automotive-skills-suite: ISO 26262/21434 Deliverables as Paired Builder + Reviewer Skills

Most safety and quality deliverables in automotive programs live as xlsx workbooks, hand-edited across handoffs. jherrodthomas/automotive-skills-suite turns each phase into a builder plus a matching reviewer, and chains them so a change upstream propagates downstream without re-keying.

Why This Skill Matters

The README describes the suite as "76 builder skills + 76 matching confirmation reviewer skills" totaling 152 installable Claude skills (note: the 100+ figure that appears in skills.json for this repo is an older description snapshot — the README itself shows the current 152). Every artifact-producing skill is paired with a confirmation reviewer whose output is a visual dashboard with KPI tiles, charts, and findings tables.

The repository has 2.4k stars on GitHub and is MIT-licensed. The repo README explicitly says it was "Built with Claude" — the suite targets Claude Desktop / Cowork with the "Save skill" install path; it is not framed as a portable Agent Skills manifest.

The standards covered are wide: ISO 26262:2018 (functional safety), ISO/SAE 21434:2021 (cybersecurity, mandated by UN R155), ISO 21448:2022 (SOTIF for ADAS/AV), IATF 16949:2016 / AIAG-VDA (APQP, FMEA, Control Plan, PPAP, MSA, SPC), Automotive SPICE PAM 3.1 / 4.0, AUTOSAR Classic and Adaptive R22-11, ISO 14229-1 (UDS), SAE J2012 (DTCs), ISO 22901-1 (ODX), and the communication-protocol stack (CAN, LIN, FlexRay, Automotive Ethernet). The full chain diagram in the README spells out how HARA feeds FSC feeds TSC feeds HW/SW reqs feeds FMEDA feeds the Safety Case.

The "chain" is the moat: every downstream skill consumes the upstream skill's xlsx as a stable file-format contract. A change to the HARA automatically propagates to FSC, TSC, FMEDA, and the safety case.

Installation

The repo provides each skill as a standalone .skill file under skills/, with a triggering description in its frontmatter so the agent picks it up by phrasing. The README's quickstart is:

Example: install the HARA builder, then ask Claude — "Build a HARA for a new ECU project — Electronic Stability Control"

In practice that means:

1. Download the .skill file you need from skills/ (or the full bundle from Releases).
2. In Cowork or Claude Desktop, click "Save skill" to install it.
3. Trigger it by phrasing — every skill declares its triggering description in its frontmatter, so the agent auto-picks.

Because this is a Claude-targeted bundle, you will not find an npx skills add … install command or a portable SKILL.md manifest. The skills ship as .skill files optimized for Claude Desktop / Cowork.

Real Workflow: Build a HARA, Then Review It

You have a new ECU project for an Electronic Stability Control function. The chain expects a HARA at the top.

Step 1. Install hara-builder and its reviewer from skills/. Each skill is a Claude-targeted .skill file; the repo's reviewers follow the X-checklist-reviewer naming pattern (e.g. hara-checklist-reviewer).

Step 2. Trigger the builder by phrasing:

Build a HARA for a new ECU project — Electronic Stability Control.

Step 3. The builder emits an xlsx workbook with: item definition, hazard list (14 malfunction guide words plus Cartesian combinations of function × malfunction × environment), severity/exposure/controllability scoring, ASIL determination, and the safety goals rolled up for downstream consumption.

Step 4. Run the matching reviewer — same project context — to produce a confirmation dashboard. Reviewer outputs include visual KPI tiles, a compliance bar against ASIL thresholds, a stacked breakdown by section, and a findings table with recommended actions.

Step 5. Hand the xlsx to the next pair (fsc-builder + reviewer, then tsc-builder + reviewer, then the HW/SW lanes), which read it as the stable contract. A change to a hazard upstream is picked up automatically — no re-keying.

Real Workflow: Diagnose a Why-Did-ASIL-Move Conversation

A safety auditor asks why your FMEDA shows ASIL C for a HW element that the HARA marked ASIL B. You need to trace the decomposition.

Step 1. The fmeda-builder produces SPFM / LFM / PMHF and auto-verifies against per-ASIL targets. Open that xlsx and find the element's row.

Step 2. Run the matching reviewer skill (fmeda-checklist-reviewer) — its dashboard lists every element that falls short of the per-ASIL threshold, plus the residual risk classification.

Step 3. To trace upstream, follow the chain: tsc-builder consumes the FSC; the FSC consumes the HARA's safety goals; the HARA's ASIL came from the Cartesian (function × malfunction × environment) scoring. The chain diagram in the README (HARA → FSC → TSC → HW-SR / SW-SR → FMEDA → Safety Case) maps the file you are reading back to the input that set the ASIL.

Step 4. If the ASIL decomposition is right, the FMEDA threshold check is the system catching your hand-entered SPFM numbers; re-run the builder with corrected diagnostic coverage.

Step 5. If the ASIL decomposition is wrong, the fix lives in the HARA — re-score the malfunction, regenerate the xlsx, and let the downstream chain re-derive.

Tips

• Read the README's chain diagram once; it tells you which pair goes next without reading the reference docs.
• Always run the matching reviewer right after a builder — the chain assumes builder xlsx has been confirmed before downstream consumption.
• Many reviewers auto-verify quantitative thresholds (FMEDA SPFM/LFM/PMHF vs ASIL; PPAP Cpk vs 1.33/1.67; MSA Gauge R&R % vs AIAG; SPC Cpk vs target). Let the dashboard flag the gap; the recommended-actions column is the to-do.
• Reviewers never modify source artifacts — gaps land in "Recommended Actions" instead. Do not edit a builder's xlsx by hand to silence a reviewer; rerun the builder with corrected inputs.

When Not to Use This

If your team does not use Claude Desktop or Cowork, this suite is the wrong shape — it is not packaged as a portable Agent Skills manifest and has no npx install path. For a portable multi-agent list of automotive-relevant skills (smaller scope), look elsewhere; this suite is Claude-only and runs only against Claude tooling. If you are looking for a single safety-analysis skill rather than a 152-skill chain, this suite is overkill.

See the leaderboard for more skills.