build-coherence
À propos
La compétence `build-coherence` aide les développeurs à choisir entre plusieurs solutions valides en appliquant un processus de raisonnement structuré et multi-voies, inspiré de l'intelligence en essaim. Elle évalue indépendamment les approches concurrentes, utilise un raisonnement explicite ("waggle dance") pour justifier les choix, et emploie des seuils de confiance pour la prise de décision. Utilisez-la lorsque vous devez vous engager sur un choix architectural unique, justifier la sélection d'un outil, ou avant une action irréversible à enjeux élevés.
Installation rapide
Claude Code
Recommandénpx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/build-coherenceCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Build Coherence
Evaluate competing approaches through independent assessment, explicit reasoning-out-loud advocacy, confidence-calibrated commitment thresholds, and structured deadlock resolution — producing coherent decisions from multiple reasoning paths.
When to Use
forage-solutionshas identified multiple valid approaches and a selection must be made- Oscillating between two approaches without committing to either
- Needing to justify a decision with structured reasoning (architecture choice, tool selection, implementation strategy)
- When a previous decision was made by gut feeling and needs evidence-based validation
- When internal reasoning is producing contradictory conclusions and coherence must be restored
- Before an irreversible action (merging, deploying, deleting) where the cost of the wrong choice is high
Inputs
- Required: Two or more competing approaches to evaluate
- Optional: Quality assessments from prior scouting (see
forage-solutions) - Optional: Decision stakes (reversible, moderate, irreversible) for threshold calibration
- Optional: Time budget for the decision
- Optional: Known failure mode (oscillation, premature commitment, groupthink)
Procedure
Step 1: Independent Evaluation
Assess each approach on its own merits before comparing them. The critical rule: do not let the assessment of approach A bias the assessment of approach B.
For each approach, evaluate independently:
Approach Evaluation Template:
┌────────────────────────┬──────────────────────────────────────────┐
│ Dimension │ Assessment │
├────────────────────────┼──────────────────────────────────────────┤
│ Approach name │ │
├────────────────────────┼──────────────────────────────────────────┤
│ Core mechanism │ How does this approach solve the problem? │
├────────────────────────┼──────────────────────────────────────────┤
│ Strengths (2-3) │ What does this approach do well? │
├────────────────────────┼──────────────────────────────────────────┤
│ Risks (2-3) │ What could go wrong? What is assumed? │
├────────────────────────┼──────────────────────────────────────────┤
│ Evidence quality │ How well-supported is this approach? │
│ │ (verified / inferred / speculated) │
├────────────────────────┼──────────────────────────────────────────┤
│ Quality score (0-100) │ Overall assessment │
├────────────────────────┼──────────────────────────────────────────┤
│ Confidence (0-100) │ How confident in this assessment? │
└────────────────────────┴──────────────────────────────────────────┘
Fill this out for each approach separately. Do not write a comparison until all individual evaluations are complete.
Got: Independent evaluations where each approach is assessed on its own terms. The evaluation of approach B does not reference approach A. Quality scores reflect genuine assessment, not ranking.
If fail: If the evaluations are contaminated (you find yourself writing "better than A" while assessing B), reset. Assess A completely, then clear the framing and assess B from scratch. If the scores are all identical, the evaluation dimensions are too coarse — add domain-specific criteria.
Step 2: Waggle Dance — Reason Out Loud
Advocate for each approach proportionally to its quality. This is the AI equivalent of the bee waggle dance: making implicit reasoning explicit and public.
- For each approach, state the case for it — as if presenting to a skeptical user:
- "Approach A is strong because [evidence]. The main risk is [risk], which is mitigated by [mitigation]."
- Advocacy intensity should be proportional to quality score:
- High-quality approach: detailed advocacy with specific evidence
- Medium-quality approach: brief advocacy with acknowledged limitations
- Low-quality approach: mentioned for completeness, not actively advocated
- Cross-inspection: after advocating for A, actively look for evidence that supports B instead. After advocating for B, look for evidence that supports A. This counteracts confirmation bias
The purpose of reasoning-out-loud is to make the decision auditable — to yourself and to the user. If the reasoning cannot be articulated, the assessment is shallower than the score suggests.
Got: Explicit reasoning for each approach that would be persuasive to a neutral observer. Cross-inspection reveals at least one consideration that was initially overlooked.
If fail: If advocacy feels perfunctory (going through motions), the approaches may not be genuinely different — they may be variations of the same idea. Check: do the approaches differ in mechanism, or only in implementation detail? If the latter, the decision may not matter much — pick either and move on.
Step 3: Set Quorum Threshold and Commit
Set the confidence threshold required to commit, calibrated to the decision's stakes.
Confidence Thresholds by Stakes:
┌─────────────────────┬───────────┬──────────────────────────────────┐
│ Decision Type │ Threshold │ Rationale │
├─────────────────────┼───────────┼──────────────────────────────────┤
│ Easily reversible │ 60% │ Cost of trying and reverting is │
│ (can undo) │ │ low. Speed matters more than │
│ │ │ certainty │
├─────────────────────┼───────────┼──────────────────────────────────┤
│ Moderate stakes │ 75% │ Reverting has cost but is │
│ (costly to reverse) │ │ possible. Worth investing in │
│ │ │ evaluation │
├─────────────────────┼───────────┼──────────────────────────────────┤
│ Irreversible or │ 90% │ Cannot undo. Must be confident. │
│ high-stakes │ │ If threshold not met, gather │
│ │ │ more information before deciding │
└─────────────────────┴───────────┴──────────────────────────────────┘
- Classify the decision stakes
- Check: does the leading approach's quality score × confidence reach the threshold?
- If yes: commit. State the decision, the reasoning, and the key risk being accepted
- If no: identify what additional information would raise confidence to the threshold
- Once committed, do not revisit unless new disqualifying evidence emerges
Got: A clear commitment moment with stated reasoning. The decision is made at an appropriate confidence level for its stakes.
If fail: If the threshold is never met (can't reach 90% on an irreversible decision), ask: is the decision truly irreversible? Can it be decomposed into a reversible test phase + an irreversible commit? Most apparently irreversible decisions can be staged. If staging is impossible, communicate the uncertainty to the user and ask for guidance.
Step 4: Resolve Deadlocks
When two or more approaches have similar scores and the quorum threshold is not met for any single one.
Deadlock Resolution:
┌────────────────────────┬──────────────────────────────────────────┐
│ Deadlock Type │ Resolution │
├────────────────────────┼──────────────────────────────────────────┤
│ Genuine tie │ The approaches are equivalent. Pick one │
│ (scores within 5%) │ and commit. The cost of deliberating │
│ │ exceeds the cost of picking the "wrong" │
│ │ equivalent option. Flip a coin mentally │
├────────────────────────┼──────────────────────────────────────────┤
│ Information deficit │ The tie exists because evaluation is │
│ (scores uncertain) │ incomplete. Invest one more specific │
│ │ investigation — a targeted file read, a │
│ │ quick test — then re-score │
├────────────────────────┼──────────────────────────────────────────┤
│ Oscillation │ Scoring keeps flip-flopping depending on │
│ (scores keep changing) │ which dimension gets attention. Time-box:│
│ │ set a timer, evaluate once more, commit │
│ │ to the result regardless │
├────────────────────────┼──────────────────────────────────────────┤
│ Approach merge │ The best parts of A and B can be │
│ (compatible strengths) │ combined. Check for compatibility. If │
│ │ merge is coherent, use it. If forced, │
│ │ don't — pick one │
└────────────────────────┴──────────────────────────────────────────┘
Got: Deadlock resolved through the appropriate mechanism. The resolution is decisive — no lingering doubt that undermines execution.
If fail: If the deadlock persists through all resolution strategies, the decision may be premature. Ask the user: "I see two equally strong approaches: [A] and [B]. [Brief case for each.] Which aligns better with your priorities?" Delegating a genuine tie to the user is not a failure — it is acknowledging that the decision depends on values the AI cannot infer.
Step 5: Assess Coherence Quality
After committing to a decision, evaluate whether the process produced genuine coherence or just a decision.
- Was the decision evidence-based, or was it rubber-stamping an initial preference?
- Test: was the preference the same before and after evaluation? If so, did the evaluation change anything?
- Were the losing approaches genuinely considered, or were they straw men?
- Test: can you articulate the strongest case for the losing approach?
- What signal would trigger reassessment?
- Define a specific observation that would invalidate the decision ("If I discover that the API doesn't support X, then approach B becomes better")
- Is there useful information from the losing approaches that should inform implementation?
- A risk identified in approach B might apply to approach A as well
Got: A brief quality check that either confirms the decision or identifies it as weak. If weak, return to the appropriate earlier step rather than proceeding on shaky ground.
If fail: If the quality check reveals that the decision was preference-based rather than evidence-based, acknowledge it honestly. Sometimes preference is all that is available — but it should be labeled as such, not dressed up as analysis.
Validation
- Each approach was evaluated independently before comparison
- Advocacy was proportional to quality (not equal attention regardless of merit)
- Cross-inspection was performed (looking for counter-evidence after advocacy)
- Quorum threshold was calibrated to decision stakes
- If deadlocked, a specific resolution strategy was applied
- Post-decision quality check was performed
- A reassessment trigger was defined
Pitfalls
- Premature commitment: Deciding before evaluating all approaches. The first approach considered has an anchoring advantage — it gets more mental attention by being first. Evaluate all before comparing
- Equal advocacy for unequal approaches: If approach A scored 85 and approach B scored 45, spending equal time advocating for both wastes effort and creates false equivalence
- Rubber-stamping: Going through the evaluation process to justify a decision already made. The test is whether the evaluation could have changed the outcome. If not, the process was theater
- Threshold avoidance: Lowering the confidence threshold to make the decision easier rather than gathering the information needed to meet the appropriate threshold
- Ignoring the losing side: The losing approach often contains warnings that apply to the winning one. Risks identified in approach B don't disappear because approach A was chosen
Related Skills
build-consensus— the multi-agent consensus model that this skill adapts to single-agent reasoningforage-solutions— scouts the solution space that coherence evaluates; typically precedes this skillcoordinate-reasoning— manages information flow during multi-path evaluationcenter— establishes the balanced baseline needed for unbiased evaluationmeditate— clears assumptions between evaluating different approaches
Dépôt GitHub
Compétences associées
content-collections
MétaCette compétence propose une configuration éprouvée en production pour Content Collections, un outil axé sur TypeScript qui transforme des fichiers Markdown/MDX en collections de données typées de manière sûre avec une validation Zod. Utilisez-la lors de la création de blogs, de sites de documentation ou d'applications Vite + React riches en contenu pour garantir la sécurité de typage et la validation automatique du contenu. Elle couvre tout, de la configuration du plugin Vite et de la compilation MDX à l'optimisation des déploiements et la validation des schémas.
polymarket
MétaCette compétence permet aux développeurs de créer des applications avec la plateforme de marchés prédictifs Polymarket, incluant l'intégration d'API pour le trading et les données de marché. Elle fournit également une diffusion de données en temps réel via WebSocket pour surveiller les transactions en direct et l'activité du marché. Utilisez-la pour mettre en œuvre des stratégies de trading ou pour créer des outils traitant les mises à jour de marché en direct.
creating-opencode-plugins
MétaCette compétence aide les développeurs à créer des plugins OpenCode qui s'interconnectent avec plus de 25 types d'événements tels que les commandes, les fichiers et les opérations LSP. Elle fournit la structure du plugin, les spécifications de l'API événementielle et les modèles d'implémentation pour les modules JavaScript/TypeScript. Utilisez-la lorsque vous avez besoin d'intercepter, de surveiller ou d'étendre le cycle de vie de l'assistant IA OpenCode avec une logique personnalisée pilotée par les événements.
sglang
MétaSGLang est un framework de service LLM haute performance spécialisé dans la génération rapide et structurée pour les workflows JSON, regex et agentiques grâce à son cache de préfixe RadixAttention. Il offre une inférence nettement plus rapide, particulièrement pour les tâches avec des préfixes répétés, ce qui le rend idéal pour les sorties complexes et structurées ainsi que les conversations multi-tours. Choisissez SGLang plutôt que des alternatives comme vLLM lorsque vous avez besoin d'un décodage contraint ou que vous construisez des applications avec un partage étendu de préfixes.
