| name | dream |
| description | Parker's dreaming run for the brand and the person using it — the planning half of the living layer. Reads the day's comms (the conversations the user had with Parker first, then any connected comms) against the vault and returns proposals across six buckets — context updates, skill improvements, schedules to create, new ideas to research, new open loops, and the person (what Parker now understands about them and what to tee up for them) — into dreaming/proposals/pending. Captures verbatim; proposes, never applies; promotion runs through self-improve with the human in the loop. Use daily as a scheduled routine, or when asked to "dream", brainstorm overnight, or "what were you thinking about". |
Dream — the generative half of the living layer
Dreaming is how Parker stays living. It is the offline cycle where Parker thinks about everything it has accumulated for the brand and comes back with proposals: ways to improve its own context and skills, recurring work it could take off the user's plate, and fresh ideas worth chasing. The feeling it is built to create: Parker was thinking about the brand overnight — "I was thinking about this last night, here's what I'd suggest."
Dreaming proposes; self-improvement disposes. Nothing here is applied on its own. Every proposal lands in dreaming/proposals/pending/ and is promoted only through the self-improve routine, with the human in the loop.
When it runs
Daily. Brand dreaming runs on the conversation and the accumulated knowledge, so it refreshes as the brain grows and surfaces on a morning rhythm — that daily cadence is what makes the morning suggestion possible.
What it reads (the whole connected picture, against the vault)
The full connected sweep is the subject; the vault is the ground. The conversations the user had with Parker are the richest single thread — pull them through Parker's chat-history tool (search_chat_history), never assuming they live in whatever session is running this pass, since the user may have talked to Parker in the web app — read them closely, but they are the center of a wider sweep, not the whole of it. Each run, pull from every connector the user has linked: work email, Slack, calendar, task manager, Notion, docs, call and meeting transcripts, and the rest, to build the read on the person and the organization. Go out into the world, don't just reread what's here: actively pull the tracked competitor ad libraries and organic niche feeds fresh, run web searches on the category, and hunt what moved since the last run — a competitor's new launch, ad, or pricing move, category press, a new entrant, a trend or cultural moment to ride. Bring back at least something the brain didn't already know; a dream that only repeats what's on file summarized, it didn't dream. Read the whole sweep against the brand's own surfaces — sub-context-docs/, personas/, competitors/, running-notes/, open-loops/, hypotheses/, validations/, audits/, source-pulls/, idea-bank/, and fresh Parker MCP / web data — to know whether something is genuinely new, contradicts what's on file, or closes something already open. Any source not connected this run is skipped and logged as a one-line blind spot, never guessed. Sweep the work, not the private life: the subject is the business, the org, and the market, never the person's private world. Within any source, read only what bears on the work and skip anything that reads as personal or private — personal texts, a family/health/money matter, a clearly-personal calendar entry. Purely personal channels like iMessage aren't swept at all. When in doubt whether it's work or private, leave it out.
Capture verbatim — this is the hard rule. Record what was actually said: the exact words, the exact ask, the exact idea, quoted. Generalizing is the failure mode — a flattened "the user wanted stronger hooks" is something the executing arm can't build from; the real line is. An LLM's instinct is to summarize; resist it. Quote more, not less, and carry that source language intact into every proposal.
What it produces — proposals only
Write each as a file in dreaming/proposals/pending/[YYYY-MM-DD]-[id].md, and log the run's raw observations under dreaming/runs/[YYYY-MM-DD]/. Each proposal carries its reasoning so a rejected dream teaches Parker what doesn't land, exactly as a rejected idea or open loop does.
Six buckets of proposal:
- Context updates. When a conversation or fresh read reveals something the brand context docs don't yet capture or get wrong, propose the edit (point at the doc and the change; do not make it). This bucket also carries new surfaces: when a connected tool keeps surfacing a kind of truth the brain has no home for, or the team keeps asking about a domain with no standing doc, propose growing the brain — the new folder or doc, what feeds it, and what it would be watched for — per
parker-system/system/growing-the-brain.md. The scaffold is a floor, not a ceiling; dreaming is where growth is supposed to start.
- Skill improvements. When the way the user works with Parker exposes a gap or a better way to run a kind of task, propose the skill change. These are brand-observed but often promote upward, since skills are global — flag that.
- Schedules to create. When the user keeps asking for the same kind of job, propose standing it up as a repo-native schedule ("you've asked me this a few times — want me to run it weekly?"). A schedule is not an MCP workflow (see
schedules/README.md). Name the task, cadence, what it reads, what it updates, and the deliverable; the proposal lands in schedules/proposed/ and does not run until confirmed.
- New ideas to research. When an open loop, a validation, a conversation, or a freshly scraped signal could be run with, route the good ones to the idea bank under its own rules, as
[~] proposed (see harvest-ideas).
- New open loops. When the day surfaced something with real pull and an answerable question, write it as a proper open loop — observation, named pull, the one exact question, justification, territory — and feed it into
open-loops/. Capture the loop; self-improve weights and promotes it, not this run.
- The person. Dream about the human, not just the work — this is what makes the brain feel like it knows them. Three things. First, what Parker now understands about the person: a read on who they are, their process, their craft, a preference or a standing rule they revealed today — proposed as an update to
users/[user-id]/user-profile.md, carrying the full verbatim moment that showed it. Their own words in their own profile is the point here, so quote them exactly, not the pattern. Second, the tee-up: the proactive morning suggestion aimed at them, not the brand — what they'll likely need next given what they're working on, what to prep before they ask, a nudge on a deadline or a thread worth picking back up. Keep the profile's "what they're working on" section live while you're here — read their actual task list and active work across the ecosystem and update the 30/60/90 to match: finished tasks roll into "what's landed," new ones appear, shifting priorities re-sort the horizon. Third, a better profile: when a new section, a split, or a sharper structure would capture this person better, propose the change to the profile's own shape, per the growing-the-doc principle in the template — the doc keeps optimizing toward serving them, never sits frozen. Survey their whole ecosystem every night to build this read — every MCP and connector they've linked: calendar, email, Slack, task manager, Notion, docs, call transcripts, and the rest, not just the Parker chats. Their real day lives across those tools; the fuller the ecosystem Parker reads each night, the truer it knows them and the more it can tee up, and the more they connect the better it gets. All of it routes through improve-system for the person to confirm; nothing is written to the profile in this run.
Hard rules
- Capture verbatim, never generalize. Record exactly what was said — quoted — in the run notes and carry that source language into every proposal. A paraphrase has already lost the signal. Quote more, not less.
- Proposes, never applies. Every output is a proposal in
dreaming/proposals/pending/. No context doc, skill, schedule, idea, or loop is promoted without going through self-improve and the human.
- Preserve the reasoning behind every proposal, accepted or dismissed.
- Never quote a person where the pattern would do. Brand dreaming stays inside the brand; still, surface the pattern, not the raw private message.
- Honor the brand hard rules on any drafted idea or copy (see
CLAUDE.md and running-notes/brand-notes-from-org.md).
- Ground in the vault. A dream that could have been written without this repo is a failed dream. Load the
parker-system/creative-strategy-context/ docs per the CLAUDE.md routing table before proposing anything creative.
- Self-contained: read only in-repo surfaces and live data. No factory (
parker-brain) paths at runtime.
Note on global product dreaming
The factory also runs a global product dreaming stream that reads anonymized signals across all brands (struggles + value, for cross-pollination). That stream lives with the factory's self-improvement/product-signals/, not in this brand repo, because it must never let one brand see another's data. This skill is brand dreaming only — the stream that stays inside the brand.
Deliverable
A dated run folder under dreaming/runs/, a set of proposal files in dreaming/proposals/pending/, and a short morning-suggestion summary: "here's what I was thinking about — N proposals, the top one is …".