AI knowledge map
Where this was learned, applied, and proven
The graph is a deliberate index, not a force-directed blob. The list is the accessible source of truth.
Prompt
A single instruction that produces a response.
Configured assistant
Persistent role, context, rules, and domain constraints. Primarily conversational.
LLM workflow
A predefined sequence of LLM and non-LLM steps. Reliable, but the path is largely fixed.
Agentic workflow
Inspects state, selects approved actions, manages retries, and adapts its path to complete a goal.
Autonomous agent
Broader authority over time. Requires significantly stronger governance. None of these projects claim this class.
AI concepts × projects
LLMs as bounded components
Prompt and role configuration
Context engineering
LLM workflows
Agentic workflows
Tool use and allowlists
Human-in-the-loop
Trusted retrieval / source timing
Structured outputs
Validation and evaluation
Model limitations
Authority boundaries
Observability
AI-assisted development
LLMs as bounded components
Models draft, classify, and summarize. They do not own authority, publication, or remediation.
Where learned: First as specialized GPTs on a service-desk publishing workflow, then as Gemini inside TinyClaw and a domain Gem.
Evidence: Every project labels where the model sits on the maturity spectrum rather than calling it an agent by default.
Applied in: Multi-Stage Knowledge Publishing Workflow · TinyClaw · Domain-Specific AI Expert
Prompt and role configuration
Persistent role, rules, and output contracts — not a one-shot instruction.
Where learned: Applied as a specialized Gemini Gem for a specific vehicle-maintenance domain.
Evidence: Capability snapshot: Prompt & role configuration — Applied.
Applied in: Domain-Specific AI Expert · Multi-Stage Knowledge Publishing Workflow
Context engineering
Separate owner context, configuration, and reference instead of blending them into one prompt.
Where learned: The Jeep assistant made this the whole product.
Evidence: Lesson: usefulness was almost entirely context and constraints, not model choice.
Applied in: Domain-Specific AI Expert
LLM workflows
A predefined sequence of LLM and non-LLM steps. Reliable, path largely fixed.
Where learned: Chelan County PUD Self Help pipeline; TinyClaw scheduled reporting.
Evidence: Operator-driven stages; scheduled collect → normalize → analyze → deliver.
Applied in: Multi-Stage Knowledge Publishing Workflow · TinyClaw
Agentic workflows
Inspect state, select approved actions, manage retries, adapt the path. Still not autonomous.
Where learned: Specified in the publishing agent; built in the home-lab triage agent.
Evidence: Triage agent plans collection from the case rather than running a fixed script.
Applied in: Home-Lab Endpoint Triage Agent · Knowledge Publishing Agent
Tool use and allowlists
Tools are capabilities. If it is not on the allowlist, completing the task “the easy way” is a failed control.
Where learned: Triage red-team included privilege-escalation prompts that would have been easier with extra tools.
Evidence: Adversarial study: did the agent stay inside its granted authority?
Applied in: Home-Lab Endpoint Triage Agent · Knowledge Publishing Agent
Human-in-the-loop
Consequential actions require a person. The loop is a state machine, not a polite request in a prompt.
Where learned: Publishing, packaging, containment, and deploy all use explicit gates.
Evidence: Clearward human approval before packaging; triage agent never remediates.
Applied in: Home-Lab Endpoint Triage Agent · Clearward · Basaltborne · Knowledge Publishing Agent
Trusted retrieval / source timing
Normalize and timestamp sources before the model sees them. Forum consensus is not a specification.
Where learned: TinyClaw freshness path; Jeep source-verification instruction.
Evidence: Retry → fallback → flag stale → skip, in that order.
Applied in: TinyClaw · Domain-Specific AI Expert · Basaltborne
Structured outputs
Contracts the next stage can consume: HTML stages, morning reports, triage JSON, workbook tables.
Where learned: If formatting can drift, downstream consumption silently breaks.
Evidence: Triage output contract: JSON, executive summary, timeline, cited risk score.
Applied in: Multi-Stage Knowledge Publishing Workflow · TinyClaw · Home-Lab Endpoint Triage Agent · Clearward
Validation and evaluation
Happy path is not enough. Missing, malformed, stale, contradictory, injected, and timed-out inputs are the real suite.
Where learned: Triage functional plus adversarial layers; Clearward 49 automated checks; Basaltborne CI layers.
Evidence: Documented red-team findings; 97/97 Vitest, 120/120 pgTAP, 28/28 Playwright at Phase 3.
Applied in: Home-Lab Endpoint Triage Agent · Clearward · Basaltborne
Model limitations
Models will guess, blur evidence and inference, and treat injected text as instructions unless the design forbids it.
Where learned: Most clearly when attacking the triage agent.
Evidence: Prompt-only controls were the ones that needed hardening; code-enforced controls held.
Applied in: Home-Lab Endpoint Triage Agent · TinyClaw
Observability
Tool calls, retries, confidence, approvals, and stop reasons have to be reconstructable later.
Where learned: Adversarial testing asked whether each attempt was visible in the audit log.
Evidence: If the attempt is not in the log, the control did not hold.
Applied in: Home-Lab Endpoint Triage Agent · Knowledge Publishing Agent
AI-assisted development
Agents draft under persistent instructions and phase gates. They do not merge, package, or deploy themselves.
Where learned: Clearward AGENTS.md; Basaltborne phase-gated lifecycle with PR review of AI-authored changes.
Evidence: Never claim success without reopening and verifying the resulting state.
Applied in: Clearward · Basaltborne