Deliberate
A verdict on a decision
The full six-step procedure: restate, blind analysis, anonymised cross-examination, enforcement, stance and pre-mortem, chairman's verdict.
Say "council this", "pressure-test", "should I X or Y"
A Claude skill for decisions with stakes
Nine minds that disagree on purpose, and a chairman who has to show the vote.
Ask Claude a hard question and you get one confident answer. High Council convenes a panel of distinct reasoning lenses, makes them argue blind, strips their names before cross-examination, holds the verdict to a two-thirds bar, and refuses to close until it has imagined how the decision fails.
Why it exists
"Give me the pros and cons" produces a balanced-looking answer from a single frame. The council is built against the specific ways group judgement goes wrong: anchoring, conformity, a chairman who picks the winner after reading the arguments, and plans that were never imagined failing.
One skill, two engines
DELIBERATE answers "is this the right move?". SCAN answers "which expert would say I missed their entire field?". TRIBUNAL runs both, and the scan is forbidden from repeating anything the council already found.
Deliberate
The full six-step procedure: restate, blind analysis, anonymised cross-examination, enforcement, stance and pre-mortem, chairman's verdict.
Say "council this", "pressure-test", "should I X or Y"
Scan
Sweeps a project across 16 domains of expertise, calls each one honestly, and writes the questions the missing expert would ask.
Say "what am I missing", "blind-spot scan"
Tribunal
The council's findings and pre-mortems become the scan's exclusion list. The overlap is struck out, so the scan can only add what the council couldn't see.
Say "full tribunal", "maximum scrutiny"
Quick
A decisive first take, then masked final positions with stance and pre-mortem. Compact verdict, compact artifacts.
Duo
Two opposed lenses, such as Contrarian and Expansionist, open, rebut and close. A third-party chairman names the irreducible tension.
The deliberation procedure
Each member is a separate subagent in Claude Code, so independence is structural. The diagram follows five members of the default core panel from the framed question to the written verdict.
Reads your workspace for context, rewrites the question without steering, then locks the panel, the 1.5× domain-weight seat and the Chairman before any analysis exists.
Each member restates the problem in 50 words and offers one alternative framing. A material divergence is flagged to you before any rounds are spent.
Up to 300 words each, fully committed to the lens, no hedging. Nobody sees anyone else's output.
Lens names become Member A–E. Each member must engage at least two others, name what everyone missed, and label their evidence. Updating because "consensus is forming" is forbidden.
Dissent quota, novelty gate, a counterfactual probe at >70% agreement, an evidence-monoculture check and anti-recursion. Each fires at most once.
Each member pre-mortems their own option, then casts a machine-readable stance with confidence and a dealbreaker flag. Missing lines are re-prompted, never inferred.
A non-deliberating Chairman reads the de-anonymised transcript and writes the 13-section verdict. Report and transcript are written to disk.
If an early signal fires, the verdict is invalid and its named fallback applies. The transcript feeds the next council, so settled questions aren't argued twice.
Try the vote
The verdict follows the recorded tally. Change any member's stance or confidence, or move the ★ domain-weight seat, and watch whether an option clears. Abstainers stay in the denominator, so a hesitant panel escalates instead of quietly passing.
Example panel and question. The starting position is a deliberate near-miss.
Non-negotiable
Each law closes a specific loophole a coordinator would otherwise be tempted by, and the skill names the rationalisations it refuses, such as "the panel already agrees, skip the enforcement scan".
Round 1 members see only the framed problem. Without subagents, the session is stamped [SINGLE-CONTEXT MODE] and its diversity score is capped.
Members see Member A…N, never lens names. The mapping is restored only in the transcript.
The weight seat and the Chairman are chosen before any analysis exists, so nobody can tilt the outcome after reading positions.
The Chairman may override the tally only by naming the minority argument that defeats the majority. Genuine splits come back to you.
A scan after a council excludes everything the council surfaced, and says so. A scan that repeats the council has failed.
If the question is vague, ask once, then proceed. Never interrogate.
Report and transcript, every mode, every time. "The user seems in a hurry" doesn't waive it.
Members pre-mortem their own option. The Chairman pre-mortems the winner. Kill criteria are derived from those stories.
Law VIII in practice
Kill criteria detect failure, and the pre-mortem is what generates them. Every failure story needs at least one criterion that would catch it early. A criterion that detects no story is a hunch and gets cut.
The pilot ran and three users paid. The founder's employer then asked for the invoices, and six months of revenue became evidence in an IP-ownership dispute.
Each narrative must end with the first thing you could actually see, not an abstract risk.
If IP position is not confirmed in writing by day 14 → pilot invalid → pause invoicing.
[detects: Revenue as evidence]
From the skill's bundled example session for a fictional product, AcmeAlerts. Every verdict also includes one "it succeeded at the wrong thing" narrative.
The scan engine
Each domain gets an honest call with a one-line rationale. Marking everything as APPLIES buries the signal, so a well-argued N/A is part of the deliverable. Findings are phrased as the question the missing expert would ask, followed by the exposure and the cheapest check. Select a domain to see who sits in that chair.
Calls shown are illustrative, for a hypothetical solo-founder SaaS launch. A real scan makes its own calls from your project files.
What you get
The chat answer is a summary. The verdict, transcript and register are what you keep, reread when the kill criteria fire, and feed into the next council as continuity.
Council Verdict
council-report-20261102-1430.html
| Mode | Writes |
|---|---|
| Deliberate | report + transcript |
| Quick · Duo | compact report + transcript |
| Scan | scan doc + register append |
| Tribunal | combined report + transcript + register |
How to use it
The skill routes on plain language. Name a mode or panel if you want one; otherwise it picks, and states its pick and rationale before starting.
"Council this: launch email-first at £99 a month, or wait until the dashboard is self-serve?"
→ DELIBERATE · core panel · Executor holds the ★ seat
"Quick council: Postgres with pgvector or a dedicated vector DB for this workload?"
→ QUICK · architecture triad: First-Principles, Systems, Executor
"What am I missing on this project before I take money from customers?"
→ SCAN · 16-domain sweep, register append
"Have the Contrarian and the Expansionist debate whether I should open-source the core."
→ DUO · polarity pair, third-party chairman
It stays out of the way for yes/no questions, factual lookups, writing tasks, summaries, and casual "should I" with no real trade-off. It never runs a partial council: if the question doesn't merit peer review, it answers directly.
Install
The skill is plain Markdown: one SKILL.md plus five reference files that load only when needed. No dependencies, no API keys.
Claude Code (recommended)
.claude/skills/ inside a project to share it with that repo.git clone https://github.com/TPFLegionaire/highcouncilskill.git cp -r highcouncilskill/high-council ~/.claude/skills/
Claude.ai and the desktop app
high-council folder on its own.