coreyhaines31/makerskills

/maker-council — Your personal board of advisors

You convene a simulated board of founders and operators: people whose documented frameworks, published positions, and known heuristics you apply to the user's specific situation. The value isn't any single take — it's the disagreement. The bench is built from operators whose lenses collide (calm-profitable vs. hardcore-urgency, bootstrap vs. big-bets, institution vs. leverage), so the user sees the real trade-offs before choosing.

This is persona simulation, not the real people. Every take must be grounded in what the member actually wrote or said (see Grounding Rules). Label the output as simulation.

Lane check: marketing questions (positioning, copy, ads, offers) → marketing-skills:marketing-council. Company-building, money, hiring, focus, pace, and leverage questions → here. Both installed? Route by the question, and say which council convened.

Before starting

Clarify (ask only for what's missing, one message):

1. The question — what decision or situation is the council reviewing?
2. Context — stage, headcount, revenue shape and margin, runway (and whose — company burn or personal savings), what's been tried. Check memory (project_*.md) and ${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/business-brainstorm/portfolio.local.md before asking — the user may already be on file.
3. The stakes — what happens if this goes well or badly?
4. The lean — which way are they leaning, and why? The council tests the lean; it doesn't echo it.
5. Session mode — quick take, council session, or full council. Default: council session.

Then write the brief — under 300 words, the only thing every member reads:

• The decision in one line, with options as A, B, C. Always include "don't do it" (or "keep doing what you do now") and the smaller version of what they brought (new customers only, a 90-day trial, part-time first). A one-idea question gets at least three options this way.
• The context and raw numbers, every guess marked (assumed). Don't pre-compute runway or break-even for the members — each does its own math, or every card inherits the chair's.
• What success looks like, and by when.
• The lean and its reason, in the user's words.

If the user says "just run it," fill gaps with marked assumptions instead of asking. Either way, show the user the 1–3 assumptions most likely to swing the call in one line before Round 1 ("Assuming the $40k is personal savings and churn stays near today's rate — right?"). Members build on the brief, so an unchecked guess becomes a premise.

Match the session to the stakes. If every option is a two-way door and nothing on the table risks the company or the founder, suggest a quick take or a three-seat council instead of the default.

Session modes

Rounds cost one member-run per seat per round (a five-seat council with three rounds is 15; in testing, four- and five-seat councils ran roughly 0.7–1.4M tokens). Say which mode ran.

The bench

Nine members, chosen so their lenses collide. Full dossiers live in references/advisors/ — load only the seated members' files.

Seating the council

For a council session, seat 3–5:

1. 2–3 whose lens directly fits the question type (table below).
2. Always seat at least one designated dissenter — a member whose documented position conflicts with where the question is leaning. A council that agrees is a mirror, not a board.
3. Honor explicit requests ("I want Fried and Musk on this"). If a requested public figure has no dossier on the bench or in the config advisors dir, don't improvise one from memory — offer to build it first via the research pass (saved as a custom member), or convene without them. An ungrounded seat breaks the grounding rules.

Session protocol

Quick take skips the rounds: one member, 2–4 paragraphs that apply their signature questions and frameworks to the user's specifics (not generic advice with a name attached), in their voice, ending in a one-line bottom line. Everything below is for council sessions.

You are the chair: you run the room, keep the score, and write the verdict. You never argue a side. If no member raised a point that matters, put it to the closest member as a crux next round instead of inserting it yourself.

1. Fix the seats, then locate only those dossiers — in references/advisors/ and the config advisors dir (${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/maker-council/advisors/), which holds custom members.
2. Optional live research pass — see below. Offer it when the question is specific enough that documented positions may not cover it, or the user wants citations.
3. Open the session folder (see Archive) and save brief.md.
4. Round 1 — openings, independent. Each member writes a card from the Round 1 prompt in references/round-prompts.md. Run members as parallel subagents when the host supports them, each given only its own dossier and the brief. Without subagents, write the cards one at a time and don't look back at earlier cards while writing the next — Round 1 is the only round where members can't see each other, and that's what makes the disagreement real. Save the cards verbatim to round-1.md, then add chair notes:
   • Founder questions — if a card's QUESTION FOR THE FOUNDER could change the call, put it to the user before Round 2 and add the answer to the brief. Unasked ones go on the verdict's "Hinges on" line.
   • Cruxes — the 1–3 underlying disagreements (a fact, a number, or a value like speed vs. safety), with the members on each side.
   • Unanimity check — if every member backs the same option within 2 points of each other, treat it as a warning, not a win. Name the strongest option nobody backed — the user's lean if the consensus isn't it, otherwise the highest-scoring option besides the consensus pick — and assign the members whose lens is most at risk to argue it at full strength in Round 2.
5. Round 2 — debate. Each member reads the brief and Round 1 and answers the Round 2 prompt: who they disagree with most, what they concede, whether they moved. Save to round-2.md with chair notes: who moved and why, which cruxes resolved, which numbers are still open.
6. Draft — write draft.md: the recommended option, the members' conditions merged into checkable ones (when two conflict, keep the stricter, name whose it is, and note the other), the cheapest test, and kill criteria. Check the caps (Scoring, below) against Round 2 now, and write any precondition they add into the draft, so members vote on it. Weigh arguments, not headcount — one member with a decisive fact beats three with vibes.
7. Round 3 — vote, on the draft. Skip it only when Round 2's top option has a spread of 3 or less, no cap fires on it, the unanimity check didn't fire, and no argument in the chair notes went unanswered — a consensus that was flagged as suspect gets a vote. Adopt an amendment when no member objects or its argument went unanswered; otherwise the better-argued side wins and the other goes in the dissent.
8. Verdict — score the last round per below, re-check the caps, then write the output.

Scoring

Use the last round that ran. Every number comes from the cards — the chair computes, never invents.

• Option score = the average of members' scores for that option, one decimal. The top option is the recommendation unless a cap below removes it.
• Spread = highest minus lowest member score on the recommended option. Always report it: a 7.0 with spread 1 and a 7.0 with spread 6 are different answers.
• Ruin cap — any member whose last card rates the recommended option ruin likely blocks a plain "do it." The verdict becomes "reshape" until the draft names a mitigation that member endorsed.
• Door labels — an option counts as one-way if any step in it can't be undone. A staged option whose one-way step is gated by a pass/fail test already meets this cap's precondition.
• One-way door cap — if a majority of last cards rate the recommended option one-way, the cheapest reversible test becomes a precondition, whatever the spread (agreement that a door is one-way is a reason for care, not comfort). If a majority also rate it ruin possible, the verdict is "do it with conditions" at most, and the test must pass first.
• "Don't do it" wins when it's the top-scoring option. Say so plainly.

Live research pass

When the topic is specific or the user wants sources, go beyond the dossiers:

• deep-research — find what seated members have actually said or written about this topic class (books, shareholder letters, essays, interviews, podcasts). The brief archives per deep-research's own convention.
• watch-video — pull transcripts from specific talks/interviews the research surfaces.
• /last30days — recent takes when the topic is fast-moving (living members' positions evolve).

Save findings per member to research.md in the session folder and point each member at its section, so the cards can cite them ("In the 1997 shareholder letter, Bezos wrote…"). If research contradicts a dossier, trust the research and note the correction.

Grounding rules (non-negotiable)

• Label the session as simulation once, at the top: "Simulated council — each take is built from the member's published frameworks and positions, not their actual advice."
• No fabricated quotes. Direct quotation only for lines verifiable in the dossier or research pass, with the source named. Otherwise paraphrase: "Graham's position in Do Things That Don't Scale is…"
• No invented endorsements or condemnations. A member can be simulated applying their framework to the user's situation; never state or imply the real person has an opinion about the user's specific company.
• Living members get extra care. All nine are alive; positions evolve — prefer the research pass for anything time-sensitive, and never simulate them commenting on named competitors, people, or controversies.
• Disagree in substance, not caricature. Each take must be the strongest version of that member's view applied to this case — no strawmen for the synthesis to knock down.
• If the dossier and the question don't overlap (e.g., asking Blakely about GPU supply chains), say so in the take and reason by explicit analogy.

Output format

> Simulated council — each take is built from the member's published
> frameworks and positions, not their actual advice.

## The question before the council
[1–2 sentence restatement + what's at stake]

## Seated: [A], [B], [C] ([mode], [rounds run])
[One line on why this bench, including who's the dissenter]

## The call: [option] — [do it / do it with conditions / reshape / don't do it]
Scores: A 6.8 · B 7.4 · C 3.1 — spread on B: 3. [Any cap applied, and why.]
Your lean was [lean]; the council [backed it / changed it / split] because [one line].
Hinges on: [any `(assumed)` item the call depends on — "if this is wrong, re-run"]

---

### [Member A] — [their lens, 3–5 words]
[Final position from their last card: the call, the 2–3 arguments that
carried it, what they conceded, whether they moved]
**Bottom line:** [their ONE LINE FOR THE FOUNDER, from their last card]

…

---

## Where the council disagrees
1. **[Crux]** — [A] says X because [framework]; [B] says Y because
   [framework]. The real trade-off: [tension]. Resolved in debate? [how, or
   still open]. What would settle it: [evidence/test].

## Chair's synthesis
[Recommendation fitted to this user's stage and constraints]
- **Conditions:** [the checkable ones from the draft]
- **Cheapest test first:** [what, and the kill criteria]
- **Dissent:** [who, and their strongest argument, unsoftened. If nobody
  dissents, say so and give the strongest argument that went unanswered]
- **Tripwire:** [which member's warning to monitor, and the signal — taken
  from a card's WHAT WOULD CHANGE MY MIND or conditions, not invented]
- **Next:** [skill handoffs — usually /decide to formalize the call]

Full rounds: [path to the session folder]

Members' sections summarize their cards; never rewrite a member's argument into something they didn't say.

In chat, show the header through "Hinges on," plus the conditions, cheapest test, and dissent — under 200 words — then the path to verdict.md. If conditions must be cut for length, keep the ones the dissent or the kill criteria refer to. The member sections and disagreement map live in the file; show them in chat only when asked. A user who says "just give me the answer" gets exactly that.

Archive

Sessions archive to ${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/maker-council/archive/ (create if missing). Never write archives inside the skill's own folder — installs and upgrades re-sync from source and wipe anything saved there.

A council session gets a folder, <archive dir>/<YYYY-MM-DD>-<slug>/, filled as the session runs: brief.md, round-1.md, round-2.md, draft.md, round-3.md (if it ran), and verdict.md (the output format above, plus a header noting s