Panning For Gold

Panning for Gold

Turn raw, unstructured capture into an evaluated, actionable inventory of threads. Four phases: Front-gate, Extract, Evaluate, Synthesize.

Triggers

When to Use

Use this skill when:

• You have a voice transcript, stream-of-consciousness note, or brain dump that contains many ideas mixed with noise.
• You need a deterministic procedure to triage which ideas merit deeper analysis.
• You want a permanent, structured artifact that the analyst agent or the spec-generator skill can consume.

Use a different skill when:

• The input is already structured (specs, ADRs, code). Use the analyst agent or the spec-generator skill instead.
• You only need a one-off summary. Use a generic summarizer.

Workspace Layout

The skill operates inside a workspace root. The default is ./.panning/ in the current working directory; override via --workspace <path> or the PANNING_WORKSPACE environment variable.

<workspace>/
|-- transcripts/          # Raw input (read-only after capture)
|-- inventories/          # Pass1 + final inventories
|-- evaluations/          # Per-thread evaluation files
`-- gold-found/           # Synthesized gold-found files

The script creates missing subdirectories on demand. It never overwrites an existing file unless --force is passed.

Process

Phase 0: Front-gate (run before Extract)

Phase 0 is documentation-only and runs in the agent or human prompt before any pan.py invocation. There is no script enforcement; the LLM or human invoking the pipeline runs the six questions inline (or via the gstack front-gate-before-pipeline skill if it is available in the workspace). pan.py is unchanged and starts at Phase 1.

Run the six forcing questions against the input itself:

1. Demand Reality: Is anyone (including you) actually waiting on insight from this brain dump, or did you capture it because something was on your mind?
2. Status Quo: What happens if this transcript sits untriaged for another month? If "nothing breaks," that's a signal.
3. Desperate Specificity: Name the specific decision, person, or artifact that needs the gold-found output.
4. Narrowest Wedge: Could you extract one High-Signal thread by hand in 10 minutes instead of running the full four-phase pipeline?
5. Observation: Have you re-read the transcript recently, or are you triaging from memory of why it felt important when captured?
6. Future-fit: If extraction produces nothing actionable, will you delete the gold-found file or let it accumulate?

Halt criteria: if Demand Reality is aspirational ("might be useful someday") or Desperate Specificity can't name a downstream consumer, the right move is to discard the transcript or archive it unprocessed. Triaging unmotivated capture produces gold-found files nobody reads.

Phase 1: Extract

Build a pass1 inventory from a transcript without filtering. Every thread that surfaces is recorded.

Each thread block uses this structure:

## Thread N: <short title>

- **Signal**: high | medium | low
- **Quote**: "<verbatim quote that surfaced this thread>"
- **Context**: <one sentence on where this came from>
- **Initial take**: <one sentence on why this might matter>

The pass1 inventory is provisional. A second pass refines and merges into a final inventory.

Phase 2: Evaluate

For each thread in the final inventory, write an evaluation file. The evaluation captures depth: what evidence supports the thread, what would falsify it, what action it implies.

Evaluations are independent. They can be authored serially or in parallel.

Phase 3: Synthesize

Combine the final inventory and the evaluation files into a gold-found markdown file. The gold-found file groups threads by signal level (High-Signal, Medium-Signal, Low-Signal, matching the section headers in references/gold-found-template.md) and includes a metadata block at the top.

Elaboration gate (documentation-only, mandatory for High-Signal threads): Like Phase 0, this gate is LLM-applied, not script-enforced. scripts/synthesis.py appends the raw evaluation content for each thread; it does not parse for Connects to: or fail synthesis when the line is missing. The author of the evaluation (or a manual post-synth edit on the gold-found file) is responsible for producing the line. Future work may add a pan.py validate --gold-found check; until then, the acceptance checklist below is the gate.

For each High-Signal thread, write one explicit connection to an existing artifact. Search these paths:

• Skills: .claude/skills/<name>/SKILL.md
• ADRs: .agents/architecture/ADR-*.md
• Serena memories: walk .serena/memories/** (topic subdirectories under .serena/memories/<topic>/<memory-name>.md) and load via mcp__serena__read_memory("<topic>/<memory-name>.md"). mcp__serena__list_memories returns top-level indexes only and does not enumerate atomic memories; see .serena/memories/README.md.
• Open issues: GitHub issues in the current repo
• Prior session logs: .agents/sessions/YYYY-MM-DD-session-*.json

Write the connection as a one-liner under the thread in the evaluation file (consumed by synth) or add it manually to the gold-found file after synthesis: Connects to: <artifact name> (<one-line why>).

If no connection exists, treat that as a flag, not a feature. Genuinely novel insights are rare; usually "no connection" means either (a) the thread isn't as High-Signal as it looked, or (b) you haven't searched hard enough. Run a vault/skill/session search before promoting it. The forcing function catches false positives and produces compounding knowledge instead of orphan notes.

This is the elaboration principle (Make It Stick) applied at the synthesis layer: new findings are only durable knowledge once wired to existing entities. Standalone artifacts decay; connected ones compound.

CLI

pan.py init      --workspace <path>
pan.py validate  --inventory <file>
pan.py merge     --pass1 <file> --final <file> --output <file>
pan.py synth     --inventory <file> --evaluations <dir> --output <file>

The script delegates to inventory.py for parsing and merging, and synthesis.py for gold-found generation.

Scripts

Verification

• Phase 0 front-gate was answered before invoking pan.py (six questions, halt criteria evaluated).
• Inventory thread blocks contain Signal, Quote, Context, Initial take fields.
• Final inventory deduplicates threads from pass1 by title.
• Gold-found file has a metadata block plus High-Signal, Medium-Signal, and Low-Signal sections in that order.
• No threads from the inventory are silently dropped.
• Every High-Signal thread in the gold-found file has a Connects to: line pointing at a skill, ADR, Serena memory, open issue, or prior session log (elaboration gate).

References

• references/inventory-template.md - thread inventory skeleton.
• references/gold-found-template.md - gold-found skeleton.