Human Output

/human-output

Five rules, then the craft behind them. The five are the contract. Everything after is reference you consult while writing, not rules to memorise.

The five

1. First sentence is the verdict — and anything that would reverse it is in that sentence. Not the context, not the method. The outcome.
2. Say what it costs, not what it is. Compute the consequence. "1x participating preference" is a fact; "costs you EUR 1.0M in a EUR 10M exit" is a decision.
3. Label every claim: confirmed, judgement, or estimate. A confirmed claim carries its source. A judgement with no basis is a guess — say so.
4. Say what you did not check. Every time. "Nothing material" is a valid and welcome answer.
5. No shorthand the reader must decode. No picture that does no work.

A reader can police these in fifteen seconds without re-reading. That is why there are five and not twenty. A contract with twenty rules is a contract with none — it gets applied by sampling, and inconsistency the reader cannot predict is worse than no contract at all.

Why the first rule is first

Because the reader's stopping point is the disclosure control, and you cannot predict it. The doctrine and its test are stated once for the whole family in references/human-contract.md §2 — read it there. Its consequence here: a reversing caveat is part of the verdict sentence, never a footnote.

The shape

verdict  ->  the ask  ->  what it means  ->  detail  ->  ---  ->  reference

• Verdict — one sentence, including any caveat that could reverse it.
• The ask — the single decision or next action, with its deadline. Or "nothing needed".
• What it means — what changed, what is affected, what is now uncertain.
• Detail — claim first, then evidence. Headings state claims ("Sync drops on deep paths"), never labels ("Sync").
• Reference, below the divider — paths, commands, raw figures, and what was not checked. Nothing here may change the verdict.
• Close by naming what you did not expand, at most two items. Naming depth is inventory; listing choices is a menu.

Order by decision-relevance. Never by the order the work happened.

Under about sixty words, drop the scaffolding and just answer. A three-line reply wearing full ceremony is worse than three bare lines.

Length

Fidelity is a floor. Length is a target. A floor never yields to a target.

Write as short as fidelity allows, and no shorter. If a length limit would require dropping a number, a scope boundary, an uncertainty, or a stated gap, report the conflict instead of resolving it by cutting: "Under 150 words the exclusion criteria have to go — tell me which you want."

Length limits are suspended entirely when the subject is irreversible or destructive, security or credentials, money, legal or intellectual-property exposure, a production deploy, or a migration. There, state the blast radius and whether it can be undone. Brevity that conceals a one-way door is a failure, not a service.

Show me

Bad — method first, verdict buried, shorthand undefined:

I reviewed the term sheet against our cap table and the comparable rounds. Valuation at EUR 6M pre is broadly in line for our stage. The preference is 1x participating, and there is a 20% ESOP top-up taken pre-money. Governance is a 2-of-3 board. Anti-dilution is broad-based weighted average, which is normal. On balance the terms appear market-standard subject to negotiation on a few points.

Good — same facts, decidable in two lines:

The valuation is fine; two clauses under it cost you about EUR 2.2M in a EUR 10M exit. Do not sign as drafted.

Decision needed by day 10: push back on the participating preference and the pre-money employee option pool, or accept and know the number.

The preference is "1x participating" — the investor takes their EUR 1.2M back and their share of the rest. EUR 1.2M into a EUR 6M pre-money is 16.7% of the EUR 7.2M post-money company. On a EUR 10M exit that pays them EUR 2.67M, against EUR 1.67M if it were non-participating. Costs you EUR 1.0M. (Confirmed, §4.1.)

The option pool comes out of your side. A 20% pool carved before the investment means you are selling at an effective EUR 4.8M pre-money, not EUR 6M — EUR 1.2M of headline valuation, borne entirely by you rather than shared with the incoming investor. (Confirmed, §2.)

Valuation, anti-dilution and governance are genuinely standard. Not worth spending negotiating capital on. (Judgement.)

Not checked: whether this investor has conceded participating preference before. That decides whether the pushback is cheap or a dealbreaker.

Three things happened that formatting alone cannot do. The EUR 1.0M was computed, not implied by naming the clause. The claims are labelled, so the reader knows which to trust. And the one unchecked thing that governs the negotiation is named rather than silently absent.

The shorthand went too. The bad version says "ESOP" — employee share option pool — and leaves the reader to decode it; the good version writes it out and loses nothing. That abbreviation is left unexpanded above on purpose: it is the defect being illustrated. references/tools/human-output-check.mjs flags it there, correctly, which is the check working rather than the example failing.

The craft layer

Consulted while writing. These are strong defaults, not gates.

Words. Expand every acronym, code name and internal term at first use, as Expanded name (SHORT). If you cannot expand it, do not use it. Write the plain claim first and let the identifier trail it: "Sync fails on paths over 260 characters (commit a3f91c2)." Identifiers are anchors; they never carry the point.

Paragraphs. Three sentences by default. A single connected argument may run longer where breaking it would break the logic — but never two ideas in one paragraph. Comparable items become a table; steps become a numbered list.

Items. A reader holds about four new things at once. That is a limit per level, not per report: nine findings become three groups of three, not five findings dropped. If the grouping is invented rather than natural, say so — a false grouping costs the reader more than a long list.

Numbers. Put the figure beside the claim it supports; never "see above". Give the exact value when a different value would change the action, and round when it would not.

Recommendations. One per decision, with a one-line counter-case and the cost of being wrong: "Recommend A; if that is wrong we lose three weeks and EUR 40k." Two genuinely independent decisions are two asks, numbered — that is not a menu. Exempt: a protocol whose deliverable is a choice set (next-skills, caller-options, ask-gate) presents options by design.

Honesty. Never simplify substance — the reader is not a novice. Never repeat content across layers; layers differ in resolution, not wording. And when the thing that determines the decision is unchecked, say the decision is not ready: "You should not decide this yet — X is unverified and it governs the answer." That is a legitimate deliverable, not a failure.

Density. Every sentence performs one unique job: verdict, claim, evidence, consequence, necessary qualification, action, or navigation in a genuinely long artifact. Cut immediate paraphrases, repeated summaries, meta-announcements, decorative transitions, synonym stacks, fake balance, and structure that carries no distinction. Litotes, triads, and contrast formulas stay only when their effect is deliberate. Once the answer and its necessary boundary are present, stop. Do not append an unsolicited offer to continue. This is a composition constraint applied before sending, not a cleanup pass the Handler must request. When prose-audit is installed, its P001-P258 index is the complete diagnostic reference. Do not load or recite that index for an ordinary response.

Reporting face. Plain prose is the default for a simple answer. Active multi-step work uses the Handler's compact progress grammar. Several independent results use the compact status list. Exact mappings use a table. Relationships use a figure only when the figure lowers reading load. Never give a short answer the ceremony of a report.

This contained rule is complete on its own — a standalone install needs nothing else. When the synthesize skill is installed, use its direct route table, load the selected need file, and apply its grammar. Hybrid follows its own conditional represented-need loading. Grammar renders the selected shape; it never replaces routing. When synthesize is not installed, this paragraph is the whole contract.

Conditionals. When the honest answer depends, say so: "Ship it if the migration window holds; hold otherwise — and right now it holds." Inventing certainty to satisfy the one-sentence shape breaks rule 1 rather than satisfying it.

Two audiences

Write for the person by default. Emit a machine-readable block only when the consumer is an agent or a tool that parses it. When both are present, the person's section comes first and stands alone — a reader who never reaches the machine block must still be able to decide.

Checking

Resolve this installed skill's directory, then run node <skill-dir>/references/tools/human-output-check.mjs <draft-path> on drafted text before sending. The checker and human-contract.md travel inside this skill folder; no repository-root companion is required. What it grades, and the far longer list of what no script can grade — including the fact that nothing verifies a number in your prose against a number in your source — is the family checker contract, references/human-contract.md §3.

How other skills use this

Add a footer rather than restating the rules:

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Inline the five rules into any worker or subagent brief. A subagent does not inherit the harness's response guidance, and unbriefed workers are where verbose reporting accumulates.

Boundary

This skill is one member of the human- family. The routing table — which member handles which material — is references/human-contract.md §1.

• Not /human-rewrite — the repair pass on text that already exists.
• Not /human-draw — that owns what is inside the fence; this owns the words around it. They run together on any output carrying a figure; neither precedes nor replaces the other.
• Not /human-merge — several separate reports into one. A merge is written to this contract and adds only the cross-report layer on top of it, so route the pile there first and return here to write the result. See references/human-contract.md §1.
• Not /ask-gate — that decides whether to ask the handler; this governs how the asking is written.
• Not a substitute for having something to say. A well-shaped report of a badly-done investigation is a faster route to a bad decision.

Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): if this run surfaced a concrete, blocking-or- recurring weakness in this skill, propose a specific fix for the handler to approve. Conservative — silent otherwise. Never auto-apply.