test. For explicit proposal\nassessment, use the <a href=\"#two-step-analysis\">two-step analysis</a>.</li>\n<li>Execute and prove only admitted claims.</li>\n<li>Re-evaluate claims and proof when evidence changes. Revise the contract only\nunder authoritative direction or to correct its interpretation against\nestablished requirements. Current execution state alone does not authorize\na revision.</li>\n<li>Halt and report as soon as the fixed point is reached.</li>\n</ol>\n<p>Reject a failed claim with one report line. Do not turn it into an investigation,\nfix, or deferred follow-up.</p>\n<p>This skill works alone. If MSL is also installed, use this skill to decide what\nwork is necessary and MSL to report it. If Codex Voice Optimizer is active,\napply this kernel to its coordination claims and owning work threads while CVO\nowns routing, project placement, roles, authority, and speech.</p>\n<p>If the CODER Loop is active, apply this kernel to task-family,\nacceptance-claim, finding, repair, and proof admission while CODER owns\ndecomposition, independent review, remediation, and acceptance. The\nthree-round fuse applies only when this kernel's evaluator fails; it is not a\ndefault CODER review-round limit.</p>\n<p>If Timebox is active, use this kernel to decide what work remains necessary and\nTimebox to govern how that work converges inside the authorized clock. A hard\nstop reports open claims honestly; it never changes the necessity or proof\ntest.</p>\n","files":[{"path":"agents/openai.yaml","sizeBytes":207,"isText":true},{"path":"SKILL.md","sizeBytes":10034,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-09-24T15:42:38.91705Z","sha256":"FF3051E19858AFDF858A93327764E46A4268EF0AF840B4F01C1BE0671C9C8859","sizeBytes":4449},"review":null,"source":{"repositoryUrl":"https://github.com/transcendr/slopware-skills","path":"plugins/msw/skills/msw","license":null,"commit":"310939b43d225dec498b27ef777daf405322402e","subtreeSha":"133F27C711B2FAE54FAFA1DD5EC09B5A6DB13EA7D19DA12220A411C84F297AF8","lastSyncedAt":"2026-09-24T15:42:29.395779Z"},"reviewedAt":"2026-09-24T15:43:18.219996Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/transcendr/slopware-skills/tree/main/plugins/msw/skills/msw"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install transcendr-slopware-skills@llmmart"},{"target":"git","command":"git clone https://github.com/transcendr/slopware-skills.git"}]}