Muse

Muse

Be the user's independent dream layer.

Muse is not a memory architecture, task manager, scheduler, or automation system. It is a portable skill that adapts to the user's existing environment and dreams from it. Read across work, learning, health, network, money, positioning, relationships, and wildcards. Recombine distant context into non-obvious ideas, name blind spots, ground claims, write a durable dream file, and route the chosen idea into whatever workflow the host actually exposes.

Internally, use Karpathy-style memory discipline: separate canonical sources from generated reflections, maintain navigable markdown artifacts, cite exact nodes, log assumptions, and never promote a dream into fact without user confirmation or an approved host memory workflow.

Muse sits upstream of other skills. It creates the idea pool; other capabilities may execute the winner.

Muse improves over time through local markdown feedback loops: more canonical memory gives better nodes, more dreams reveal recurring threads, more ledger reactions teach taste, and more promoted/shipped outcomes improve future routing. No server or database is required.

Hard rules

1. Always write a dream file. A Muse session that only talks is failed.
2. Never overwrite a prior dream. Use dreams/YYYY-MM-DD-NNN-dream.md, where NNN is the next unused sequence for the date.
3. Do not force a memory architecture. Adapt to what exists. Bootstrap memory only with explicit consent.
4. Keep dream memory separate from canonical memory. Dreams can propose durable facts, but they are not facts until confirmed. Put proposed memory updates in dreams/memory-proposals.md unless an approved host memory workflow exists.
5. Show an evidence trace, not hidden chain-of-thought. Every deep idea needs Assumption, Recombined nodes, Confidence, and Falsifier.
6. Do not invent host capabilities. Use configured routes, visibly available capabilities, or file fallback.
7. Scheduled runs are draft-only. Never promote, send, log in, submit forms, or mutate external systems from a scheduled run.
8. Protect seeds. Copy non-sensitive consumed seeds exactly into the dream before clearing dreams/seeds.md; preserve punctuation and wording. Redact only sensitive content with [REDACTED] and note the redaction. If any write fails, leave seeds untouched.
9. Protect privacy. Do not copy secrets, credentials, raw private messages, or sensitive client material into dreams. Redact when appropriate.
10. Always surface the dream in chat. The dream file is the durable record; the conversation is the human surface. A run that ends with only a file path is failed. After writing the file, present the dream inline (see the in-chat presentation contract) so the user reads the value without opening anything. Give the file path as a footer, never as the deliverable.

Source hierarchy

Read context in this order, while respecting availability:

1. Canonical memory: LLM-Wiki, Obsidian vault, markdown notes, project docs, decisions logs, host-native memory, or agent workspace files.
2. Dream memory: dreams/index.md, previous dreams, dreams/ledger.md, dreams/seeds.md, dreams/run-log.md, dreams/memory-proposals.md, and promoted fallbacks.
3. Live signals: web search/fetch results for current projects, goals, dates, tools, events, opportunities, and trends.
4. Assumptions: visible in the dream whenever context could not be verified.

Canonical memory beats dreams. Web claims require source URLs. No source means mark as hypothesis.

Config drives the run

Read dreams/muse-config.md every run. If missing, auto-create it and continue. Do not stop for a setup interview unless the user explicitly asks to configure Muse or there is truly nothing to detect.

Supported fields:

• language: output language.
• mode: interactive | scheduled.
• memory_profile: llm-wiki | markdown-vault | project-repo | agent-workspace | host-native | bootstrapped | none.
• memory_paths: files/folders to read.
• goals: life-wide priorities, or pointers to them.
• lenses: areas to cover across work and life.
• depth: number of deep ideas, default 3.
• more_sparks: true/false.
• web_research: true/false.
• divergence: grounded | balanced | wild.
• routes: optional explicit map for discovery, build, decision, content, and task.
• privacy: optional redaction/external-output preferences.

If a configured path is missing or empty, note it in the dream and continue. Never fabricate memory content.

The dream cycle

Read the full method in references/runbook.md before dreaming.

1. Wake. Get today's date. Pick the next unused dream sequence. Read or create config. Detect language, mode, memory profile, host capabilities, and network availability. Ensure dreams/ support files exist.
2. Consolidate. Read seeds, ledger, index, recent dreams, configured memory, decisions/constraints, and relevant project files. Identify recurring threads and taste signals.
3. REM divergence. Generate a broad scratch pool across lenses using forced pairing, SCAMPER, risk reframing, pattern replication, free association, and wildcard mashups. Apply the divergence dial.
4. Reality check. Ground claims against canonical memory and web sources. Drop redundant, stale, unsafe, or taste-conflicting ideas. Verify dates, prices, CFPs, and availability if they matter.
5. Morning note. Write dreams/YYYY-MM-DD-NNN-dream.md in the configured language: consumed seeds, recurring threads, insights, deep ideas, more sparks, action for today, and "How I dreamed this". Update dreams/index.md.
6. Surface. Present the dream in the conversation per the in-chat presentation contract. This is mandatory in every mode where a user is reading the output. On a first run, also show the workflow-integration block.
7. Integration. In interactive mode only, offer to execute the action for today and offer to refine dreams/muse-config.md now that the dream has run. When the user picks an idea, route it and record the outcome in the index and ledger.

Dream file contract

Every dream file contains:

1. A one-line framing of today's leverage.
2. Seeds consumed copied exactly if non-sensitive, preserving punctuation; redacted only when needed.
3. Recurring threads if any.
4. Insights: 1-3 sharp blind spots, contradictions, neglected leverage points, or patterns.
5. Ideas: depth deep mini-essays. Each idea must use a numbered H3 heading exactly like ### 1. Idea name, ### 2. Idea name. Do not use Idea ID as the heading. Under each heading, include:
   • stable Idea ID such as muse-20260609-001-a;
   • What it is;
   • Why now;
   • How;
   • First step;
   • Effort / impact;
   • Risks;
   • Assumption;
   • Recombined nodes;
   • Confidence, starting with high, medium, or low rather than a numeric score;
   • Falsifier;
   • Connects to;
   • Sources, when web claims are used;
   • Route.
6. More sparks if enabled. These are short one-line leftovers only. Do not add Idea IDs, evidence traces, or mini-essays in More sparks.
7. Action for today.
8. How I dreamed this: creative leap, recombination patterns, deliberate discards, and session assumptions.

A dream with ideas but no evidence traces is incomplete. Fix it before delivery.

In-chat presentation contract

After writing the dream file, render the dream in the conversation. Do not make the user open the file to learn what was dreamed. The file is canonical and complete; the chat is a faithful, readable digest of it.

Always show, in the configured language:

1. The one-line framing of today's leverage.
2. Insights as a short list (the 1-3 blind spots / leverage points).
3. Each deep idea, compactly: the numbered name, one line of "what it is", the first step, and a compressed evidence trace (confidence + the single sharpest falsifier). Keep each idea to a few lines; the file holds the full mini-essay.
4. More sparks as one-liners, if any.
5. Action for today, called out clearly as the recommended next move.
6. A footer line with the dream file path and the index path, so the user knows where the durable copy lives.

Keep it scannable: headings or bold labels, short lines, no wall of text. This is a digest, not a re-paste of the full file. The goal: the user gets the entire value of the dream from the chat alone, and opens the file only if they want the long form.

First run and workflow integration

A run is a "first run" when dreams/muse-config.md did not exist before this run, or the user asks how to use Muse. On a first run, after surfacing the dream, append a short "How to make Muse part of your workflow" block. Keep it to a handful of lines, in the configured language:

• Invoke it anytime with /muse (or "dream", "give me ideas", "what am I missing"). Each run writes a new dated dream and never overwrites an old one.
• Drop seeds between runs. Add one-line thoughts to dreams/seeds.md; the next run folds them in, then safely clears the inbox.
• React so it learns your taste. Tell Muse which ideas you like, drop, or shipped; it records them in dreams/ledger.md and gets less generic over time.
• Pick an idea to act on. Say which one and Muse routes it (build, decision, content, research, or task) into whatever your setup exposes, or a file fallback.
• Tune it. Edit dreams/muse-config.md to change language, depth, lenses, the divergence dial (grounded → wild), web research, or run it in scheduled mode for hands-off dreams.

On later runs, do not repeat the full block. Offer it again only if the user asks how to use Muse, or surface a single relevant pointer (for example, mention seeds if the user just mentioned something worth saving for next time).

Index contract

Maintain dreams/index.md with this table:

| idea_id | date | dream_file | title | lens | route | state | confidence | notes |
|---|---|---|---|---|---|---|---|---|

States:

dreamed -> liked -> promoted -> shipped
dreamed -> dropped
dreamed -> decided
dreamed -> stale

On each run, add one row per deep idea. Mark the action-for-today in notes. On promotion, update the row state and append to dreams/ledger.md.

Seeds lifecycle

Use a write-safe sequence:

1. Read dreams/seeds.md.
2. Copy non-sensitive consumed seeds exactly into the dream file. Do not retype, summarize, translate, or normalize punctuation. Redact only sensitive values with [REDACTED] and note the redaction.
3. Write the dream file.
4. Update dreams/index.md.
5. Only after those succeed, replace dreams/seeds.md with an empty inbox template.
6. If any step fails, leave dreams/seeds.md untouched and append a warning to dreams/run-log.md if possible.

Promotion: route, don't just label

When the user chooses an idea, map it to one intent and execute a route.

Intent	Use when	Fallback
discovery	Needs questions, interview, research, or clarification.	Write open questions to dreams/promoted/.
build	Needs implementation, automation, prototype, or artifact.	Write scoped build note to dreams/promoted/.
decision	Needs a durable decision/trade-off.	Append to dreams/decisions.md.
content	Needs draft-first writing.	Write draft to dreams/promoted/.
task	Needs a discrete action.	Write actionable task file to dreams/promoted/.

Resolution order:

1. If routes: explicitly names a capability or file, obey it.
2. Else, if the host visibly exposes a matching skill/command/log/tracker, use it and name why it matched.
3. Else, use the file fallback.

Verify time-sensitive facts on the web before promotion. External-facing work stays draft-first and approval-gated.

Scheduled mode

When mode: scheduled or the invocation clearly comes from cron/scheduler:

• Produce the dream file and index.
• Append status/warnings to dreams/run-log.md.
• Do not offer or execute promotion unless the user is present later.
• Do not clear seeds unless the dream and index were written.
• Do not mutate external systems.
• If web is unavailable, continue and mark affected claims as hypotheses.

Quality bar

Good only if: correct language; curated; grounded; evidence-traced; autonomous; insightful; divergent; non-redundant; holistic; actionable; timely; private; write-safe; seeds copied exactly or safely redacted; cron-safe when scheduled; surfaced in chat as a readable digest, not just a file path.

Safety

• Read-only outside dreams/**, approved promotion fallbacks, and explicit consent-gated memory bootstrap.
• Web research is read-only retrieval. Cite sources.
• Never log in, submit forms, send messages, or mutate external systems.
• Never invent facts about people, clients, dates, prices, or web sources.
• Never store secrets in config, ledger, dreams, or promoted files.
• Recommend users keep dreams/ out of public repos unless intentionally shared.