[lukemaj]

What: triggers.md no longer credits karpathy-wiki's SKIP and anti-rationalization shape to superpowers. It now also covers when a description is not enough, how to word routing rows, and how to test a trigger. A new page, docs/06-testing/trigger-benchmarks.md, gives the full trigger-benchmark method.
Why: The repo's description reference got its main shape's source wrong, misstated length and over-firing, and had no valid method for testing a trigger. AgentsMD has since measured all of this.
So what: A reviewer checks the exact head, then the user merges.

Closes #5
Refs #7

Changes

• docs/05-authoring/triggers.md: four corrections; new sections "When a description is not enough", "Routing rows" and "Test a trigger" (replaces the pressure-test loop); sources pinned.
• docs/06-testing/trigger-benchmarks.md (new): decision, fixture, cases and moments, confinement, instrument validity, interleaved arms, records, instrument checks. Every rule cites the case study or an AgentsMD PR or results file.
• docs/06-testing/unit-tests.md: trigger benchmarks are a fourth kind of test. Superpowers removed tests/skill-triggering/ in 8611a4e, and the page now says so. The "no CI" sentence is updated because Make the doc checks run: link check, llms.txt freshness and CI #10 merged: CI runs npm test, not the validator.
• docs/06-testing/benchmark-integrity.md: one sentence linking the new page.
• Navigation: docs.json, README Testing list, the loader Skill's map (new entry; the triggers entry was also updated because it described the old two-shape page). GLOSSARY.md gains Arm. public/llms.txt regenerated.

Superpowers lines re-checked (obra/superpowers 8ca22db, superpowers-evals e64684c)

Claim	Line	Result
No SKIP / anti-rationalization in superpowers	repo-wide grep for "SKIP when", "Do NOT skip", "forbidden rationali"	0 hits
Before-clause anchor	skills/writing-skills/SKILL.md:171, :546-552	confirmed
Under 500 characters	:103 "Keep under 500 characters if possible"	confirmed
Description lengths	15 skills/*/SKILL.md descriptions, parsed	79-375, not 92-392 as the Issue says; page uses the measured range
"an agent may follow"	:152-154	confirmed
Section renamed to Skill Discovery Optimization	:140, RELEASE-NOTES.md:264	confirmed
One reported case	:154-156	confirmed (single code-review example)
Bootstrap	AGENTS.md:76	confirmed
No-guidance control, 5+ reps	:582-583	confirmed
skill-not-called, vacuous-pass -before verbs	evals docs/scenario-authoring.md:520-524, :660-666	confirmed
Checkbox over-trigger ~35-40% on Opus 5	evals docs/experiments/2026-09-02-opus5-signature.md:311-320	2/5 on main, 7/20 earlier sweep; Opus 4.8 0/51
tests/skill-triggering/	superpowers 8611a4e	removed, "covered by drill triggering-* scenarios"

AgentsMD numbers and sources (records at 06b7af6)

Number	Source
Opus 0/40 naive, 15/40 negatives loaded operations	docs/work/164-trigger-audit/results.md
Pointer 3/20 to 20/20, 417 characters	docs/work/174-routing-injection/results.md
Verification 0/5 to 5/5 with the before-clause	case study; #181 midtask.jsonl, #182 --summary
Bare conjunct: SKILL-MECHANICS.md 15/15, prose.md 4/15 (attempts), then 5/5	recomputed from 174-routing-injection/records-prefix.jsonl (Opus, arms A/B/C) and records.jsonl (B2, B3)
Diagnosis 8/10 to 5/5; fix1 verification 3/5	182-routing-sweep/records.jsonl --summary, regenerated
OpenCode question negative 0/5 to 5/5 clean	case study (#182 results)
349 tests, 17/20 with 10 no-edit, 9/10 escaped runs, 12 verdicts rescored	case study

The Issue's "prose.md 7/22 against 22/22" comes only from the non-public #174 analysis quoted in #6. The public records do not reproduce it, so the page uses the recomputable 4/15 against 15/15 and marks those runs as read attempts.

Elon record

• Requirements and who asked: the user (2026-09-29, Clean up the repository and apply the routing and benchmark learnings #7 and Correct triggers.md and add the trigger-benchmark method #5): correct the four superpowers findings, add routing-row guidance and the harness pitfalls from Case study: the AgentsMD routing benchmarks (toolboxmd/agentsmd#164, #174, #182) #6, and add a reusable trigger-benchmark method.
• Deleted: a separate routing-rows page (it lives in triggers.md); retelling the case study inside the method page (it links the case study); the Issue's 7/22 figure, which is not reproducible from public records.
• Bottleneck: the description reference misattributed its main shape and had no valid method for testing a trigger.
• Checked myself: re-opened every cited superpowers and superpowers-evals line at the pinned commits; recomputed description lengths and the bare-conjunct and sweep numbers from AgentsMD records extracted with git archive 06b7af6; ran npm test.

Proof

• npm ci && npm test: exit 0 (links, llms.txt freshness and coverage, absolute paths, bare-path gate).
• npm run build:llms regenerated public/llms.txt.
• Independent review: Luna (gpt-6-luna, max) REQUEST_CHANGES on 5d7e16f (qualify the #174 3/20 baseline), fixed in ccaa79f, then APPROVE on ccaa79f; review/independent is success.

🤖 Generated with Claude Code

[lukemaj]

REQUEST_CHANGES

• P2 - Qualify the #174 baseline. docs/05-authoring/triggers.md:79 says the pointer took naive prompts from 3/20 to 20/20. The cited #174 results later explain that two base-arm test-design runs read the required file through a chained cat that exited non-zero; one read was before the first edit, so the corrected baseline may be 5/20. Please identify 3/20 as the originally scored value and include that caveat (or report the recomputed value).

npm test passes in the specified worktree. The remaining checked acceptance criteria, changed-file scope, internal links, glossary terms, and textual constraints pass.

[lukemaj]

APPROVE

The #174 baseline wording at docs/05-authoring/triggers.md:79 is fixed: 3/20 is explicitly the scored baseline, and the paragraph notes the possible higher count from the chained-cat reads with a link to the case-study caveat. I rechecked the exact head ccaa79f928064eeb05ec6610ad79c5e4dc8fa93c; I found no remaining actionable issues in the reviewed scope.

npm test passes in the specified worktree with Node 22.23.1. This repository's Mintlify link check does not support the environment's default Node 26.10.0.