agentleFS
qengine13qengine13/forge-skills/skills/crucible-forge/SKILL.md

crucible-forge

Skill1 starsChanged 8 days agoApache-2.0View on GitHub

Portable hard-decision tribunal for architecture, strategy, trading-risk, migration, pricing, kill-or-keep, and pre-commit gate decisions. Use when a decision is costly, contested, evidence-sensitive, identity-sensitive, or the user asks to stress-test, run a tribunal, forge a verdict, decide between options, or prove a plan clears a 10x bar.

The file

Show its rules and text

We split this skill into the rules it gives an agent

23 rules. About 0 are common, 16 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈16

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill to turn a consequential decision into an auditable verdict.

    line 12
  2. Distinctive:

    Read model and tool selection before selecting models or delegating.

    line 17
  3. Distinctive:

    Use the current host's authorized capabilities.

    line 18
  4. Distinctive:

    Do not use for trivial reversible choices, simple lookups, or already-decided implementation unless the implementation carries high …

    line 37
  5. Distinctive:

    Check late-arrival, backfill, revision, corporate-action, survivorship, dedupe, cluster, and train/test contamination paths.

    line 73
  6. Very specific:

    If any critical feature fails this check, verdict is FAIL or INSUFFICIENT_EVIDENCE.

    line 76
  7. Distinctive:

    Use when subagents are not explicitly authorized or the decision can be resolved locally.

    line 82
  8. Distinctive:

    Use only when the user explicitly asks for agents, panelists, delegation, or parallel tribunal work.

    line 92
  9. Distinctive:

    Do not pass other panelists' drafts into round 1.

    line 95
  10. Distinctive:

    Run cross-examination, inversion for high-stakes splits, then final sealed vote.

    line 97
  11. Distinctive:

    If the user provides a schema, obey the user's schema exactly.

    line 118
  12. Distinctive:

    No unverifiable expected-impact claim.

    line 122
  13. Distinctive:

    No future-data feature without point-in-time proof.

    line 123
  14. Distinctive:

    No strategic recommendation that conflicts with the identity or strategy file unless the conflict is explicit.

    line 124
  15. Distinctive:

    No adoption verdict without kill conditions.

    line 125
  16. Very specific:

    No consensus theater: if the split is real, return INSUFFICIENT_EVIDENCE.

    line 126
  17. Distinctive:

    No implementation hidden inside a verdict unless the user explicitly asked to execute.

    line 127
  18. Very specific:

    Read references/panelists.md when selecting tribunal roles.

    line 131
  19. Very specific:

    Read references/checklists.md for decision-specific gates.

    line 132
  20. Very specific:

    Use templates/verdict.schema.json when the user does not provide a schema.

    line 133
  21. Very specific:

    Use scripts/aggregate_verdicts.py to combine saved panel JSON files offline.

    line 134
  22. Distinctive:

    Treat confidence as an estimate, not a calibrated probability.

    line 135
  23. Very specific:

    Run scripts/validate_verdict.py <json-file> before claiming a stored verdict is valid.

    line 136

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<!-- Public edition: adapted for portable installation and subscription-aware model selection. See collection NOTICE. -->

Crucible Forge

Use this skill to turn a consequential decision into an auditable verdict. Map claims to sources, preserve dissent, and refuse to approve when material evidence is missing.

Runtime compatibility

Read model and tool selection before selecting models or delegating. Use the current host's authorized capabilities. No provider dispatcher or paid API is bundled.

What it adds

  • A decision-time evidence ledger: source, date, uncertainty, and consequence if wrong.
  • Point-in-time checks for data decisions: could the input have been known when the decision was made?
  • Independent first passes when separate agents are authorized and available.
  • The strongest opposing argument before rejection.
  • Explicit reversal conditions and preserved dissent.
  • Offline JSON validation and conservative aggregation of independently obtained panel responses.

When To Use

  • Architecture or stack tradeoffs.
  • Strategy, positioning, pricing, offer, hiring, firing, or kill-or-keep decisions.
  • ML/data/trading plans where lookahead, leakage, calibration, or validation discipline matters.
  • High-blast-radius code gates, schema migrations, live trading thresholds, launch gates, or money-moving paths.
  • Any user wording like "should we", "worth it", "stress test", "run tribunal", "10x", "go/no-go", or "forge verdict".

Do not use for trivial reversible choices, simple lookups, or already-decided implementation unless the implementation carries high irreversible risk.

Mandatory Intake

  1. Restate the decision in one sentence.
  2. Classify stakes:
    • reversible: cheap to undo, low blast radius.
    • expensive: meaningful time, money, customer, architecture, data, or reputation cost.
    • existential: live capital, legal/compliance, core strategy, irreversible migration, or company-level bet.
  3. Load the project's identity or strategy file (for example identity.md) before strategy-level guidance if present. If missing and identity is required, ask the minimum question needed.
  4. Load project instructions (AGENTS.md, CLAUDE.md, README.md) only as needed.
  5. Define the baseline and the 10x claim. If it is correctness/blocker work, say so instead of pretending it is 10x strategy.

Evidence Ladder

Score evidence by the strongest available tier:

  • A: live source, primary docs, reproducible command, dataset row, file line, contract, or current system output.
  • B: internal docs, recent project history, validated benchmark, or directly relevant prior incident.
  • C: plausible inference from adjacent evidence.
  • D: training-data prior or unsupported expert intuition.

Rules:

  • A verdict depending on D evidence is capped at confidence 0.55.
  • A verdict depending on C evidence is capped at confidence 0.70.
  • Missing evidence required to prevent major downside yields INSUFFICIENT_EVIDENCE.
  • Web or live-doc verification is mandatory for current prices, laws, APIs, model capabilities, product limits, external benchmarks, and time-sensitive research claims.

Leakage Gate

For trading, ML, analytics, replay, corpus, or data-product decisions, run this check before recommending adoption:

  1. Name every inference-time feature, label, threshold, or ranking.
  2. For each one, identify T_decision.
  3. Prove the value was knowable using only information available before T_decision.
  4. Check late-arrival, backfill, revision, corporate-action, survivorship, dedupe, cluster, and train/test contamination paths.
  5. Demand tests for point-in-time invariance, embargo/purge, walk-forward or CPCV, and training/inference parity.

If any critical feature fails this check, verdict is FAIL or INSUFFICIENT_EVIDENCE.

Tribunal Modes

Compact Mode

Use when subagents are not explicitly authorized or the decision can be resolved locally.

  1. Scout: gather direct evidence.
  2. Red Team: strongest failure case.
  3. Steel Man: strongest adoption case.
  4. Constraint Keeper: fit against the identity or strategy file, project rules, blast radius, and constraints.
  5. Judge: issue verdict with confidence, evidence, objections, kill conditions, and next steps.

Sealed Multi-Agent Mode

Use only when the user explicitly asks for agents, panelists, delegation, or parallel tribunal work.

  1. Dispatch independent subagents with disjoint roles.
  2. Do not pass other panelists' drafts into round 1.
  3. Reveal after all round-1 outputs are complete.
  4. Run cross-examination, inversion for high-stakes splits, then final sealed vote.
  5. Synthesize with dissent preserved.

Output Shapes

Default verdict:

{
  "panelist": "crucible_forge",
  "verdict": "PASS|FAIL|INSUFFICIENT_EVIDENCE",
  "decision": "accept|reject|accept_with_modification|defer",
  "confidence": 0.0,
  "evidence_provided": [],
  "evidence_demanded": [],
  "outstanding_objections": [],
  "kill_conditions": [],
  "next_3_steps": []
}

If the user provides a schema, obey the user's schema exactly.

Fail-Closed Rules

  • No unverifiable expected-impact claim.
  • No future-data feature without point-in-time proof.
  • No strategic recommendation that conflicts with the identity or strategy file unless the conflict is explicit.
  • No adoption verdict without kill conditions.
  • No consensus theater: if the split is real, return INSUFFICIENT_EVIDENCE.
  • No implementation hidden inside a verdict unless the user explicitly asked to execute.

References

  • Read references/panelists.md when selecting tribunal roles.
  • Read references/checklists.md for decision-specific gates.
  • Use templates/verdict.schema.json when the user does not provide a schema.
  • Use scripts/aggregate_verdicts.py to combine saved panel JSON files offline. This does not call models.
  • Treat confidence as an estimate, not a calibrated probability; never invent a Brier score without outcomes.
  • Run scripts/validate_verdict.py <json-file> before claiming a stored verdict is valid.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.