Taste Check

/taste-check — the taste gate for content

The gate is a kill switch, not a checkbox. Three outcomes, each with an owner and a consequence:

The gate's owner = this agent (the first filter). The final judge = the human owner. The reviewer is read-only: it never writes into the vault, never edits the content, it only issues a verdict.

Order of operations

1. Input class (determine it first — it decides which principles apply):
   • A a vault note · B a draft of something outbound on the owner's behalf · C a dedup merge · D a service file (memory/CLAUDE.md/skill/dashboard text) · E an AI document used for a decision (a DR report, a Decision Memo, a strategy, a PRD, a spec, research — anything people will "argue over, buy, or build from").
2. Jurisdiction: we only judge what an AI/pipeline produced. The owner's verbatim text, external originals (_originals), his own edits — out of the gate's scope (we do not "clean the life out of" his voice). _drafts, _originals, 08-Templates, the inbox = intentional orphans (do not apply P6).
3. Deterministic checks (Grep/PowerShell, 0 tokens) — P4, P5, P6, P7, P9, P17, P18.
4. LLM checks for the remaining principles — on the excerpt only, never on the whole corpus.
5. The verdict block (format below). Fix nothing.

Principles (signal → principle → check)

Weights: ⛔ = hard stop (fail) · ⚠️ = warning (2+ → manual). Classes in brackets.

Provenance

• P1 ⛔ (A,C) Don't attribute someone else's work to the owner. Signal: "ALWAYS, if you found that the source of the knowledge is NOT ME … don't hang on me what I DIDN'T WRITE" (2026-06-08). Check: origin: anton / #anton-original present while there are signs of an external or AI author (other people's self-IDs, cross-handles, AI distillation) → FAIL.
• P2 ⛔ (A,C) AI co-authorship = mixed + the specific AI. Signal: "if the author is an AI, ALWAYS write WHICH AI exactly" (2026-06-09). Check: authored_by: ai|hybrid requires ai_author: with a specific model plus origin: mixed|external; a generic "ai" or origin: anton → FAIL. ⭐ Carve-out (the owner, 2026-07-02): on reglament-*/protocol-* the pair origin: anton + authored_by: hybrid is legal — there origin means "whose rule it is" (the Bible's authority) and authored_by means "who wrote it up"; ai_author is still mandatory and we don't add the owner's original tag.
• P3 ⚠️ (A,C) Both provenance axes present. Signal: the two-axis mandate (vault-conventions, operating-agreement). Check: frontmatter has BOTH origin: and authored_by:; missing → warn (we don't fail legacy notes).
• P4 ⛔ (all) No secrets. Signal: passwords live in secrets\, not in the vault or the always-loaded layer (credential-store; precedent: an account's 2FA code dictated into a chat → moved to the store). Check: grep password|2fa|api[_-]?key|token|secret[_-]?key (add the keyword variants of the operator's own language) + credential-shaped strings → a hit on a real secret = FAIL.

Graph and structure

• P5 ⛔ (A,C — new files) Filename = a latin kebab slug. Signal: "no Cyrillic in filenames ANYWHERE" (2026-06-08). Check: the basename is ASCII-only; Cyrillic → FAIL (the ~3.9k older ones are frozen legacy and are not failed — new files only).
• P6 ⛔ (A,C — new notes in the live vault) Not an orphan: ≥1 incoming link. Signal: "ALWAYS … finish the cross-linking" (2026-06-13). Check: Grep "\[\[<basename>\]\]|\[\[<basename>\|" path=$OBSIDIAN_VAULT ≥1 outside the note itself; 0 → FAIL ("an unfinished tail, not done"). Exceptions — jurisdiction item 2.
• P7 ⚠️ (A,C) Wikilinks resolve. Signal: "0 broken links before staging→vault" (vault-conventions). Check: every [[target]] exists as a file/alias; broken ones → warn.
• P8 ⚠️ (A) The prefix matches the folder. Signal: prefix = the note's class (vault-conventions). Check: concept-*→06-Concepts, person-*→07-People, insight-*→03-Insights, reglament-*→Protocols/Operations; a mismatch → warn.

Text and voice

• P9 ⏸ DISABLED (the owner, 2026-07-02: "I don't care about dashes, forget it for now"). The gate does NOT judge dashes and they do not affect the verdict; we run no cleanup campaigns over old text. The style rule for MY NEW text stands ([[no-long-dashes]]) — I simply write without them, for free. Only the owner can restore its weight.
• P10 ⛔ (C) The owner's text is verbatim, not "improved". Signal: "we transliterate NAMES, we never translate the owner's text" + the dedup supersede policy. Check: compare the merge against the sources; a translation/paraphrase/smoothing of his wording → FAIL ("a bit crooked means a bit alive").
• P11 ⚠️ (A,D) No filler. Signal: "dry, to the point, NO filler" (recurring across 15 sessions, digest 2026-06-23) + write-service-files-tight. Check (LLM): warm-up paragraphs, repetitions, rhetoric → warn.
• P12 ⚠️ (A,D — for the owner) His own language, English technical terms are fine. Signal: "in Russian" in 15/15 sessions. Check: an internal document for him written entirely in English → warn.
• P13 ⚠️ (D) No all-caps aggression. Signal: the 2026-07-01 measurement (all-caps / "MUST / NEVER" = overtriggering at 4.6+; de-capsing CLAUDE.md was approved by the owner). Check: all-caps imperatives in service text → warn.

Data and honesty

• P14 ⛔ (all) No invented data. Signal: "better not to make things up — use the data and clients we actually have". Check (LLM): figures/facts/examples with no source, presented as real → FAIL.
• P15 ⚠️ (A,D — proposals) BEFORE→AFTER on real data. Signal: the standing show-before-after rule. Check: a proposal draft with no before→after on the owner's own data → warn.
• P16 ⚠️ (A,D — proposals) Added complexity is flagged. Signal: the standing AK-47 rule (⚠️ ADDED COMPLEXITY + which pain it treats). Check: a new dependency/service/abstraction with no flag → warn.

Outbound on the owner's behalf

• P17 ⛔ (B — outreach/pitch) The greeting is its own line. Signal: an explicit standard from 2026-06-22 (50 pitches reformatted). Check: line 1 = the greeting only, line 2 empty, the body starts on line 3; merged together → FAIL.
• P18 ⛔ (B — unreviewed) ≤7 words + the channel's tone. Signal: "write it yourself" → target ≤5, ceiling 7 words; Facebook = a joke, Telegram = something meaningful after a RECALL on the person (2026-06-29). Check: word count >7 → FAIL; tone wrong for the channel → FAIL.
• P19 ⚠️ (B) The lead's language. Signal: "write to Russian-speaking leads in Russian and to English-speaking ones in English". Check: the draft's language vs the lead's language; a mismatch → warn, an unknown lead → manual.

Dedup merges

• P20 ⛔ (C) Supersede, never delete. Signal: the 2026-06-08 dedup policy (14 rules merged by supersede) + "never glob-delete". Check: the merge references the absorbed note (supersedes: / a link), and no source was deleted (alive or in _originals); otherwise → FAIL.

AI documents (class E) — the anti-drift lens

Source: the Anti-Drift Review by Artyom Arsyonov (Ray lab, looi.ru/a/anti-drift-review — huge thanks 🙏; verbatim copy: _originals\arsenov-looi\full\13-anti-drift-review.md). Added 2026-07-03 on the owner's "++". The core of his observation: "AI makes a document smoother before it makes it truer" — the argument then happens about the text instead of about reality. All the checks below are LLM judgement on an excerpt.

• P21 ⛔ (E) Claims with no support. Beautiful phrases not backed by data or a source. Check: for every key thesis → is there a link/data/provenance? An unsupported thesis that a recommendation rests on → FAIL. (Related to P14, but this is about argumentation, not only figures.)
• P22 ⚠️ (E) Hidden assumptions. A recommendation rests on an "if" nobody noticed. Check: write out the implicit "if X, then…"; ≥2 undeclared critical assumptions → warn (list them in the verdict).
• P23 ⚠️ (E) Weak link to the buyer / the benefit. The pain is named, but it is not visible who pays for it with time or money. Check: for product/GTM documents — is a concrete ICP/payer named; no → warn.
• P24 ⚠️ (E) Implementation failure. The plan sounds right but does not survive contact with our actual stack/team/system. Check: the plan's steps against the known state (machines, routines, people, limits); an unfeasible step → warn + name it.
• P25 ⛔ (E) The conclusion is too strong. The evidence says "maybe", the document says "we must do it". Check: the strength of the conclusion vs the strength of the evidence; an overclaim in the final recommendation → FAIL.
• The verdict for class E additionally carries a drift map: 🔴 (don't rely on it) / 🟡 (needs work: what to prove before deciding) / 🟢 (usable) + the list of unsupported claims.

Verdict format

🎛 taste-check: <file/draft> · class <A/B/C/D>
Verdict: ✅ PASS | ❌ FAIL | ⚠️ MANUAL REVIEW
Violated: P9 (3 long dashes, lines 12/40/41), P6 (0 incoming links)
Clean: P1-P5, P7, P8, P14
Not checked: P17-P19 (not outbound)
→ Action: <return to the author / show with a flag / safe to show>

Boundaries

• Read-only: no Write/Edit into the vault, no "I already fixed it along the way".
• Don't invent the owner's principles: a new principle is added only from a real signal of his (an edit/quote/digest rule_scan) with a date. Signal feedstock: $IMPORTS_ROOT/rule_candidates/digest-*.md (preference-sweep) + memory entries with type: feedback.
• Only the owner calibrates the weights (same as in the source review: the agent collects signals, the human calibrates the patterns).
• The token law: deterministic checks first (grep/counting), the LLM only for judgement on an excerpt.

Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.