Valet

The problem

AI assistants assume the user has reliable executive function. They assume you’ll remember to use the tool, decide what to delegate, hold context across sessions, and catch errors before anything ships. That’s a heavy cognitive contract. For operators whose attentional systems are the constraint (whether from neurodivergence, sustained burnout, or the structural chaos of solo product work), generic AI tools don’t close the gap. They widen it.

The second problem is trust. AI drafts read smoothly. Smooth is dangerous. I shipped a document with a wrong number because the prose was confident enough that I read for flow instead of fact. A colleague caught it. The failure wasn’t a bad draft. It was a system that hadn’t thought carefully enough about when to make me slow down.

Both problems point at the same thing: the orchestration layer. Most AI tools ignore it entirely.

Who it’s for

High-stakes operators managing multiple contexts whose internal scaffolding is unreliable. Solo PMs. Founders running product and people simultaneously. Knowledge workers who are very good at their jobs and intermittently bad at the mechanics of sustaining them.

The design constraint was specific: build for someone who cannot always trust their own memory, prioritization instincts, or attention span to catch what the AI got wrong. The resulting framework is more rigorous than most operators need. That rigor is the product.

Built first for me. Designed to generalize.

How it helps

Valet inverts the relationship between person and tool. Most AI tools wait for the person to hold everything in their head and then ask the right question. Valet holds the structure. The planning hierarchy, the memory, the routing logic, the escalation rules live in the system. The human’s job is to make decisions and author the narrative. The system’s job is to gather, structure, draft, and remind.

Five principles ground the design. Default to action: if the user doesn’t decide, the system decides based on known preferences. Context-first organization: nothing lives in a general “tasks” bucket. Progressive autonomy: the system earns more independence as trust is established. External scaffolding over internal willpower: systems beat motivation for anything that has to happen reliably. And “human stays sharp.” The person does the thinking, Valet does the assembly, and the design actively resists reversing that.

The last principle is the one that matters most. If the AI starts authoring the narrative instead of the human, the human becomes brittle and dependent. Valet fights that drift deliberately. Every workflow is designed to surface my judgment, not substitute for it.

What it does

Valet runs as a persistent layer over Claude Code. Not a product you open. A system that’s already running when you start working. I ship with it daily. It’s how I run sprint planning, vendor evaluations, stakeholder communications, and product strategy. Twenty-five production skills, each a named protocol that fires on contextual signals rather than a slash command you have to remember.

Planning. Multi-horizon hierarchy from decade goals through daily execution. Annual planning protocol that runs as a 5-phase structured process. Every output connects back to a quarterly goal. The link is explicit, not assumed.

Memory. Three-layer architecture with different time horizons. Append-only durable logs for completed work and decisions that never get overwritten. Ephemeral context snapshot rebuilt fresh each morning. Canonical task source in Jira for anything that involves engineering coordination. The layers don’t compete. They serve different purposes.

25 production skills. Morning standup, evening review, deep work protocol, decision journal, meeting prep, meeting followup, sprint planning, sprint review, vendor evaluation, communication coaching, slide generation, Confluence page builder, inbox processing. Each fires on context, not command.

Methodology contributions. These are the artifacts worth understanding independently.

Strategic Intent Taxonomy. Nine named communication intents, defined before the words start: building capital, creating accountability without assigning it, seeding an idea, protecting a timeline, lowering the temperature. Most communication coaching operates at grammar and tone. The leverage is one level up: what are you actually trying to accomplish with this message?

Trust Tier System. Green (logistics, just runs), Yellow (briefing card, 30-second scan), Red (briefing card plus forced retrieval). Every Yellow and Red deliverable gets a companion briefing card formatted for speaking out loud, not reading on a screen. The tier system came directly from the failure above. The fix wasn’t better AI output. It was forced retrieval before anything ships.

Research Integrity Protocol. Every factual claim gets tagged: [VERIFIED], [SOURCED], [INFERRED], or [UNVERIFIED]. Two unverified claims in a single output trigger automatic web search before the work continues. The tagging makes every document defensible without a second pass.

5-level escalation system. Nudge, Friction, Forced Choice, Consequence, Auto-Execute. Built on one operating principle: inaction is not the same as nothing happening. Designed for the times when willpower is the unreliable variable.

What I learned

Most “AI productivity” systems are wrappers around chat. The interesting work is in the orchestration layer: when does the system push back, when does it draft, when does it wait?

The Trust Tier system came from a real failure. The wrong number. The smooth draft. The colleague who caught what I missed. The fix was structural, not editorial: before anything ships, I should be able to answer a specific question about it cold. If I can’t, the tool has failed regardless of how good the prose is.

The hardest design decision wasn’t a feature. It was a constraint. The system is not allowed to conclude without me. Valet can gather, draft, remind, and escalate. It cannot decide what matters. That line sounds obvious. Holding it under time pressure is harder than it sounds. The whole system is built to hold that line even when I’m the one trying to cross it.

Building this also taught me what agentic AI fluency actually means in practice. It’s not prompt engineering. It’s system design: what gets automated, what stays human, where the handoff lives, and how you catch drift before it compounds. Those questions have answers. Most AI tools don’t ask them.

Status

Actively maintained. Running my work daily. Currently private; the memory layer holds enough personal material that a clean public release needs real work. The methodology contributions are the public artifacts for now. Framework essays are in progress.