ai-laziness-traps — T1-T21 trap catalogue
The canonical AI-laziness trap catalogue rule: 21 named traps with trigger/tempted-output/counter, kickoff-author citation obligations, and the anti-patterns that review rejects.
Census id: F3 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin
b069c59328aa8e08671d3a19fd25efa618dfc1c5(stagingFETCH_HEAD, read 2026-09-11; re-anchored from census pina1337cb301). Everyfile:linebelow is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.
| Field | Value |
|---|---|
| Status | raw reference (census row F3: MISSING → drafted E4) |
| Ships-to | clone only (setup.d/20-agents.sh:44 «(rules/ is not shipped)») |
| Fires-at | «any R-phase, audit, sample-based investigation, or open-ended AI task.» (.claude/rules/ai-laziness-traps.md:16) — path-scoped inject (5 globs) |
What it bans
The producing of «the shape of the expected output without performing the underlying work» — AI-laziness failure modes, formalised as a project invariant: .claude/rules/ai-laziness-traps.md:34 «AI agents under time pressure (their own session quota, perceived user impatience, or absence of an external done-check) reliably converge to path of least resistance: producing the shape of the expected output without performing the underlying work.» The general law it states: .claude/rules/ai-laziness-traps.md:36 «any AI task whose stopping criterion is «when I'm done» rather than «when an external check passes» will under-shoot the actual stopping point.» The fix: .claude/rules/ai-laziness-traps.md:38 «The fix is operationalising the laziness traps and forcing kickoff authors to instantiate them, not just reference them generically.»
Scope: .claude/rules/ai-laziness-traps.md:28 «The rule applies to any AI session running an R-phase, audit, sample-based investigation, decision-register exercise, doc-creation / doc-revision of a discipline-bearing artefact … or other open-ended task whose output is bounded only by «when do I stop». It does NOT apply to mechanical edits with explicit done-criteria». Header: .claude/rules/ai-laziness-traps.md:10 «# AI laziness traps — discipline rule»; §2 is the catalogue: .claude/rules/ai-laziness-traps.md:40 «## §2 Canonical trap catalogue» — T1-T21, each with trigger / tempted output / counter (T1 at :44-:48, T21 at :167-:181).
Never (fires)
A kickoff that cites the catalogue without instantiating it. The §3 obligations require citation + enumeration + a domain trap: .claude/rules/ai-laziness-traps.md:188 «2. Enumerate which T-numbers apply to its specific R-phase by listing them: Active traps for this R-phase: T1, T3, T4, T7, T11, T13, T15, T20.» — and the firing shape is named as anti-pattern: .claude/rules/ai-laziness-traps.md:195 «- #trap-catalogue-blanket-reference — kickoff cites this rule but does not enumerate T-numbers or add domain-specific traps. Treats the catalogue as decoration, not discipline. Counter: review-phase rejection.» The rule closes the escape hatch itself: .claude/rules/ai-laziness-traps.md:191 «Blanket reference («see ai-laziness-traps.md» without enumeration or domain extension) is insufficient and itself a form of T7 (pattern-matching the prompt instead of reasoning against it).»
Always (clean)
A kickoff that enumerates its active T-numbers and extends the catalogue with ≥1 domain trap (.claude/rules/ai-laziness-traps.md:189 «3. Add ≥1 domain-specific trap that is NOT in the canonical catalogue but is structurally possible in this R-phase's context.»), and findings that carry verification — the T3 counter as the canonical clean shape: .claude/rules/ai-laziness-traps.md:60 «Counter: every finding must have ONE of: (a) command + output, (b) file:line citation + the line's actual content, (c) explicit INCONCLUSIVE-needs-LLM or INCONCLUSIVE-needs-human if you cannot verify mechanically. No prose-only findings.»
Enforcement channels
- Rule-index row:
.claude/rules/00-rule-index.md:11«|ai-laziness-traps.md| A | any R-phase, audit, sample-based investigation, or open-ended AI task. | paths:(5), edit-time inject |» - Edit-time inject, 5 path globs:
.claude/rules/ai-laziness-traps.md:12«» with the inject text at:13«» - Companion principle test (CI, shipped):
.claude/rules/ai-laziness-traps.md:15«> Class: A — companion principle test shipped at packages/core/principles/12-ai-laziness-traps.test.ts (#74, 2026-05-17).» — a mechanical check on kickoff files for the §3 citation + T-enumeration syntax (.claude/rules/ai-laziness-traps.md:207). - Not backend-rendered — no FF diagnostic (FF7001/FF7002,
packages/core/diagnostics/registry.ts:321/:329) applies; the honest status is: edit-time inject + CI principle test + review-time rejection.
ai-laziness-digest — resident hot digest
The always-on resident digest of the AI-laziness trap catalogue: exact-quote T-number counters, rollback trigger, and the anti-drift principle test that pins it to the catalogue.
attention-is-not-a-mechanism — hope-as-gate prohibition
Class C prose rule: a load-bearing check must be a deterministic gate or a named cold-agent protocol — bare human/AI attention may hold merge authority but is never the detection layer.