Skip to content

[recipes] panning-for-gold: update to v3.0.0 - #481

Draft
jaredirish wants to merge 4 commits into
NateBJones-Projects:mainfrom
jaredirish:recipes/panning-for-gold-v3
Draft

[recipes] panning-for-gold: update to v3.0.0#481
jaredirish wants to merge 4 commits into
NateBJones-Projects:mainfrom
jaredirish:recipes/panning-for-gold-v3

Conversation

@jaredirish

@jaredirish jaredirish commented Aug 21, 2026

Copy link
Copy Markdown
Contributor

Summary

Updates the panning-for-gold recipe (merged previously) with five months of production lessons, folded into the existing v2.0.0 structure rather than a rewrite. All additive, nothing removed or contradicted.

New in v3.0.0:

  • Stage 0: Open Brain Pre-Load. Builds a context dossier (people, projects, recent activity) before extraction starts, so every thread gets tagged with its prior connection instead of being read cold. Capped query budget, graceful degradation when Open Brain is sparse.
  • Deadline Audit. Three-step check (literal quote / inferred from a named anchor / the model's own extrapolation) that stops the skill from inventing urgency a transcript never established. Conflicting source dates default to the conservative reading.
  • Gap Detection. After evaluation, an explicit pass for what got missed, what Open Brain already knew but nobody said out loud, and who owes whom what.
  • Hardened speaker consolidation. A zero-speaker-labels presence check (some exports have none at all), a rule for when a diarizer's speaker ceiling is exceeded by the actual recording, and a mandatory biographical-attribution quiz for any claim that could get projected onto the wrong person.
  • Advisory-only dedup on writeback. A hard-blocking dedup guard is worse than one that lets an occasional duplicate through — this version never hard-skips a capture on a similarity score.

What Changed

File Change
recipes/panning-for-gold/panning-for-gold.skill.md +5 new phases/sub-steps, expanded Lessons Log, 2.0.03.0.0
recipes/panning-for-gold/metadata.json Version bump, description updated
recipes/panning-for-gold/README.md New "What's New in v3.0.0" section, updated setup steps and troubleshooting for the new phases

Validation

  • metadata.json validated against .github/metadata.schema.json directly (not just json.load) — the schema is additionalProperties: false, which caught a companion_recipes field I'd initially added, pointed at a flywheel-retrieval recipe that turns out was never actually merged upstream (only exists in my own local unmerged checkout). Removed the field, but the first fix over-corrected and degraded the corresponding links to unlinked prose instead of noticing recipes/live-retrieval/ (merged, PR [recipes] Live Retrieval: ambient Open Brain context surfacing #128) was the actual companion recipe under a different name. A follow-up commit repoints all 4 references to it.
  • Ran markdownlint-cli2 against this repo's actual .github/.markdownlint.jsonc config (not the tool's defaults) — clean.
  • Grepped all three files for personal names, client names, and file paths per the repo's existing scrubbing checklist — clean except the author byline (unchanged, matches the original submission's attribution).
  • Diffed against upstream/main to confirm this branch changes only recipes/panning-for-gold/*.

Testing

Each capability in this update (Stage 0 pre-load, deadline audit, gap detection, the speaker-consolidation edge cases, advisory-only dedup) is a rule the underlying skill has been running under in production on my own Open Brain instance for the past ~5 months — the Lessons Log inside panning-for-gold.skill.md documents the specific production incident each rule traces back to, dated as it happened. This PR anonymizes and packages that live skill; it is not a from-scratch draft.

Requires: an Open Brain MCP connection (search_thoughts, capture_thought, list_thoughts) and an AI coding tool that supports skills/system prompts (Claude Code or similar). No paid services required beyond what the original v2.0.0 recipe already listed.

Marked as draft, opening for maintainer review before marking ready.

Generated with Claude Code

Five months of production lessons folded into the merged v2.0.0
recipe, in the same voice and phase structure:

- Stage 0 Open Brain pre-load (context dossier before extraction,
  capped query budget, graceful degradation when OB1 is sparse)
- Deadline Audit (literal quote / inferred-from-anchor / extrapolation,
  stops the skill from inventing urgency a transcript never established)
- Gap Detection (missed threads, unmentioned prior context, WAITING_FOR)
- Hardened speaker consolidation: a zero-labels presence check, a
  diarizer-speaker-ceiling rule, and a mandatory biographical-
  attribution quiz for any claim that could get projected onto the
  wrong speaker
- Advisory-only dedup on writeback (a guard that blocks is worse than
  one that lets a duplicate through)
- companion_recipes now links both directions with flywheel-retrieval

All additive on top of the merged version, nothing removed or
contradicted. Scrubbed of personal names, client names, and file
paths per the repo's existing checklist.
@github-actions github-actions Bot added the recipe Contribution: step-by-step recipe label Aug 21, 2026
@github-actions

Copy link
Copy Markdown

Hey @jaredirish — welcome to Open Brain Source! 👋

Thanks for submitting your first PR. The automated review will run shortly and check things like metadata, folder structure, and README completeness. If anything needs fixing, the review comment will tell you exactly what.

Once the automated checks pass, a human admin will review for quality and clarity. Expect a response within a few days.

If you have questions, check out CONTRIBUTING.md or open an issue.

- Wrap fenced code blocks with blank lines (MD031, 3 spots)
- Drop companion_recipes from metadata.json — not in this repo's
  .github/metadata.schema.json (additionalProperties: false), so it
  would fail schema validation
- Replace the flywheel-retrieval markdown links (4 spots) with plain
  prose — that recipe was never actually merged upstream, only exists
  in a local unmerged checkout, so the relative links were dead
Code review caught this: the prior fix commit correctly removed dead
links to a nonexistent "flywheel-retrieval" recipe (that name only
ever existed in a local unmerged checkout), but overcorrected by
degrading them to unlinked prose. recipes/live-retrieval/ is already
merged upstream (PR NateBJones-Projects#128) and is exactly the companion this recipe
was describing. Repointed all 4 references to it.
Small technical flowchart (126KB, well under the repo's 1MB binary
limit) matching the process's own dot graph 1:1. Embedded right after
"What It Does" so it's visible before the reader hits the step-by-step.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

recipe Contribution: step-by-step recipe

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant