Plaud

Plaud (agent-first CLI)

Use this skill to pull content out of Plaud in a way that’s reliable for agents: stable --json envelopes, local search, explicit paging/filters, and honest coverage boundaries.

Prime directive: optimize for recall

Completeness and accuracy matter more than efficiency. Treat every search result as a candidate set until you have checked the coverage warnings and run enough compensating searches.

Hard filters and tidy result lists are not proof of completeness:

• A speaker/name search only finds recordings where the name was detected, labeled, or appears in searchable text. It can miss unlabeled Speaker 1 / Speaker 2 recordings.
• An exact project search can miss allusions, nicknames, old names, misspellings, and conversations where the project is implied but never named.
• Date filters can hide older context or recordings with surprising timestamps.
• A clipped top-N result list can hide additional candidates. If data.coverage.riskFactors.truncated is true, increase --limit and keep triangulating.
• Generic speaker terms, very short aliases, and broad multi-term fuzzy queries are high-risk candidate generators, not evidence of completeness.
• --ids-only protects tokens and privacy, but it does not make the result exhaustive.
• --snippets validates candidates; it does not prove that non-returned recordings are irrelevant.

Default workflow (read-only)

• Sanity check: plaud doctor --json
• Browse recent files: plaud files list --limit 25 (table) or plaud files list --json --limit 25
• Drill in: plaud files get <id> --json
• Pull content into an explicit scratch/export directory: plaud files download <id> --out <dir> --what transcript,summary,json

Common tasks

• Download audio: plaud files download <id> --out <dir> --what audio --audio-format opus
• Bulk export a scoped range: plaud files export --zip --since 2026-01-01 --until 2026-02-01
• Local store search: plaud files sync --max 200, then plaud files search "keyword" --ids-only
• Include snippets only when transcript/summary excerpts are appropriate for stdout: plaud files search "keyword" --snippets
• Trash / restore: plaud files trash <id> and plaud files restore <id>
• Tags: plaud files tags list --json, plaud files tags add <tagId> <id>, plaud files tags clear <id>
• Speaker labels (per recording): plaud files speakers list <id> --json and plaud files speakers rename <id> --from "Speaker 2" --to "Person A"
• Re-run transcript/summary: plaud files rerun <id> --wait

Recall-first search workflow

For comprehensive retrieval, do not stop after one exact query or one filter.

1. Start with a broad local store query:
   • plaud files search "topic or person" --ids-only --json
2. Read meta.coverageWarnings and data.coverage.riskFactors before judging completeness.
   • If truncated is true, increase --limit and treat the first response as incomplete.
   • If warnings mention generic speaker terms, short aliases, or multi-term fuzzy search, add more precise/adjacent queries before making a claim.
3. Run adjacent searches:
   • aliases: Jordan, J. Lee, Jordan Lee
   • project variants: Project Atlas, Atlas project, old names, related people, adjacent domain terms
   • people/context: cofounders, advisors, customers, topic terms, dates
   • unlabeled-speaker risk: Speaker 1, Speaker 2, unknown, title-known names
4. Use --from / --to to narrow only after a broad pass has established the likely time range.
5. Use --snippets only on likely candidates or query variants that need validation.
6. For final answers, state the coverage receipt: queries used, broad counts, high-confidence IDs, suspected misses, and why filters may be lossy.

Bad pattern:

plaud files search "Project Atlas" returned two results, so these are all Project Atlas conversations.

Good pattern:

Exact search for Project Atlas returned two direct hits. Adjacent searches for related people and domain language returned more candidates, so I am treating the exact hits as confirmed direct matches and the adjacent hits as review candidates. Coverage remains non-exhaustive because allusions, unlabeled speakers, and missing transcript/summary text can hide relevant recordings.

For hard fuzzy asks such as "conversations involving disagreement, urgency, or a difficult decision" or "entries with unlabeled speakers whose identities may be inferable from context", build a small query matrix first. Include direct terms, aliases, adjacent people, emotional or domain vocabulary, and generic speaker terms. Review snippets or transcripts for the resulting candidate set before making a completeness claim.

The following names and projects are synthetic examples.

Example for a person:

plaud files search "Jordan Lee" --ids-only --json
plaud files search "Jordan" --ids-only --json
plaud files search "J. Lee" --ids-only --json
plaud files search "Speaker 1 Jordan" --ids-only --json

Example for an alluded-to project:

plaud files search "Project Atlas" --ids-only --json
plaud files search "billing migration" --ids-only --json
plaud files search "invoice reconciliation" --ids-only --json
plaud files search "pilot customer rollout" --snippets --json

Paging and filtering (precision tools, not proof)

• Prefer filters over dumping everything only after you have considered recall risk:
  • Last N: plaud files list --json --last 30d --limit 50
  • Date range: plaud files list --json --from 2026-02-01 --to 2026-03-01 --limit 50
• Page forward: plaud files list --json --skip 50 --limit 50
• If results change during paging, anchor with --to (or use a narrower range) and retry.
• If a user asks for "all", "complete", "comprehensive", or "thorough", filters are only one evidence source. Add a broad recall pass and explicitly report known blind spots.

Auth (only when needed)

If plaud doctor --json indicates missing auth:

• Best UX (interactive): plaud auth login (opens a browser, captures token, closes when done)
• Headless/remote:
  • Copy ~/.config/plaud/config.json from a trusted machine, or
  • plaud auth set --stdin (paste token via stdin), or
  • PLAUD_AUTH_TOKEN=... plaud doctor --json (ephemeral; avoid saving)
• HAR import exists (plaud auth import-har ...) but is a last resort (HARs often contain sensitive data).

Hard constraints (security)

• Never print/paste full tokens or HAR contents into chat/logs.
• Never pass tokens via CLI flags (use login/stdin/config/env).
• Prefer --json outputs; treat non-JSON text as human-only.
• plaud files sync stores readable local JSON details, transcript text, and summary text unless narrowed with --what; set PLAUD_STORE_DIR for scratch/temporary runs.
• Do not paste transcript/summary snippets into public issues, docs, or logs unless the user explicitly asks.

JSON contract (for tool use)

• Stable machine-readable behavior is documented in docs/CONTRACT_V1.md.
• files search --json includes data.coverage and meta.coverageWarnings; read them before making claims about completeness.

Public-package note

This is an unofficial CLI. Prefer Plaud's official tooling for supported auth and ordinary terminal use; use this CLI when you need JSON-first export, local store search, or de-duplication.