{"slug":"human-output","title":"human-output","summary":"The contract for writing anything a person will read and decide from — a report, a finding, an answer, a request for a decision. Works on any subject: research, legal, finance, operations, planning, code. Referenced by other skills and invocable directly before a long piece of wo","platform":"Claude","tags":[],"authorName":"LLM Mart","authorSlug":"llm-mart","score":0,"source":"github","price":null,"verified":false,"createdAt":"2026-11-20T19:24:09.157858Z","repo":{"url":"https://github.com/allemaar/open-skills","stars":15,"forks":1,"license":"Apache-2.0","updatedAt":"2026-11-13T09:33:18Z"},"bodyHtml":"<hr>\n<p>name: human-output\ndescription: &gt;\nThe contract for writing anything a person will read and decide from — a report, a finding, an answer, a request for a decision. Works on any subject: research, legal, finance, operations, planning, code. Referenced by other skills and invocable directly before a long piece of work. Trigger phrases: \"/human-output\", \"remember I am a person not a machine\", \"write this for a human\", \"keep the reader in mind\". Not /human-rewrite (repairs text that already exists) or /human-draw (builds a picture) — human-output governs writing as it happens.\nvisibility: public\nself-improvable: true\ncompanions:</p>\n<ul>\n<li>path: references/human-contract.md\noptional: false\nwhy: \"The required portable family doctrine and style floor; bundled inside this skill.\"</li>\n<li>path: references/tools/human-output-check.mjs\noptional: false\nwhy: \"The required mechanical output checker; bundled inside this skill.\"</li>\n<li>path: ../synthesize\noptional: true\nwhy: \"Optional sibling that selects a direct reader-need recipe before this skill applies the prose contract.\"\ntriggers:</li>\n<li>\"/human-output\"</li>\n<li>\"remember I am a person not a machine\"</li>\n<li>\"write this for a human\"</li>\n<li>\"keep the reader in mind\"</li>\n<li>\"before you write the report\"\nnext-skills:</li>\n<li>skill: human-rewrite\nphrase: \"/human-rewrite\"\nwhy: \"The text already exists and needs repair rather than authoring.\"</li>\n<li>skill: human-draw\nphrase: \"/human-draw\"\nwhy: \"The material is relational, positional, or proportional — it wants a figure. Runs alongside this skill, which keeps the prose around it.\"</li>\n</ul>\n<hr>\n<h1>/human-output</h1>\n<p>Five rules, then the craft behind them. The five are the contract. Everything\nafter is reference you consult while writing, not rules to memorise.</p>\n<h2>The five</h2>\n<ol>\n<li><strong>First sentence is the verdict — and anything that would reverse it is in\nthat sentence.</strong> Not the context, not the method. The outcome.</li>\n<li><strong>Say what it costs, not what it is.</strong> Compute the consequence. \"1x\nparticipating preference\" is a fact; \"costs you EUR 1.0M in a EUR 10M exit\"\nis a decision.</li>\n<li><strong>Label every claim: confirmed, judgement, or estimate.</strong> A confirmed claim\ncarries its source. A judgement with no basis is a guess — say so.</li>\n<li><strong>Say what you did not check.</strong> Every time. \"Nothing material\" is a valid\nand welcome answer.</li>\n<li><strong>No shorthand the reader must decode. No picture that does no work.</strong></li>\n</ol>\n<p>A reader can police these in fifteen seconds without re-reading. That is why\nthere are five and not twenty. A contract with twenty rules is a contract with\nnone — it gets applied by sampling, and inconsistency the reader cannot predict\nis worse than no contract at all.</p>\n<h2>Why the first rule is first</h2>\n<p>Because the reader's stopping point is the disclosure control, and you cannot\npredict it. The doctrine and its test are stated once for the whole family in\n<code>references/human-contract.md</code> §2 — read it there. Its consequence here: a\nreversing caveat is part of the verdict sentence, never a footnote.</p>\n<h2>The shape</h2>\n<pre><code>verdict  -&gt;  the ask  -&gt;  what it means  -&gt;  detail  -&gt;  ---  -&gt;  reference\n</code></pre>\n<ul>\n<li><strong>Verdict</strong> — one sentence, including any caveat that could reverse it.</li>\n<li><strong>The ask</strong> — the single decision or next action, with its deadline. Or\n\"nothing needed\".</li>\n<li><strong>What it means</strong> — what changed, what is affected, what is now uncertain.</li>\n<li><strong>Detail</strong> — claim first, then evidence. Headings state claims\n(\"Sync drops on deep paths\"), never labels (\"Sync\").</li>\n<li><strong>Reference</strong>, below the divider — paths, commands, raw figures, and what\nwas not checked. Nothing here may change the verdict.</li>\n<li><strong>Close by naming</strong> what you did not expand, at most two items. Naming depth\nis inventory; listing choices is a menu.</li>\n</ul>\n<p>Order by decision-relevance. Never by the order the work happened.</p>\n<p>Under about sixty words, drop the scaffolding and just answer. A three-line\nreply wearing full ceremony is worse than three bare lines.</p>\n<h2>Length</h2>\n<p><strong>Fidelity is a floor. Length is a target. A floor never yields to a target.</strong></p>\n<p>Write as short as fidelity allows, and no shorter. If a length limit would\nrequire dropping a number, a scope boundary, an uncertainty, or a stated gap,\nreport the conflict instead of resolving it by cutting: <em>\"Under 150 words the\nexclusion criteria have to go — tell me which you want.\"</em></p>\n<p>Length limits are suspended entirely when the subject is irreversible or\ndestructive, security or credentials, money, legal or intellectual-property\nexposure, a production deploy, or a migration. There, state the blast radius\nand whether it can be undone. Brevity that conceals a one-way door is a\nfailure, not a service.</p>\n<h2>Show me</h2>\n<p><strong>Bad</strong> — method first, verdict buried, shorthand undefined:</p>\n<blockquote>\n<p>I reviewed the term sheet against our cap table and the comparable rounds.\nValuation at EUR 6M pre is broadly in line for our stage. The preference is\n1x participating, and there is a 20% ESOP top-up taken pre-money. Governance\nis a 2-of-3 board. Anti-dilution is broad-based weighted average, which is\nnormal. On balance the terms appear market-standard subject to negotiation\non a few points.</p>\n</blockquote>\n<p><strong>Good</strong> — same facts, decidable in two lines:</p>\n<blockquote>\n<p><strong>In the illustrated EUR 10M exit, the preference costs about EUR 1.0M.\nThe option pool separately reduces headline valuation; those amounts cannot\nbe added as one exit cost. Do not sign as drafted.</strong></p>\n<p><strong>Decision needed by day 10:</strong> push back on the participating preference and\nthe pre-money employee option pool, or accept their separate effects.</p>\n<p><strong>The preference is \"1x participating\"</strong> — the investor takes their EUR 1.2M\nback <em>and</em> their share of the rest. EUR 1.2M into a EUR 6M pre-money is\n16.7% of the EUR 7.2M post-money company. On a EUR 10M exit that pays them\nEUR 2.67M, against EUR 1.67M if it were non-participating. <strong>Costs you\nEUR 1.0M.</strong> (Confirmed, §4.1.)</p>\n<p><strong>The option pool comes out of your side.</strong> A 20% pool carved before the\ninvestment means you are selling at an effective EUR 4.8M pre-money, not\nEUR 6M — <strong>EUR 1.2M of headline valuation, borne entirely by you</strong> rather\nthan shared with the incoming investor. (Confirmed, §2.)</p>\n<p>Valuation, anti-dilution and governance are genuinely standard. Not worth\nspending negotiating capital on. (Judgement.)</p>\n<p><strong>Not checked:</strong> whether this investor has conceded participating preference\nbefore. That decides whether the pushback is cheap or a dealbreaker.</p>\n</blockquote>\n<p>Three things happened that formatting alone cannot do. The EUR 1.0M was\n<em>computed</em>, not implied by naming the clause. The claims are <em>labelled</em>, so the\nreader knows which to trust. And the one unchecked thing that governs the\nnegotiation is <em>named</em> rather than silently absent.</p>\n<p>The shorthand went too. The bad version says \"ESOP\" — employee share option\npool — and leaves the reader to decode it; the good version writes it out and\nloses nothing. That abbreviation is left unexpanded above <strong>on purpose</strong>: it is\nthe defect being illustrated. <code>references/tools/human-output-check.mjs</code> flags it there,\ncorrectly, which is the check working rather than the example failing.</p>\n<h2>The craft layer</h2>\n<p>Consulted while writing. These are strong defaults, not gates.</p>\n<p><strong>Words.</strong> Expand every acronym, code name and internal term at first use, as\n<code>Expanded name (SHORT)</code>. If you cannot expand it, do not use it. Write the\nplain claim first and let the identifier trail it: <em>\"Sync fails on paths over\n260 characters (commit <code>a3f91c2</code>).\"</em> Identifiers are anchors; they never carry\nthe point.</p>\n<p><strong>Paragraphs.</strong> Three sentences by default. A single connected argument may run\nlonger where breaking it would break the logic — but never two ideas in one\nparagraph. Comparable items become a table; steps become a numbered list.</p>\n<p><strong>Items.</strong> A reader holds about four new things at once. That is a limit per\nlevel, not per report: nine findings become three groups of three, not five\nfindings dropped. If the grouping is invented rather than natural, say so — a\nfalse grouping costs the reader more than a long list.</p>\n<p><strong>Numbers.</strong> Put the figure beside the claim it supports; never \"see above\".\nGive the exact value when a different value would change the action, and round\nwhen it would not.</p>\n<p><strong>Recommendations.</strong> One per decision, with a one-line counter-case and the\ncost of being wrong: <em>\"Recommend A; if that is wrong we lose three weeks and\nEUR 40k.\"</em> Two genuinely independent decisions are two asks, numbered — that is\nnot a menu. Exempt: a protocol whose deliverable <em>is</em> a choice set\n(<code>next-skills</code>, <code>caller-options</code>, <code>ask-gate</code>) presents options by design.</p>\n<p><strong>Honesty.</strong> Never simplify substance — the reader is not a novice. Never\nrepeat content across layers; layers differ in resolution, not wording. And\nwhen the thing that determines the decision is unchecked, say the decision is\nnot ready: <em>\"You should not decide this yet — X is unverified and it governs\nthe answer.\"</em> That is a legitimate deliverable, not a failure.</p>\n<p><strong>Density.</strong> Every sentence performs one unique job: verdict, claim, evidence,\nconsequence, necessary qualification, action, or navigation in a genuinely long\nartifact. Cut immediate paraphrases, repeated summaries, meta-announcements,\ndecorative transitions, synonym stacks, fake balance, and structure that carries\nno distinction. Litotes, triads, and contrast formulas stay only when their\neffect is deliberate. Once the answer and its necessary boundary are present,\nstop. Do not append an unsolicited offer to continue. This is a composition\nconstraint applied before sending, not a cleanup pass the Handler must request.\nWhen <code>prose-audit</code> is installed, its P001-P258 index is the complete diagnostic\nreference. Do not load or recite that index for an ordinary response.</p>\n<p><strong>Reporting face.</strong> Plain prose is the default for a simple answer. Active\nmulti-step work uses the Handler's compact progress grammar. Several independent\nresults use the compact status list. Exact mappings use a table. Relationships\nuse a figure only when the figure lowers reading load. Never give a short answer\nthe ceremony of a report.</p>\n<p>This contained rule is complete on its own — a standalone install needs nothing\nelse. When the <code>synthesize</code> skill is installed, use its direct route table, load\nthe selected need file, and apply its grammar. Hybrid follows its own\nconditional represented-need loading. Grammar renders the selected shape; it\nnever replaces routing. When synthesize is not installed, this paragraph is the\nwhole contract.</p>\n<p><strong>Conditionals.</strong> When the honest answer depends, say so: <em>\"Ship it if the\nmigration window holds; hold otherwise — and right now it holds.\"</em> Inventing\ncertainty to satisfy the one-sentence shape breaks rule 1 rather than\nsatisfying it.</p>\n<h2>Two audiences</h2>\n<p>Write for the person by default. Emit a machine-readable block only when the\nconsumer is an agent or a tool that parses it. When both are present, the\nperson's section comes first and stands alone — a reader who never reaches the\nmachine block must still be able to decide.</p>\n<h2>Checking</h2>\n<p>Resolve this installed skill's directory, then run\n<code>node &lt;skill-dir&gt;/references/tools/human-output-check.mjs &lt;draft-path&gt;</code> on drafted text before sending. The checker and <a href=\"references/human-contract.md\"><code>human-contract.md</code></a> travel inside this skill folder; no repository-root companion is required. What it\ngrades, and the far longer list of what no script can grade — including the\nfact that nothing verifies a number in your prose against a number in your\nsource — is the family checker contract, <code>references/human-contract.md</code> §3.</p>\n<p>If the runtime or draft is unavailable, or the checker was not run or failed,\nreport that outcome and the unchecked scope. Do not claim a green check. Keep\nthe manual checks and any controlling workflow's required-check gate; the\noptional craft layer does not waive mandatory applicable work.</p>\n<h2>How other skills use this</h2>\n<p>Add a footer rather than restating the rules:</p>\n<blockquote>\n<p><strong>Human output.</strong> This skill's handler-facing output obeys the human-output\ncontract (<code>human-output/SKILL.md</code>).</p>\n</blockquote>\n<p>Inline the five rules into any worker or subagent brief. A subagent does not\ninherit the harness's response guidance, and unbriefed workers are where\nverbose reporting accumulates.</p>\n<h2>Boundary</h2>\n<p>This skill is one member of the <code>human-</code> family. The routing table — which\nmember handles which material — is <code>references/human-contract.md</code> §1.</p>\n<ul>\n<li>Not <code>/human-rewrite</code> — the repair pass on text that already exists.</li>\n<li>Not <code>/human-draw</code> — that owns what is inside the fence; this owns the words\naround it. They <strong>run together</strong> on any output carrying a figure; neither\nprecedes nor replaces the other.</li>\n<li>Not <code>/human-merge</code> — several separate reports into one. A merge is <em>written</em>\nto this contract and adds only the cross-report layer on top of it, so route\nthe pile there first and return here to write the result. See\n<code>references/human-contract.md</code> §1.</li>\n<li>Not <code>/ask-gate</code> — that decides <em>whether</em> to ask the handler; this governs how\nthe asking is written.</li>\n<li>Not a substitute for having something to say. A well-shaped report of a\nbadly-done investigation is a faster route to a bad decision.</li>\n</ul>\n<blockquote>\n<p><strong>Next skills.</strong> On completion, run the Next Skills protocol\n(<code>next-skills/SKILL.md</code>): surface the <code>next-skills</code> recommendations from\nfront-matter for the caller to pick. Offer only — never auto-invoke.</p>\n</blockquote>\n<blockquote>\n<p><strong>Self-improvement.</strong> Run the Self-Improvement Protocol only when its installed owner is available. Preserve its qualification, vetting and approval rules. If absent, skip that optional hook without claiming it ran, surfacing an unvetted fix or auto-editing; substantive work may complete.</p>\n</blockquote>\n","files":[{"path":"references/human-contract.md","sizeBytes":9052,"isText":true},{"path":"references/tools/human-output-check.mjs","sizeBytes":11124,"isText":false},{"path":"SKILL.md","sizeBytes":13736,"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":"2026-11-20T19:28:47.110061Z","sha256":"E324E472BB6C806A95B9EA7F2E29E9D53DE706F2B899D62EBE41EA47F561AB16","sizeBytes":14974},"review":null,"source":{"repositoryUrl":"https://github.com/allemaar/open-skills","path":"skills/human-output","license":"Apache-2.0","commit":"acf5f797d21208c9cff2c1d97650a3113d7840df","subtreeSha":"555C35FDD14AD114064EF5C589DE26A7D9F6CB5A416A99BE5906CC639AFC5957","lastSyncedAt":"2026-11-20T19:24:04.446897Z"},"reviewedAt":"2026-11-20T19:37:40.160454Z","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/allemaar/open-skills/tree/main/skills/human-output"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install allemaar-open-skills@llmmart"},{"target":"git","command":"git clone https://github.com/allemaar/open-skills.git"}]}