ai-config-kit-core

name

ai-config-kit-core

description

Route into the albrand/ai-config-kit operating framework and load only the relevant doctrine for planning, implementation, debugging, review, security, harness design, quality convergence, repository adoption, journaling, skill learning, token economy, or context acceleration. Use when the user asks to apply, inspect, adopt, or update ai-config-kit; asks for a rigorous agent operating model or harness; or when an installed ai-config-kit specialist skill requires shared framework context.

AI Config Kit Core

Use the bundled framework as an on-demand operating library. Keep current system, developer, user, and repository instructions authoritative. Treat framework rules as guidance unless the repository explicitly adopts them.

Source And Compatibility Gate

1. Read repository instructions and current task constraints first.
2. Record the available capabilities that materially affect the task: local reads/writes, validation commands, network, external systems, sub-agents, model routing, and context accelerators.
3. Do not claim unavailable harness capabilities. Do not launch sub-agents, mutate external systems, run active security tests, or install dependencies unless current instructions and user authorization permit it.
4. Challenge framework directives for current fit. In particular, do not turn optional board access, journaling, cross-agent critique, sibling-repository scans, or context accelerators into universal blockers.
5. Prefer current source files, tests, runtime evidence, accepted criteria, and authoritative external state over framework examples or prior conclusions.

Load The Narrowest Profile

Load only the references required for the active task:

• Orientation or ordinary work: references/AI_BOOTSTRAP.md, references/OPERATING_MODEL.md, and relevant repository instructions.
• Implementation or architecture: add references/ARCHITECTURE_AND_CODE_QUALITY.md and references/QUALITY_GATES.md.
• Debugging: add references/QUALITY_GATES.md; require evidence, reproduction when feasible, a root-cause account, a focused fix, and regression protection.
• Review or PR work: use the installed $high-signal-pr-review skill. Load references/REVIEW_AND_PR_FRAMEWORK.md only for shared doctrine.
• Security: use $adversarial-security-sweep for static review or $pentest-specialist for authorized active testing. Load references/SECURITY_AND_PENTEST.md for shared doctrine.
• Roadmap, technical bootstrap, or whole-project hardening: use the matching $roadmap-terraform, $tech-terraform, or $assess-then-harden skill.
• Module planning: use $plan-module-delivery.
• UX work: use $ux-design-agent.
• Harness, delegation, routing, or cost design: load references/HARNESS_STRATEGY.md, references/AGENT_ORCHESTRATION.md, and references/TOKEN_ECONOMY.md. Add references/ADAPTIVE_MODEL_ORCHESTRATION.md when multiple models, reasoning efforts, or an external sidecar are in scope. Add references/CROSS_AGENT_COORDINATION.md only for actual multi-tool work.
• Quality iteration: load references/QUALITY_CONVERGENCE.md and references/QUALITY_GATES.md.
• Repository adoption: load references/REPO_ADOPTION_PLAYBOOK.md, references/REPO_AGENTS_TEMPLATE.md, and references/FRAMEWORK_MANIFEST.md.
• Long-lived task journaling: load references/SESSION_JOURNALING.md only when the task benefits from durable resume state.
• Skill promotion or framework learning: load references/CONTINUOUS_SKILL_LEARNING.md and references/SKILLS_CATALOG.md.
• Optional graph, wiki, symbol-index, or review-graph evaluation: use $context-acceleration and load references/CONTEXT_ACCELERATION.md.
• Full framework audit or upstream reconciliation: start with references/FRAMEWORK_MANIFEST.md and references/README.md, then inspect only the profiles and files named there.

Core Execution Loop

1. Classify the work and identify the expected outcome, scope, non-goals, affected surfaces, assumptions, and approval boundaries.
2. Discover from indexes and metadata before opening broad source. Trace enough callers, callees, tests, configs, docs, schemas, and runtime boundaries to understand impact.
3. For non-trivial changes, state a concise plan before writes. Re-plan when evidence changes coupling or risk.
4. Execute the smallest coherent change that satisfies the current source of truth. Preserve unrelated user changes.
5. Apply proportional validation. Distinguish passed, failed, blocked, skipped, and not run. Never imply an unrun check passed.
6. For auth, authorization, secrets, crypto, external input, file handling, outbound requests, dependencies, or build/config changes, add a focused security pass and supply-chain inspection.
7. Close with what changed, what was validated, what was not validated, and residual risk or the next action.

Framework Practices Worth Preserving

• Evidence before claims; source-of-truth precedence over memory.
• Progressive disclosure and deterministic filtering before broad context.
• Business-rule and regression thinking in implementation and review.
• High-signal findings tied to concrete behavior and practical impact.
• Approval gates for external mutations and active security testing.
• Independent critique only when available, useful, and authorized.
• Journals as resumable evidence, not unquestionable memory.
• No AI attribution, signatures, or generated-by footers unless requested.

Guardrails

• Do not bulk-load all references for routine work.
• Do not install GLOBAL_AGENTS.md verbatim as a universal policy without an explicit adoption decision; it contains intentionally strict assumptions.
• Do not treat estimated dates as commitments or invent owners, IDs, tickets, business rules, validation results, or external state.
• Do not let the framework override a repository's real architecture, validation commands, domain rules, or higher-priority instructions.

Package Integrity

This skill is self-contained. Load only files under its bundled references/ directory after installation; do not depend on a source checkout or a personal filesystem path. Reinstall the complete package when framework references change, then run the repository's offline Codex-skill validator.