| name | idea-lifecycle |
| description | Portfolio lifecycle management for divergent minds. Processes documents into a patchable knowledge graph of ideas, concepts, projects, and insights. Operations: ingest, extract, review, relate, triage, prune, diagnose, graph. Use when the user mentions: managing ideas, brain dump, idea backlog, "too many ideas", "what should I work on", portfolio review, "what's stale", "find patterns", "what's duplicated", consolidate, "extract ideas from this", "show the graph", mosaic, "process my notes", or uploads portfolio-state.json or documents for idea extraction. Also trigger on "I keep starting things and not finishing" or "I need to get my ideas out of my head".
|
Idea Lifecycle Management
Purpose
Divergent minds produce ideas faster than they can process them. The
failure mode is not lack of creativity — it is portfolio collapse:
too many open threads, no triage discipline, no kill criteria, and
no way to see the whole landscape.
This skill provides the scaffolding to prevent that collapse. It
manages a portfolio — a persistent, structured collection of
items with typed relationships, provenance tracking, and pattern
detection — exposed as an incrementally-patchable knowledge graph.
The graph is a mosaic: each document processed, each brain dump
captured, each connection discovered adds tiles. Over time, the
mosaic reveals the user's core preoccupations, recurring patterns,
productive tensions, and dead weight.
First Steps: Load References
Before any operation, read the relevant reference files:
-
Always read first:
view <skill-path>/references/taxonomy.md
Defines item types, maturity stages, relationship types,
provenance, pattern types, priority signals, kill criteria,
and graph operations.
-
For document processing, graph ops, or diagnostics:
view <skill-path>/references/graph-operations.md
Defines extraction protocol, patching, merging, diagnostic
analysis, and visualization protocol.
-
For state file handling or artifact building:
view <skill-path>/references/schema.md
Defines JSON schemas, file conventions, visual encoding,
and artifact specifications.
Detecting the User's Intent
Map what the user says to the correct operation:
| User Says (examples) | Operation |
|---|
| "I had an idea...", "brain dump", "capture this" | ingest |
| "Process this document", "extract ideas from this PDF" | extract |
| "Show me my portfolio", "what do I have", "overview" | review |
| "How does X connect to Y", "find connections", "map" | relate |
| "What should I work on", "prioritize", "triage" | triage |
| "What should I kill", "what's stale", "clean up" | prune |
| "Find patterns", "what's duplicated", "consolidate", "recurring themes" | diagnose |
| "Show the graph", "visualize", "mosaic", "big picture" | graph |
| [Uploads portfolio-state.json] | review (load state first) |
| [Uploads document + asks about ideas] | extract |
If ambiguous, default to review.
Operations
Operation 1: Ingest
Purpose: Capture new items from direct user input with minimal
friction.
Process:
-
Accept whatever the user provides — sentence fragments, walls
of text, lists, links, stream-of-consciousness. Do not ask
the user to restructure their input.
-
For each identifiable item:
a. Preserve raw_input verbatim.
b. Classify type using taxonomy. Explain briefly.
c. Write core_statement in problem-language.
d. Set stage to raw, provenance.source_type to manual.
e. Suggest 2–4 tags.
f. Scan existing items for relationships. Propose tentative links.
-
Present items for batch confirmation.
-
Commit to portfolio state. Generate patch.
Divergent-mind rules:
- Never say "too vague." Vagueness is normal at spark stage.
- If 20 items are dumped at once, process all 20.
- Untangle compound paragraphs into separate items. Note what
was split and why.
Operation 2: Extract
Purpose: Process uploaded documents into portfolio items with
full provenance, then integrate into the existing graph.
This is the primary document-to-graph operation.
Process:
Follow the Document Extraction Protocol in graph-operations.md:
- Pre-flight: Hash check, document registration, file reading.
- Extract: Apply the strategy for the document type. Each
distinct idea, claim, question, decision, or reference becomes
a candidate item with full provenance.
- Cross-reference: Compare every extracted item against the
existing portfolio. Propose relationships.
- Dedup check: Flag potential duplicates.
- Pattern check: Run a lightweight pattern scan to see if the
new items trigger any new patterns (especially recurrence).
- Present extraction report for user confirmation.
- Commit confirmed items and relationships. Generate patch.
Multiple documents: If the user uploads several files at once,
process each sequentially. After all are extracted, run a cross-
document relationship scan and a full pattern detection pass.
Key principle: Extraction is not just ingestion — it is
integration. Every extracted item must be checked against the
existing graph. The value is in the connections, not the items
alone.
Operation 3: Review
Purpose: Show the user their portfolio state.
Modes:
- Full review: Markdown portfolio snapshot (schema.md format).
Group by type → stage. Show counts, clusters, recent session log.
- Filtered review: Subset by stage, tag, cluster, document
source, or pattern involvement.
- Interactive review: React graph artifact (see Operation 8).
Always surface during review:
- Unprocessed backlog (raw items, 3+ sessions old).
- Parked items whose reactivation trigger may have fired.
- New clusters formed since last review.
- Orphan items (zero relationships).
- Active patterns from last diagnostic.
Operation 4: Relate
Purpose: Discover and record relationships between items.
Process:
- If user specifies items: analyze and propose relationship type.
- If user asks for broad scan: pairwise comparison of all non-dead
items (prioritize shared-tag and same-cluster pairs).
- Propose relationships with reasoning and confidence level.
- Detect new clusters (3+ items, 2+ shared relationships).
- Present for confirmation. User overrides Claude's proposals.
- Commit. Generate patch.
Operation 5: Triage
Purpose: Decide what to focus on next.
Process:
-
Get current context (time, energy, constraints) — extract from
conversation if already stated, otherwise ask once.
-
Re-evaluate priority signals for all non-dead items against
current context.
-
Produce structured recommendation:
- Recommended Focus (1–3 items): why each ranks here.
- Worth a Session (next tier): what would promote them.
- Park Candidates: with proposed reactivation triggers.
- Kill Candidates: with kill criteria citation.
-
User decides. Update stages. Generate patch.
Rules:
- Triage output fits two screens max — even for 50 items.
- Respect conviction as data.
- Name the cost of switching from an active item.
Operation 6: Prune
Purpose: Remove dead weight.
Process:
- Scan all items against kill criteria (taxonomy §7).
- For each candidate: state criterion, summarize item, ask for
decision (kill / merge / one-more-cycle).
- Backup state before any kills.
- Execute approved kills and merges. Generate patch.
Operation 7: Diagnose
Purpose: Analyze the graph structure to surface patterns,
health issues, duplications, recurring themes, consolidation
opportunities, and actionable recommendations.
This is the core analytical operation for the mosaic.
Process:
Follow the Diagnostic Analysis Protocol in graph-operations.md:
- Duplication scan: Find near-identical core_statements.
- Recurrence detection: Find themes across 3+ independent items.
- Convergence / divergence analysis: Items clustering or
fragmenting.
- Dependency chain detection: Find bottlenecks.
- Staleness analysis: Find forgotten clusters and processing
backlogs.
- Consolidation opportunities: Mergeable items, promotable
concepts, orphaned references.
Produce the full Diagnostic Report (format in graph-operations.md).
After presenting findings:
- Ask which actions to take.
- Apply approved actions as a patch.
- Update pattern records (mark resolved/dismissed).
- Regenerate the graph artifact if one exists.
Scheduling: Recommend running diagnose every 3–5 sessions or
whenever the portfolio exceeds 30 non-dead items.
Operation 8: Graph
Purpose: Generate and present the interactive knowledge graph
artifact — the visual mosaic.
Process:
-
Project the portfolio state into the visualization data
structure (schema.md § Graph Visualization Data).
-
Generate a React artifact using d3-force layout with the
visual encoding defined in schema.md:
- Node shape = item type
- Node color = maturity stage
- Edge style = relationship type
- Edge width = confidence
- Cluster regions = background shading
- Pattern highlights = pulsing on involved nodes
-
Include filter controls (type, stage, tag, cluster, source,
pattern involvement).
-
Include click-to-inspect for nodes, edges, clusters, patterns.
-
Store the graph state in window.storage under key
portfolio-state for cross-session persistence.
-
If a diagnostic has been run, overlay active patterns on the
graph with visual indicators (see graph-operations.md §5).
Graph patching: When new items are added (via ingest or
extract), the graph artifact should be regenerable from the
updated portfolio state. The artifact reads from storage on
mount — it does not maintain its own separate state.
Diff view: If the user asks "what changed", render a diff
overlay: new nodes highlighted, removed nodes grayed, new edges
in a contrasting color.
State Management
Session Start
-
Check for uploaded portfolio-state.json.
- If found: load, validate, report counts.
- If not found: check
window.storage. If found there, use it.
- If neither: ask if this is a first session. If yes, create
empty portfolio.
-
Read file-reading skill if documents are uploaded alongside
the state file.
Session End
After any state-modifying operation:
- Generate the session's patch object.
- Append patch and session log entry.
- Write
portfolio-state.json to /mnt/user-data/outputs/.
- If a React artifact exists, update
window.storage.
- Present the state file for download.
- Brief summary of session: items added/killed, relationships
created, patterns detected.
Integration with Other Skills
| Need | Hand Off To |
|---|
| Deep problem decomposition on one item | problem-space-exploration |
| Choosing between competing approaches | arch-decision-analysis |
| Item matured into a full spec | codebase-requirements or bsl |
| Item needs to become a document | docx or pdf |
| Surgical edit to a markdown portfolio doc | doc-patch |
| Reading uploaded documents for extraction | file-reading / pdf-reading |
When handing off, pass: core_statement, raw_input, tags,
relationships, cluster context, and relevant pattern findings.
Anti-Patterns
- Over-structuring intake. Accept raw input. Claude structures.
- False precision. Categorical signals, not numerical scores.
- Guilt-tripping kills. Killing is portfolio hygiene.
- Ignoring the graveyard. Dead items are data. Check them
during relate and diagnose.
- Treating portfolio as a to-do list. It is a network, not
a sequence.
- Session amnesia. Always write state. Always generate patch.
- Extraction without integration. Never just add items from
a document — always cross-reference against existing graph.
- Pattern-without-action. Every diagnosed pattern must include
a concrete recommended action. Observation without prescription
is waste.