opinion

name	opinion
description	Route the current context to one expert persona and return that lens's candid second opinion. Auto-classifies to the best-fit persona, or name one explicitly. Runs the persona as a spawned subagent so its reasoning stays out of the main context. Roster spans engineering, product, go-to-market, and craft. Out of scope: generating new designs (use /design), arguing both sides (use /debate), stress-testing one plan (use /probe), scoring confidence in a claim (use /grade). Invoke explicitly with /opinion.
disable-model-invocation	true

opinion

Get one expert's take on what's on the table. /opinion reads the current context — the artifact, diff, plan, or question under discussion — picks the persona whose lens fits best, and spawns that persona as a subagent to return a candid second opinion in its own voice. One voice, not a committee: the value is a sharp read from a specific vantage, not consensus mush.

The persona runs as a spawned subagent. The main session distills what to opine on and hands it off; the persona reasons in isolation and returns its verdict. The main context stays clean and the opinion stays uncolored by the surrounding chat.

Roster

persona	lens	routes on
staff-engineer	architecture, tradeoffs, maintainability, technical risk	system design, refactors, "how should I build", scaling, build-vs-buy, tech debt
security-engineer	threat model, attack surface, data & secrets, trust boundaries	auth, tokens, input handling, permissions, PII, untrusted data, deploys
performance-engineer	latency, bundle size, rendering, memory, measurement	"slow", janky, large, "optimize", render path, perf budgets
product-manager	user value, scope, prioritization, why-now	features, roadmap, MVP, cut-or-keep, "is this worth building"
product-designer	usability, interaction, hierarchy, accessibility	UI, flows, forms, on-screen copy, "is this usable", a11y
cli-toolsmith	invocation ergonomics, composability, stdout/stderr contracts	CLI flags, --help, piping, exit codes, TUIs, defaults, "is this tool ergonomic"
marketer	positioning, messaging, audience, distribution	landing copy, naming, launch, "how do I pitch this", growth
founder	viability, ROI, opportunity cost, business risk	"should I do this at all", pricing, market, time-vs-payoff
editor	clarity, concision, structure, AI tells	prose, READMEs, posts, "does this read well", tighten
technical-writer	documentation, onboarding, completeness for outsiders	docs, API reference, guides, "is this documented", reader DX

staff-engineer is the default when context is technical but no single lens dominates.

Dispatch

dir = skill base directory
arg = <text after /opinion>

// list mode
if arg in {"list", "ls"}:
    do("print the Roster table"); stop
if arg == "" and nothing is clearly on the table in the conversation:
    do("print the Roster table; say: run /opinion <persona>, or /opinion with an artifact on screen")
    stop

// 1. find what to opine on
target = do("""identify the artifact under discussion: the file / diff / plan / question the user
  most recently put on the table. If nothing is clearly on the table, ask one question —
  'what should I get an opinion on?' — and stop until answered.""")

// 2. pick the persona
if arg names a persona (exact slug or unambiguous prefix):
    persona = that slug                      // explicit override — skip the classifier
else:
    persona = do("""classify `target` against the Roster 'routes on' cues; pick the single best
      fit. Tie, or technical-but-unclear → staff-engineer. Announce in one line:
      'Routing to <persona> — <the one cue that decided it>.'""")

// 3. run the persona as a subagent
spec  = Read(dir/personas/<persona>.md)
brief = do("""assemble the subagent brief:
  - the persona file contents (lens, heuristics, output contract) as its role
  - `target`: the distilled artifact/question PLUS the file paths it lives in, so the subagent
    can Read them directly for specifics
  - any particular angle the user asked for""")
opinion = Agent(subagent_type: general-purpose, prompt: brief)

// 4. return
do("relay the persona's opinion in its own voice; do not re-summarize, average, or soften it")
do("close with one line: 'Want another lens? /opinion <persona>.'")

Reference

Output contract (every persona)

1. Verdict — one line.
2. What matters most here — 2-4 highest-leverage observations, each tied to something concrete in the target.
3. Recommendations — what to change, ordered by leverage.
4. Confidence — 1-10 with one-line reasoning (matches the global confidence convention).

Each personas/<slug>.md restates this contract, because the subagent sees only its persona file plus the target — never this SKILL.md.

Routing notes

• The one-line route announcement is load-bearing: it makes a mis-route visible. If it's wrong, re-run /opinion <persona> to force the right lens.
• Explicit /opinion <persona> always beats the classifier.
• One persona per run, by design. For contrasting lenses, run it twice — or reach for a neighbor skill below.

Adding a persona

Drop a personas/<slug>.md with frontmatter name, lens, signals, then add a Roster row. The body is the subagent's role: what it optimizes for, the questions it always asks, what it flags, its declared blind spots, and the output contract above.

Boundary with neighbors

• /opinion — one expert lens reacts to what already exists.
• /design — generates multiple new solutions before you commit.
• /debate — argues both sides of a single open question.
• /probe — stress-tests one chosen plan for holes.
• /grade — scores confidence in a claim or decision (1-10).