[recipes] Panning for Gold v2.0: battle-tested updates from 13 production sessions

[jaredirish]

Summary

Updates to Panning for Gold (PR #21) based on 13 production sessions across 4 projects. Every change in this PR was born from a real failure or user correction during daily use.

Production Metrics

• 13 panning sessions across 4 separate projects
• 99+ threads extracted and evaluated
• 6 lessons logged from real failures (compaction loss, token waste, speaker misattribution)
• 5 critical rules, each traced to a specific production violation
• Used on: partnership strategy calls, client discovery meetings, accelerator program sessions, ambient voice captures, brainstorm processing

What Changed

New: Phase 0.5 (Speaker Consolidation)

Voice transcription software creates 3-5x more speaker labels than actual speakers present. Labels shift when environment changes (room acoustics, mic distance). This phase cleans speaker data before any thread extraction, preventing misattribution of entire conversations.

Born from: a session where voices swapped between two speakers, causing 40+ threads to be misattributed. Pain points became pitches and vice versa.

New: Phase 3.5 (Open Brain Capture)

Auto-captures each ACT NOW item and a session summary to Open Brain via capture_thought. Per-thread granularity (not just a session blob). Note for users with the Auto-Capture Protocol recipe: this phase still runs because panning captures are more granular than session summaries.

Updated: DOT Flowchart

Replaced ASCII process flow with a DOT graph diagram showing the full pipeline with decision points.

Updated: Lessons Log (6 entries)

Each entry includes the date, what failed, and what rule was added. Examples:

• Background evaluators lost to compaction, synthesis never written
• Re-reading 926-line transcript burned ~30K tokens when summary covered 90%
• "Don't be stingy": first pass had 42 threads, expanded to 82 after user pushed back

Updated: Common Mistakes (#9, #10)

• Extension Proposal: Workout & Health Log #9: Trusting auto-generated speaker labels (3-5x inflation is normal)
• Extension Proposal: Reading & Media Library #10: Being stingy on first extraction (80+ threads for 1-hour multi-topic conversation is normal)

Context

From @jaredirish. Part of the Open Brain Flywheel.

cc @matthallett1 (reviewed and merged v1.0)

[justfinethanku]

Review of PR #88: Panning for Gold v2.0

This is an impressive production-hardened update that clearly demonstrates real-world use. The 13 sessions, 99+ threads, and documented failures-turned-lessons are exactly the kind of evolution Open Brain contributions benefit from. Great work.

What's Strong

Battle-tested improvements: Every change traces back to a real failure. The Lessons Log entries (compaction loss, token waste, speaker misattribution) are specific and actionable. This is valuable learning for the community.

Speaker consolidation addition (Phase 0.5): This is a standout contribution. The 10 labels for 2 people problem is well-documented and the solution (anchor lines, scene-based re-attribution) is practical. The decision framework for re-extraction is smart.

Production metrics transparency: Listing 13 sessions, 4 projects, 6 lessons in the PR description sets a high bar for v2+ contributions. This is how recipes should evolve.

DOT flowchart upgrade: Much clearer than ASCII. Shows decision points visually.

Auto-capture integration (Phase 3.5): Per-thread granularity is more useful than session blobs. Good clarification that it still runs alongside Auto-Capture Protocol.

Issues That Need Fixing

1. metadata.json version mismatch (CRITICAL)

Current state:

• Line 9: "version": "1.0.0"
• Line 19: "updated": "2026-03-13"

Expected (based on PR title and skill file):

"version": "2.0.0",
"updated": "2026-03-19"

The skill file header correctly shows version: 2.0.0 and the diff shows these changes were intended but somehow the wrong version of metadata.json is in the PR. This will fail automated validation.

Fix: Update metadata.json to match the diff from gh pr diff 88 (lines 9 and 19).

2. File path inconsistencies (MEDIUM)

The skill file has inconsistent path references across the update:

Original (v1.0): docs/brainstorming/

• Line 30 in old version: docs/brainstorming/evaluations/
• Line 50 in old version: docs/brainstorming/YYYY-MM-DD-{source}-transcript.md

Updated (v2.0): Mix of docs/meetings/ and docs/brainstorming/

• Line 30 in new version: docs/meetings/evaluations/
• Line 183 in new version: docs/meetings/YYYY-MM-DD-{source}-inventory.md
• Line 152 in new version: docs/meetings/YYYY-MM-DD-{source}-gold-found.md

But also still:

• Phase 0 (line 45): docs/brainstorming/YYYY-MM-DD-{topic}.md

This creates confusion. Either:

1. The recipe is shifting from brainstorming/ to meetings/ because production use cases were meeting transcripts (fine, but document the rationale)
2. It supports both (fine, but be explicit that users choose)
3. It's an accidental find-replace error (needs fixing)

Recommendation: Pick one pattern or document why both exist. If meetings are the primary use case now, update Phase 0 and README examples to match. If both are valid, add a note explaining when to use each.

3. Phase 0.5 anchor line examples need generalization (MINOR)

Phase 0.5 includes very specific personal anchor examples:

• "References to family members by name"
• "Career history details only you would mention"
• "Your specific projects, tools, or frameworks"

These are written as if the user is a specific person. For a community contribution, these should be generalized or framed as examples:

Suggested rewrite:

**Speaker 1 anchors (stable across transcripts):**
- Family member references (names only they would use)
- Specific projects, tools, or frameworks unique to them
- Career history details only they would mention
- Hobby or interest references that distinguish them

**Speaker 2+ anchors (build per-meeting):**
- Workplace-specific vocabulary ("our patients", "our census")
- System knowledge (internal tools only they would reference)
...

This makes it clear these are example patterns, not instructions written for one specific user.

4. DOT graph rendering check (NICE-TO-HAVE)

The DOT graph in the skill file (lines 40-68) is well-structured, but markdown renderers vary. Consider adding a note that this renders best in tools that support DOT (Obsidian with plugins, VS Code with extensions) or a fallback PNG/SVG for GitHub display.

Alternatively, keep the DOT source but also include a rendered image for the README since that's the primary discovery path.

Metadata Validation

Field	Status	Notes
name	✅ Pass	Unchanged, correct
description	✅ Pass	Clear, accurate
category	✅ Pass	recipes
author	✅ Pass	jaredirish
version	❌ FAIL	Shows 1.0.0, should be 2.0.0
requires	✅ Pass	All fields valid
tags	✅ Pass	Comprehensive
difficulty	✅ Pass	intermediate
estimated_time	✅ Pass	20 minutes
created	✅ Pass	2026-03-13
updated	❌ FAIL	Shows 2026-03-13, should be 2026-03-19

Automated Review Rules Check

1. ✅ Folder structure: Correct (recipes/panning-for-gold/)
2. ✅ Required files: README.md, metadata.json, skill file all present
3. ❌ Metadata valid: Version and updated date don't match PR title/intent
4. ✅ No credentials: Clean
5. ✅ SQL safety: No SQL files
6. ✅ Category artifacts: Skill file + detailed instructions present
7. ✅ PR format: Correct [recipes] prefix
8. ✅ No binary blobs: Clean
9. ✅ README completeness: Prerequisites, steps, outcome, troubleshooting all present
10. N/A Primitive dependencies: None declared
11. ✅ Scope check: All changes within recipes/panning-for-gold/
12. N/A Remote MCP: Not an integration/extension

Community Value

This contribution raises the bar for v2+ recipe updates. The production metrics, failure documentation, and specific lessons learned are valuable beyond just this recipe — they model how to iterate on Open Brain workflows.

The speaker consolidation phase addresses a real pain point that will affect anyone using voice transcripts. The anchor line approach is clever and doesn't require new tools or services.

Required Changes

Must fix before merge:

1. Update metadata.json line 9 to "version": "2.0.0"
2. Update metadata.json line 19 to "updated": "2026-03-19" (or the actual last update date)
3. Resolve docs/meetings/ vs docs/brainstorming/ path inconsistency — either pick one or document when to use each

Nice to have:
4. Generalize Phase 0.5 anchor examples (currently reads like instructions for one specific user)
5. Consider a rendered image of the DOT flowchart for the README

Once the metadata.json version/date fix and path consistency issue are resolved, this is ready to merge.

---

Verdict: Minor fixes needed

The metadata version/date mismatch is a quick fix but blocks merge due to automated validation. The path inconsistency needs clarity. Once those two are addressed, this is a strong contribution with clear production value.

[justfinethanku]

Fixed the metadata version and merged — incredible production-hardened update @jaredirish!