b-open-io/visual-proposal

Visual Proposal

A visual proposal is the forward-looking sibling of a visual recap. A recap describes a change that was just made; a proposal describes a design, an options space, or a plan that has NOT been built yet, so a reviewer — often someone with no prior context, sometimes a friend or an external maintainer — can understand it, weigh it, and decide. The deliverable is a single self-contained, theme-aware HTML page, published as an Artifact when the Artifact tool is available (default-private; the user shares it), otherwise written locally and opened in the browser.

When the optional artifact-design skill is installed, load it before writing the page for additional craft guidance. Do not block when it is unavailable; apply the self-contained, theme, layout, and accessibility requirements in this skill directly. This skill governs the content and stance of a proposal.

The two rules that make or break a proposal

Rule 1 — Present options neutrally. Do not decide for the reviewer.

The single most common failure is editorializing: stamping options "Recommended" / "Reject" and collapsing a live decision into a foregone conclusion. A proposal's job is to arm the reviewer with facts, not to hand them a verdict. When the page presents alternatives:

• Give each option the SAME visual weight and the SAME structure — mechanism, how it works, what it costs, where it fits, where it doesn't.
• Present commonalities (what all options share), differences (the axes they diverge on), and challenges (each option's honest costs and open questions) — explicitly, as their own sections.
• An option that "no server implements yet" or "is non-standard" is described by that fact, not branded a loser — an implementation gap common to every path is not a disqualifier. State the fact; let the reviewer judge its weight.
• Only include an author's recommendation if the user explicitly asked for one. If they did, put it in a clearly separated "author's lean" block that names it as a lean and an open call — never as a badge on one option that visually pre-empts the choice. Default: no author recommendation at all. (This governs the author's own verdict. A judging bench — a panel of named, attributed, usually-split verdicts with flip-conditions — is a different thing and is part of the default flow for a real decision; see the judging bench section below.)

Stop after writing "the one to reject" or "best option" and replace the verdict with the underlying evidence. Leave the decision to the reviewer.

Rule 2 — Diagram over prose. Some things must be shown, not told.

A proposal earns its "visual" name by making the reader see the idea. Wherever a relationship, a structure, or a difference is clearer drawn than written, draw it. Reach for diagrams (SVG or Canvas, never external images) for:

• Structure — how the pieces fit: transaction/output layouts, data trees, component graphs, request flows. A labeled box-and-wire diagram beats a paragraph describing the same shape.
• Comparison — a matrix of options × axes, so the reader scans differences at a glance instead of reading three parallel descriptions and holding them in their head. Encode state in form (a filled cell, a chip, a check/dash), not just words.
• Before / after and flow — what a thing looks like today vs under the proposal; the sequence of steps.
• The thesis — the hero should show the most characteristic thing about the idea, drawn, before any prose.

Prose then does what prose is good at: intent, nuance, the "why." The test: if a section is three paragraphs comparing structures, it probably wants to be a diagram plus one caption.

Multi-agent advocacy — opposing representatives (the DEFAULT for real decisions)

When a proposal's core is a key decision between competing options (or a series of them), a panel of advocates is the default treatment, not an add-on — do it automatically, without being asked. The strongest neutrality comes not from one author trying to be even-handed, but from giving each option a genuine advocate from the real bОpen roster, each shown with their avatar, name, and role. Skip the panel only for a page with no real decision (a pure explainer, status update, or single-approach pitch), or when the user explicitly asks for a plain writeup. Otherwise: dispatch one real roster agent per option (see "Casting the panel" below), each assigned to steelman a different option:

• Give every advocate the SAME shared context (the problem, all options, the confirmed facts) and ask each for: the strongest case for THEIR option, its honest challenges (don't hide them), where it fits best, and a direct rebuttal to the other options. Require a structured return so the viewpoints are comparable.
• Pick advocates whose expertise or disposition genuinely leans toward the option they carry (an efficiency-minded agent for the minimal option, a standards-minded one for the standard option, a systems/composability one for the general option) — the reasoning styles should differ, not just the labels.
• In the page, present each key decision as a panel of the N viewpoints, equal visual weight, each attributed to its advocate. The reviewer reads three arguing positions per decision and decides — which is exactly the goal.
• This does NOT override Rule 1: add no author verdict. The panel is the advocates arguing; the author stays the neutral host. The judging bench (below) then rules by default — a panel of attributed, usually-split judges, which is not the author deciding.

Run the advocates concurrently because they are independent. Ground every advocate's claims in the same evidence used by the host; an advocate that invents facts is worse than no advocate.

Casting the panel — real agents and their avatars

The advocates and judges are real named agents from the bОpen roster, not invented personas — that credibility is the whole point. Two steps:

1. Pick the agents. Agent(bopen-tools:front-desk) returns the team and who fits what, or read the agents/*.md files (each has a display_name and a role). Match each option to an agent whose real expertise leans that way (a code-auditor for correctness, an optimizer for efficiency, a standards/interop agent, a domain specialist). Advocates and judges must be distinct agents.

2. Embed their avatars with the bundled script. The Artifact CSP blocks external <img src>, so avatars must be inlined as data URIs — this skill's scripts/embed-avatars.sh automates it (fetch → downscale → emit a window.AV map). Slug = display_name lowercased with non-alphanumerics → - ("Uno Satoj" → uno-satoj):

   bash scripts/embed-avatars.sh uno-satoj torque maxim
   

   Paste the printed <script> block into the page and give each panelist an <img data-a="uno-satoj">; the block wires every src on load. Show the avatar + name + one-line role on each card. If the script warns a slug has no avatar, use that agent's initials in a colored circle — never ship a faceless panel.

A premise-challenging voice

The advocates above argue among the presented options. Often the most valuable viewpoint disputes the framing itself — the whole option space rests on an assumption worth questioning ("this doesn't need to be a collection at all — mint a fungible token"; "don't build this, buy it"; "the real fork is upstream of these three"). Give such a voice a card too, on equal footing. But then make one call, and make it honestly:

• If the voice only questions the framing (no concrete, buildable alternative), keep it as a marked premise-challenge: it argues a different axis, so it is not scored against the option advocates and the bench does not rule on it. Fold what it surfaces into the open questions as an upstream decision the reviewer settles first.
• If the voice proposes a concrete, buildable alternative (a real fourth option — "mint a BSV-21 token instead"), it is NOT a sidebar. Promote it to a first-class option: full cross-examination (the other advocates attack it by name, it attacks each of them), and the judging bench re-runs to include it. Sidelining a real competing option as "just a premise-challenge" is the most common way this skill under-serves a live decision — if it can be built and compared, it gets argued and judged like everything else.

The tell is buildability: a different axis can still be a real option. When it is, verify its feasibility with the same grounding rigor as every other option (read the actual code/spec — see the grounding rule) before letting an advocate champion it, so the fourth card is as evidence-backed as the first three.

The standard flow: advocates → cross-examination → judging bench → the CEO's final call

For a real decision, run all four stages by default — advocacy, cross-examination, a judging bench, AND the CEO's final call. This is the shape that makes a proposal decision-useful rather than a bare list, and it is what this skill is for. Do NOT treat the bench or the CEO as an opt-in the user must request; running advocates but no judges is the most common way this skill under-delivers.

• Cross-examination (default). Each advocate argues for their option AND against each opponent by name — the sharpest, most specific weakness of each rival relative to their own. The page then shows, per option, both the case for it and how its rivals attack it, so the reviewer sees the collisions, not three monologues.
• The judging bench (default). After the advocates, dispatch a separate set of agents — distinct from every advocate — as judges. This runs by default and does not violate Rule 1, because a bench is a panel of attributed, usually-split verdicts with flip-conditions, not the author handing down one foregone conclusion. Give each judge the full record: the problem, options, confirmed facts, and all advocates' cases and rebuttals. Assign a distinct evaluative lens (e.g. correctness/robustness, efficiency/simplicity, ecosystem adoption/interop). Each judge names their lens, picks a winner and runner-up, states the deciding factor, and — the most decision-useful part — the single condition that would FLIP their verdict ("would flip if …"). Those flip-conditions are the real tiebreakers, so keep them: they tell the reviewer exactly which fact to go check. Present the verdicts in a clearly separated judgment section — a tally, each judge's one-line rationale, and their flip-condition. Keep verdict badges OFF the neutral option cards.

Skip the bench only when there is genuinely no decision to resolve (a pure explainer or status page), or the user explicitly asks for options-only with no assessment. Fix the judging lenses and evaluation rubric before showing judges the advocacy record. Present every judge's verdict, including dissent; never collapse a tally into objective truth. This stays within Rule 1 because the author adds no verdict of their own — the split, attributed bench does, and the human reviewer still makes the final call.

The CEO's final call (runs LAST, after everything else)

After every advocate and every judge has weighed in, Chief — the CEO (Agent(bopen-tools:ceo)) — comes in and makes the final call. He is the last stage, never run until the rest is complete. He is like one more judge, but a decisive one: he weighs the whole thing holistically for the business and the user, not the tech alone — the true cost of technical debt (priced as interest), scope creep, reversibility, and whether the change actually serves the user. Give him the FULL record: the problem, the options, the confirmed facts, every advocate's case and rebuttal, and every judge's verdict and flip-condition.

Unlike the neutral judges, the CEO does decide — that is his role, and it does not break Rule 1 because it is an attributed agent's recommendation from the business's vantage, not the author editorializing, and the human still overrides. Present his call as a distinct final section with his avatar, name, and role: his decision stated plainly, one or two sentences of business/user reasoning, and the single thing that would change his mind. The reviewer's Agree / Agree-with-caveats / Disagree control on this call is the required CEO selector in the interactive choices (below).

Show the agents — avatar, name, role. When advocates or judges appear, give each a real identity: their avatar, name, and one-line role on the card that carries their argument or verdict. It makes the page feel authored by the panel rather than narrated about them, and it lets the reviewer track a voice across the page. For a fact that several agents independently verified (e.g. all advocates confirmed it against the code), attribute it visually — cluster those agents' avatars on that fact's card (straddling its top edge reads as a group sign-off) instead of writing "everyone agreed." An avatar cluster conveys consensus and who-verified at a glance.

Readability — run humanize on every voice

The whole point is a page that's easy to follow. Text written by a panel of agents drifts into AI tics (setup-and-dismiss, tricolons, hedged filler) that make it a slog. So: ask every dispatched agent — each advocate, each judge, and the CEO — to run Skill(bopen-tools:humanize) on its returned prose before it hands it back. Then, when you assemble the page, you (the host) run Skill(bopen-tools:humanize) over the final copy — the theses, verdicts, captions, and the CEO's decision — so the whole thing reads in one clear human voice. Crisp, skimmable, no throat-clearing.

The grounding rule

Every claim, number, code shape, and diagram must trace to real research, real code, or a real spec — never invented. If a fact is unverified, label it (e.g. "unverified — needs a live check") rather than presenting it as settled. A proposal that fudges its facts wastes the reviewer's trust and their time. Pull the substance from the session's actual findings (files read, code cited, specs quoted); if the substance isn't there yet, gather it before authoring.

Structure of a proposal page

Adapt to the subject — not every section always applies — but the usual spine:

1. Hero / thesis — the one-sentence claim + the most characteristic diagram.
2. Problem / context — what's wrong or missing today, concretely.
3. Landscape / what already exists — an inventory of the relevant specs, code, and tools that exist today, where each lives (repo/file/URL), and a status chip on each (shipped, draft, branch, live, legacy, frontend-only). Frame it explicitly as orientation, NOT a judging input — options are judged on their merits, not on what happens to be built already. Include it whenever a proposal touches an existing system; it grounds a reviewer (and an external maintainer) fast and prevents "wait, does this already exist?" confusion. Make every reference in this section a clickable link — to the repo, file, PR, or doc page — so the reviewer clicks straight through to the real thing.
4. The approach / principle — the core idea, and the constraint it respects (e.g. "changes X, leaves Y untouched"). Before/after helps here.
5. The decision(s) — where options exist, the neutral comparison (Rule 1), diagram-led (Rule 2): a matrix, per-option structure diagrams, commonalities, differences, challenges. This is where the advocate panel, judging bench, and CEO's final call live (the standard flow above).
6. Architecture — how it's put together if adopted.
7. Evidence — what's confirmed vs open (cite the sources).
8. Roadmap / plan — sequenced steps, dependencies. Number only if it's a real sequence.
9. Open questions — what still needs deciding, each with the trade-off.
10. Your call — selectable option cards for every decision + a copy-response button, so the reviewer answers in-page and pastes a clean, versioned reply back to the agent. Always includes the Agree / Agree-with-caveats / Disagree control on the CEO's final call. See references/interactive-choices.md for the drop-in component (CSS + markup + copy script).

Craft (via artifact-design)

• Self-contained: inline all CSS/JS, embed assets as data URIs, no external requests (the Artifact CSP blocks them). Diagrams are SVG/Canvas, not <img>.
• Link out to browsable sources: whenever the page names something with a URL — a repo, file, PR, issue, doc page, deployed app — make it a clickable <a target="_blank" rel="noreferrer"> so the reviewer clicks straight through to the real thing. The CSP blocks embedding external resources, not <a> navigation, so outbound links open fine. Don't leave a repo/file path as dead text when it could be a link.
• Theme-aware: token-level light/dark, prefers-color-scheme plus a data-theme toggle that overrides both ways. SVG paint via CSS classes or style, never fill="var(--x)" presentation attributes — var() does not resolve there and the diagram renders black.
• Type carries it: pair a display and a body/utility face; keep prose near 65ch; a subject-true monospace for any code/technical material.
• Ground the palette and vocabulary in the subject's own world.
• Respect prefers-reduced-motion; give focus a visible state; no horizontal body scroll (wide diagrams/tables get their own overflow-x: auto).

Publish

Write the page to the scratchpad dir, then call the Artifact tool with its path, a one-line description, a stable subject favicon, and a version label. Redeploy the SAME file path to keep the URL stable across revisions; only mint a new URL for a genuinely different proposal. Remind the user it's private until they share it from the page's share menu.

Iterating

Proposals get revised as decisions settle. When the user changes an option's framing, flattens a decision, or adds facts, edit the same file and republish to the same URL with a bumped label. Keep the neutral stance unless the user asks for a recommendation.

When the user reframes the decision itself — not just picks a winner, but changes which options are in play or the principle that separates them — re-run the advocacy, don't just rewrite prose. Re-brief the advocates on the new decision points and let them re-propose (with fresh cross-examination), then re-fire the judging bench on the new points of contention. A reframe can turn a split verdict into a unanimous one (or the reverse); a proposal that only edits its conclusion text after a reframe is stale under the hood. Ground the re-run with the same rigor as the first: verify the new facts against real code or spec before an advocate argues them. Mark the prior round as superseded rather than deleting it — the progression (round 1 → reframe → round 2) is itself useful to the reviewer.