diff --git a/README.md b/README.md
index dbd102d..f95adbb 100644
--- a/README.md
+++ b/README.md
@@ -24,6 +24,12 @@ cp -R agent-skills/<skill-name> ~/.codex/skills/
 
 Replace `<skill-name>` with a skill directory name. For other agents, copy it to that agent's Skills directory.
 
+## Skills
+
+| Skill category | Description |
+| --- | --- |
+| [Zephyr agent skills](skills/zephyr/README.md) | Reusable skills for Zephyr development, debugging, exploration, patch review, and safety evidence. |
+
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=processmission/agent-skills&type=Date)](https://www.star-history.com/#processmission/agent-skills&Date)
diff --git a/skills/zephyr/README.md b/skills/zephyr/README.md
new file mode 100644
index 0000000..3e7c89b
--- /dev/null
+++ b/skills/zephyr/README.md
@@ -0,0 +1,72 @@
+# Zephyr Agent Skills
+
+Agent skills for [Zephyr Project](https://www.zephyrproject.org/) development,
+debugging, exploration, patch review, and safety evidence work.
+
+Source repository: <https://github.com/processmission/agent-skills>
+
+## Install
+
+From the root of the Zephyr repository, run:
+
+```sh
+curl -fsSL https://raw.githubusercontent.com/processmission/agent-skills/main/skills/zephyr/install.sh | bash
+```
+
+By default, the installer adds every skill in the `skills/zephyr/` catalog to
+the current project for Codex and Claude Code. It also adds `.agents/`,
+`.claude/skills/`, `.zephyr-skills/`, and `skills-lock.json` to the
+repository's local Git exclude file, so installation and agent task artifacts
+do not dirty the worktree. It does not modify the repository's `.gitignore`.
+
+To install from a local checkout, or to select an individual skill:
+
+```sh
+./skills/zephyr/install.sh --target /path/to/zephyr
+./skills/zephyr/install.sh --target /path/to/zephyr --skill zephyr-debug
+```
+
+## Repository workflow
+
+Each Zephyr skill is self-contained, can be installed independently, and does
+not require another Zephyr skill. Every skill includes the same compact audit
+workflow. For non-trivial tasks that write to the workspace, agent plans, audit
+records, command journals, decisive logs, and other temporary evidence go under
+`.zephyr-skills/<agent-task>/`. Put generated intermediate debugging documents,
+debug scripts, and Bash or Python scaffolding and harness scripts in
+`.zephyr-skills/<agent-task>/scripts/`, never in a build directory. Put
+non-Zephyr third-party source or libraries acquired solely for the agent's task,
+and temporary binaries produced by agent-only probes or tools, in
+`.zephyr-skills/<agent-task>/output/`. Do not copy normal Zephyr build artifacts
+there. Keep Zephyr-native build trees, binaries, and routine test outputs in a
+path chosen by the user or the relevant Zephyr tool's default. In a Git
+worktree, add `/.zephyr-skills/` to the repository-local exclude file returned
+by `git rev-parse --git-path info/exclude` before writing audit artifacts, and
+verify that `git status --short` contains no `.zephyr-skills/` paths.
+
+## Skills
+
+### `zephyr-debug`
+
+Reproduce and isolate Zephyr build, toolchain, QEMU, boot, board, or runtime
+failures. Use targeted Zephyr logging, QEMU GDB stubs, or OpenOCD and J-Link
+remote debugging according to the target and available tools.
+
+### `zephyr-explore`
+
+Investigate Zephyr source, branch history, external precedent, and design
+constraints before making changes. Use it to clarify current behavior,
+maintainer boundaries, and compatibility requirements.
+
+### `zephyr-patch-review`
+
+Review Zephyr diffs or patch stacks for correctness, integration impact, test
+coverage, reviewer boundaries, and merge readiness. Report only actionable,
+source-backed findings.
+
+### `zephyr-safety-evidence`
+
+Plan or review Zephyr safety evidence, requirements traceability, and
+certification gaps for ASIL, ISO 26262, IEC 61508, SIL, MC/DC, HARA,
+FMEA/FMEDA, and tool qualification work. It never presents draft or partial
+evidence as completed certification.
diff --git a/skills/zephyr/install.sh b/skills/zephyr/install.sh
new file mode 100755
index 0000000..776698c
--- /dev/null
+++ b/skills/zephyr/install.sh
@@ -0,0 +1,206 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+readonly DEFAULT_SKILL_SOURCE="https://github.com/processmission/agent-skills/tree/main/skills/zephyr"
+readonly REQUIRED_EXCLUDES=(
+    ".agents/"
+    ".claude/skills/"
+    ".zephyr-skills/"
+    "skills-lock.json"
+)
+
+usage() {
+    cat <<'EOF'
+Install the Zephyr agent skills into a Git repository.
+
+Usage:
+  install.sh [install] [--target DIR] [skills options...]
+
+Options:
+  --target DIR          Install into DIR instead of the current directory.
+  -s, --skill NAME      Install one skill instead of every Zephyr skill.
+  -a, --agent NAME...   Override the default agents (codex and claude-code).
+  -y, --yes             Skip confirmation prompts.
+  -h, --help            Show this help.
+
+Environment:
+  ZEPHYR_SKILLS_SOURCE  Override the skills source passed to `npx skills add`.
+
+By default, the installer installs every skill in the Zephyr catalog for Codex
+and Claude Code. It also adds .agents/, .claude/skills/, .zephyr-skills/, and
+skills-lock.json to the repository's local Git exclude file.
+EOF
+}
+
+die() {
+    printf 'zephyr-skills: %s\n' "$*" >&2
+    exit 1
+}
+
+require_command() {
+    command -v "$1" >/dev/null 2>&1 ||
+        die "required command not found: $1"
+}
+
+resolve_skill_source() {
+    if [[ -n "${ZEPHYR_SKILLS_SOURCE:-}" ]]; then
+        printf '%s\n' "$ZEPHYR_SKILLS_SOURCE"
+        return
+    fi
+
+    local script_path=""
+    local script_dir=""
+
+    if [[ -n "${BASH_SOURCE[0]:-}" && -f "${BASH_SOURCE[0]}" ]]; then
+        script_path="${BASH_SOURCE[0]}"
+        script_dir="$(cd "$(dirname "$script_path")" && pwd)"
+    fi
+
+    if [[ -n "$script_dir" &&
+          -f "$script_dir/README.md" &&
+          -f "$script_dir/zephyr-debug/SKILL.md" ]]; then
+        printf '%s\n' "$script_dir"
+        return
+    fi
+
+    printf '%s\n' "$DEFAULT_SKILL_SOURCE"
+}
+
+resolve_target_root() {
+    local requested_target="$1"
+    local target_dir
+    local git_root
+
+    target_dir="$(cd "$requested_target" 2>/dev/null && pwd -P)" ||
+        die "target directory does not exist: $requested_target"
+    git_root="$(git -C "$target_dir" rev-parse --show-toplevel 2>/dev/null)" ||
+        die "target directory is not inside a Git repository: $target_dir"
+
+    [[ "$target_dir" == "$git_root" ]] ||
+        die "run from the repository root or pass --target $git_root"
+
+    printf '%s\n' "$git_root"
+}
+
+add_local_exclude() {
+    local target_root="$1"
+    local entry="$2"
+    local exclude_file
+    local bare_entry
+    local escaped_entry
+    local probe_path
+
+    exclude_file="$(git -C "$target_root" rev-parse --git-path info/exclude)"
+    if [[ "$exclude_file" != /* ]]; then
+        exclude_file="$target_root/$exclude_file"
+    fi
+
+    mkdir -p "$(dirname "$exclude_file")"
+    touch "$exclude_file"
+
+    bare_entry="${entry%/}"
+    escaped_entry="$(printf '%s' "$bare_entry" | sed 's/[][\.^$*+?{}|()]/\\&/g')"
+
+    if ! grep -Eq "^/?${escaped_entry}/?$" "$exclude_file"; then
+        printf '/%s\n' "$entry" >>"$exclude_file"
+    fi
+
+    if [[ "$entry" == */ ]]; then
+        probe_path="${entry}.zephyr-skills-install-probe"
+    else
+        probe_path="$entry"
+    fi
+
+    git -C "$target_root" check-ignore --no-index -q "$probe_path" ||
+        die "failed to configure local Git exclude for $entry"
+}
+
+main() {
+    local target_dir="."
+    local skill_source
+    local target_root
+    local saw_skill=0
+    local saw_agent=0
+    local saw_yes=0
+    local arg
+    local -a forwarded_args=()
+
+    if [[ "${1:-}" == "install" ]]; then
+        shift
+    fi
+
+    while [[ "$#" -gt 0 ]]; do
+        arg="$1"
+        case "$arg" in
+            --target)
+                [[ "$#" -ge 2 ]] || die "--target requires a directory"
+                target_dir="$2"
+                shift 2
+                ;;
+            -s|--skill)
+                saw_skill=1
+                forwarded_args+=("$arg")
+                shift
+                ;;
+            -a|--agent)
+                saw_agent=1
+                forwarded_args+=("$arg")
+                shift
+                ;;
+            -y|--yes)
+                saw_yes=1
+                forwarded_args+=("$arg")
+                shift
+                ;;
+            -h|--help)
+                usage
+                exit 0
+                ;;
+            -g|--global)
+                die "global installation is not supported; run from a project root"
+                ;;
+            -l|--list|--all)
+                die "$arg is not an install option for this script"
+                ;;
+            *)
+                forwarded_args+=("$arg")
+                shift
+                ;;
+        esac
+    done
+
+    require_command git
+    require_command grep
+    require_command npx
+    require_command sed
+
+    skill_source="$(resolve_skill_source)"
+    target_root="$(resolve_target_root "$target_dir")"
+
+    if [[ "$saw_skill" -eq 0 ]]; then
+        forwarded_args+=(--skill '*')
+    fi
+    if [[ "$saw_agent" -eq 0 ]]; then
+        forwarded_args+=(--agent codex claude-code)
+    fi
+    if [[ "$saw_yes" -eq 0 ]]; then
+        forwarded_args+=(-y)
+    fi
+
+    printf 'Installing Zephyr agent skills into %s\n' "$target_root"
+    (
+        cd "$target_root"
+        npx --yes skills add "$skill_source" "${forwarded_args[@]}"
+    )
+
+    for arg in "${REQUIRED_EXCLUDES[@]}"; do
+        add_local_exclude "$target_root" "$arg"
+    done
+
+    printf 'Installed Zephyr agent skills.\n'
+    printf 'Configured repository-local Git excludes in %s.\n' \
+        "$(git -C "$target_root" rev-parse --git-path info/exclude)"
+}
+
+main "$@"
diff --git a/skills/zephyr/zephyr-debug/SKILL.md b/skills/zephyr/zephyr-debug/SKILL.md
new file mode 100644
index 0000000..3aeed17
--- /dev/null
+++ b/skills/zephyr/zephyr-debug/SKILL.md
@@ -0,0 +1,108 @@
+---
+# SPDX-FileCopyrightText: Copyright (c) 2026 Process Mission
+# SPDX-License-Identifier: MIT
+name: zephyr-debug
+description: >-
+  Use when reproducing or isolating the root cause of a Zephyr build,
+  toolchain, QEMU, boot, board, or runtime failure.
+---
+
+# Zephyr Debug
+
+## Audit workflow
+
+For every non-trivial task that writes to the workspace, choose a stable task
+slug and keep agent-only records under:
+
+```text
+.zephyr-skills/<agent-task>/
+├── audit.md        # Baseline, decisions, evidence, verification, and gaps
+├── commands.md     # Redacted commands, working directories, and results
+├── scripts/        # Intermediate debug documents, scripts, and harnesses
+├── output/         # Non-Zephyr dependencies and temporary binaries
+└── logs/           # Decisive build, test, runtime, or diagnostic logs
+```
+
+Put every generated intermediate debugging document, debug script, and
+scaffolding or harness script, including Bash and Python scripts, under
+`scripts/`, never in a build directory. Put non-Zephyr third-party source or
+libraries acquired solely for the agent's task, and temporary binaries produced
+by agent-only probes or tools, under `output/`. Do not copy normal Zephyr build
+artifacts there. Keep other agent plans and temporary evidence in the same task
+directory. In a Git worktree, add
+`/.zephyr-skills/` to the repository-local exclude file returned by
+`git rev-parse --git-path info/exclude` before writing audit artifacts. Preserve
+existing entries, avoid duplicates, and verify before handoff that
+`git status --short` contains no `.zephyr-skills/` paths. Keep Zephyr-native
+build trees, binaries, and test outputs separate: honor a user-specified path,
+otherwise use defaults such as configured `build.dir-fmt`, `build/`, or
+`twister-out/`. Record the effective paths, do not stage `.zephyr-skills/`
+unless requested, and report the task directory and unresolved gaps at handoff.
+
+## Diagnose
+
+1. Capture expected and observed behavior, the narrowest reproducer, and the
+   first decisive log.
+2. Exclude wrong revisions, boards, configuration, stale output, and generated
+   artifact or toolchain mismatches before changing source.
+3. Locate the first broken invariant across Kconfig, CMake, devicetree,
+   generated files, runner commands, boot flow, and runtime state.
+4. Test each hypothesis with the smallest reversible probe and distinguish
+   observations from inferences.
+5. For diagnosis-only requests, stop at the supported cause and remaining
+   uncertainty. Implement a fix and regression gate only when requested.
+
+## Choose a debugging path
+
+- Use Zephyr logging and targeted instrumentation when the failure remains
+  observable without stopping the system.
+- Use a QEMU GDB stub for reproducible emulated failures, early boot, exception
+  state, memory inspection, breakpoints, and instruction-level stepping.
+- Use the board's configured debug runner with OpenOCD or J-Link for hardware
+  failures. Select the runner from the board, probe, and user environment; do
+  not assume that one backend is universally available.
+- Combine logs with GDB when console history or timing explains state that a
+  stopped target cannot.
+
+## Run a remote GDB session
+
+1. Preserve the exact ELF and build configuration that reproduce the failure.
+   Use the matching toolchain GDB and keep symbols in the Zephyr build tree.
+2. Start the remote endpoint with the configured Zephyr runner: a QEMU GDB
+   stub for emulation, or a debug server backed by OpenOCD or J-Link for
+   hardware. Prefer runner-generated arguments over guessed probe, transport,
+   reset, or device settings.
+3. If installed, use Tmux for concurrent panes; otherwise use Zellij, then
+   separate terminals. Keep the debug server or QEMU in one pane, GDB in a
+   second, and optional console or log monitoring in a third.
+4. Generate a task-specific GDB command file under
+   `.zephyr-skills/<agent-task>/scripts/`, including the symbol file, remote
+   endpoint, breakpoints or watchpoints, thread inspection, and
+   evidence-producing commands. Avoid changing global GDB configuration.
+5. Connect with `target remote` or `target extended-remote` as required by the
+   server. Capture the server command, GDB command file, decisive backtraces,
+   register or memory state, and target logs in the audit directory.
+6. For a live hardware target, prefer attach semantics when flashing or reset
+   is not requested. Record any operation that changes target state.
+
+## Use Zephyr logging and instrumentation
+
+- Enable the Zephyr logging subsystem and the narrowest useful module or
+  subsystem log level through task-specific configuration or overlays.
+- Add `LOG_DBG`, `LOG_INF`, `LOG_WRN`, `LOG_ERR`, or hexdump probes at state
+  transitions, error paths, ownership changes, and relevant interrupt or
+  scheduling boundaries.
+- Include stable identifiers and state in messages, but never log secrets or
+  unbounded data. Keep probes minimal and easy to remove or gate.
+- Account for timing changes from added logging, synchronous output, transport
+  buffering, and log overflow. Recheck the failure with reduced instrumentation
+  before treating the result as causal.
+- Save decisive console output under `.zephyr-skills/<agent-task>/logs/` and
+  state which logging configuration and probes produced it.
+
+## Guardrails
+
+- Do not hide a failure by weakening assertions or tests.
+- Stop only QEMU processes started for the task; never kill broad process
+  groups.
+- Stop only debug servers and multiplexer sessions started for the task.
diff --git a/skills/zephyr/zephyr-explore/SKILL.md b/skills/zephyr/zephyr-explore/SKILL.md
new file mode 100644
index 0000000..ed91b29
--- /dev/null
+++ b/skills/zephyr/zephyr-explore/SKILL.md
@@ -0,0 +1,61 @@
+---
+# SPDX-FileCopyrightText: Copyright (c) 2026 Process Mission
+# SPDX-License-Identifier: MIT
+name: zephyr-explore
+description: >-
+  Use when Zephyr source, branch history, external precedent, or design
+  constraints must be investigated before a decision can be made.
+---
+
+# Zephyr Explore
+
+## Audit workflow
+
+For every non-trivial task that writes to the workspace, choose a stable task
+slug and keep agent-only records under:
+
+```text
+.zephyr-skills/<agent-task>/
+├── audit.md        # Baseline, decisions, evidence, verification, and gaps
+├── commands.md     # Redacted commands, working directories, and results
+├── scripts/        # Intermediate debug documents, scripts, and harnesses
+├── output/         # Non-Zephyr dependencies and temporary binaries
+└── logs/           # Decisive build, test, runtime, or diagnostic logs
+```
+
+Put every generated intermediate debugging document, debug script, and
+scaffolding or harness script, including Bash and Python scripts, under
+`scripts/`, never in a build directory. Put non-Zephyr third-party source or
+libraries acquired solely for the agent's task, and temporary binaries produced
+by agent-only probes or tools, under `output/`. Do not copy normal Zephyr build
+artifacts there. Keep other agent plans and temporary evidence in the same task
+directory. In a Git worktree, add
+`/.zephyr-skills/` to the repository-local exclude file returned by
+`git rev-parse --git-path info/exclude` before writing audit artifacts. Preserve
+existing entries, avoid duplicates, and verify before handoff that
+`git status --short` contains no `.zephyr-skills/` paths. Keep Zephyr-native
+build trees, binaries, and test outputs separate: honor a user-specified path,
+otherwise use defaults such as configured `build.dir-fmt`, `build/`, or
+`twister-out/`. Record the effective paths, do not stage `.zephyr-skills/`
+unless requested, and report the task directory and unresolved gaps at handoff.
+
+## Investigate
+
+1. Bound the question to paths, symbols, configuration, and behavior.
+2. Search current source, tests, documentation, and `MAINTAINERS.yml` before
+   using history or external material.
+3. For branch comparisons, use the user-specified target. Otherwise resolve the
+   configured upstream and use `git merge-base HEAD <upstream>`; report a
+   missing target or upstream instead of guessing.
+4. Use local history when current source does not explain intent or a
+   compatibility constraint.
+5. Consult primary external sources only when local evidence is insufficient.
+6. Stop when the evidence supports the next design, review, or verification
+   decision.
+
+## Report
+
+- Separate implemented, verified, documented, inferred, and unknown behavior.
+- Cite decisive paths, symbols, commits, or external revisions.
+- Treat external code as precedent unless import is explicitly in scope and
+  license-compatible.
diff --git a/skills/zephyr/zephyr-patch-review/SKILL.md b/skills/zephyr/zephyr-patch-review/SKILL.md
new file mode 100644
index 0000000..38c9a6e
--- /dev/null
+++ b/skills/zephyr/zephyr-patch-review/SKILL.md
@@ -0,0 +1,71 @@
+---
+# SPDX-FileCopyrightText: Copyright (c) 2026 Process Mission
+# SPDX-License-Identifier: MIT
+name: zephyr-patch-review
+description: >-
+  Use when reviewing Zephyr diffs or patch stacks for correctness,
+  integration, coverage, reviewer boundaries, or merge readiness.
+---
+
+# Zephyr Patch Review
+
+## Audit workflow
+
+For every non-trivial task that writes to the workspace, choose a stable task
+slug and keep agent-only records under:
+
+```text
+.zephyr-skills/<agent-task>/
+├── audit.md        # Baseline, decisions, evidence, verification, and gaps
+├── commands.md     # Redacted commands, working directories, and results
+├── scripts/        # Intermediate debug documents, scripts, and harnesses
+├── output/         # Non-Zephyr dependencies and temporary binaries
+└── logs/           # Decisive build, test, runtime, or diagnostic logs
+```
+
+Put every generated intermediate debugging document, debug script, and
+scaffolding or harness script, including Bash and Python scripts, under
+`scripts/`, never in a build directory. Put non-Zephyr third-party source or
+libraries acquired solely for the agent's task, and temporary binaries produced
+by agent-only probes or tools, under `output/`. Do not copy normal Zephyr build
+artifacts there. Keep other agent plans and temporary evidence in the same task
+directory. In a Git worktree, add
+`/.zephyr-skills/` to the repository-local exclude file returned by
+`git rev-parse --git-path info/exclude` before writing audit artifacts. Preserve
+existing entries, avoid duplicates, and verify before handoff that
+`git status --short` contains no `.zephyr-skills/` paths. Keep Zephyr-native
+build trees, binaries, and test outputs separate: honor a user-specified path,
+otherwise use defaults such as configured `build.dir-fmt`, `build/`, or
+`twister-out/`. Record the effective paths, do not stage `.zephyr-skills/`
+unless requested, and report the task directory and unresolved gaps at handoff.
+
+## Review
+
+1. Bound the review to an explicit base and intended behavior; separate
+   unrelated changes before judging the patch.
+2. Inspect correctness and semantic integration, including ownership, lifetime,
+   locking, error paths, generated artifacts, and 32/64-bit assumptions.
+3. Check the affected Kconfig, CMake, devicetree, architecture, board, SoC,
+   toolchain, documentation, and API-lifecycle boundaries.
+4. Verify that targeted tests and logs support each behavior claim and state the
+   remaining coverage gaps.
+5. Check patch order, reviewer-sized separation, bisectability, and applicable
+   SPDX and license conventions for new files.
+
+## Ownership and precedent
+
+- Use `MAINTAINERS.yml` for affected areas, reviewers, labels, and tests;
+  `CODEOWNERS` is not authoritative in this tree.
+- Consult local history for unfamiliar APIs, regressions, or subsystem boundary
+  changes. Prefer `git log --follow`, `git log -S`, `git blame`, and `git show`.
+- Query the current Gerrit change and patch-set state, or an upstream pull
+  request when explicitly in scope, only for merge readiness. Read current
+  submit, approval, and blocker policy instead of hard-coding it.
+- Treat unavailable history as a gap, not evidence of a defect.
+
+## Findings
+
+- Report only actionable, source-backed findings, with file and line references
+  when available.
+- Explain the failure mode, impact, and any missing or inadequate verification.
+- If there are no findings, say so and list residual verification gaps.
diff --git a/skills/zephyr/zephyr-safety-evidence/SKILL.md b/skills/zephyr/zephyr-safety-evidence/SKILL.md
new file mode 100644
index 0000000..f100978
--- /dev/null
+++ b/skills/zephyr/zephyr-safety-evidence/SKILL.md
@@ -0,0 +1,62 @@
+---
+# SPDX-FileCopyrightText: Copyright (c) 2026 Process Mission
+# SPDX-License-Identifier: MIT
+name: zephyr-safety-evidence
+description: >-
+  Use when planning or reviewing Zephyr safety evidence, safety requirements
+  traceability, certification gaps, or claims involving ASIL, ISO 26262,
+  IEC 61508, SIL, MC/DC, HARA, FMEA/FMEDA, tool qualification, or safety
+  cases.
+---
+
+# Zephyr Safety Evidence
+
+## Audit workflow
+
+For every non-trivial task that writes to the workspace, choose a stable task
+slug and keep agent-only records under:
+
+```text
+.zephyr-skills/<agent-task>/
+├── audit.md        # Baseline, decisions, evidence, verification, and gaps
+├── commands.md     # Redacted commands, working directories, and results
+├── scripts/        # Intermediate debug documents, scripts, and harnesses
+├── output/         # Non-Zephyr dependencies and temporary binaries
+└── logs/           # Decisive build, test, runtime, or diagnostic logs
+```
+
+Put every generated intermediate debugging document, debug script, and
+scaffolding or harness script, including Bash and Python scripts, under
+`scripts/`, never in a build directory. Put non-Zephyr third-party source or
+libraries acquired solely for the agent's task, and temporary binaries produced
+by agent-only probes or tools, under `output/`. Do not copy normal Zephyr build
+artifacts there. Keep other agent plans and temporary evidence in the same task
+directory. In a Git worktree, add
+`/.zephyr-skills/` to the repository-local exclude file returned by
+`git rev-parse --git-path info/exclude` before writing audit artifacts. Preserve
+existing entries, avoid duplicates, and verify before handoff that
+`git status --short` contains no `.zephyr-skills/` paths. Keep Zephyr-native
+build trees, binaries, and test outputs separate: honor a user-specified path,
+otherwise use defaults such as configured `build.dir-fmt`, `build/`, or
+`twister-out/`. Record the effective paths, do not stage `.zephyr-skills/`
+unless requested, and report the task directory and unresolved gaps at handoff.
+
+## Evidence workflow
+
+1. Read the [safety guardrails](references/safety-guardrails.md) in full before
+   producing safety-facing output.
+2. Bound the claim to a safety scope layer, owner, source baseline, and evidence
+   boundary.
+3. Create only the artifacts the task needs: traceability, evidence, scope, or
+   assumptions.
+4. For implementation claims, link requirements to code or configuration and
+   verification evidence. For product, process, or assessment claims, link the
+   applicable requirement, process, or assessment evidence and mark code or
+   test fields as not applicable or as gaps.
+5. Separate facts, assumptions, human judgments, and gaps, and state what the
+   evidence does not prove.
+6. Leave certification, risk acceptance, and scope-closure decisions to the
+   responsible humans and assessors.
+
+Never treat draft or partial evidence as proof of product ASIL compliance,
+IEC 61508 SIL/SC compliance, or completed certification.
diff --git a/skills/zephyr/zephyr-safety-evidence/references/safety-guardrails.md b/skills/zephyr/zephyr-safety-evidence/references/safety-guardrails.md
new file mode 100644
index 0000000..e930e81
--- /dev/null
+++ b/skills/zephyr/zephyr-safety-evidence/references/safety-guardrails.md
@@ -0,0 +1,64 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2026 Process Mission
+SPDX-License-Identifier: MIT
+-->
+
+# Zephyr Safety Guardrails
+
+Use this reference only when requirements or traceability are explicitly in a
+safety, certification, ASIL, SIL, or safety-case context.
+
+## Claim boundary
+
+- Describe only what cited artifacts demonstrate for the stated baseline.
+- Zephyr's public safety overview targets IEC 61508 SIL 3 / SC 3 for a limited
+  source scope that remains to be defined. Treat source areas as candidate
+  evidence until an approved scope, revision, and owner are cited.
+- Project evidence is not product certification evidence. Platform, hardware,
+  toolchain, integration, application, and assessor evidence remain with the
+  responsible integrator and organizations.
+- Tie IEC 61508 SIL statements to the applicable safety function and scope; do
+  not assign SIL to arbitrary code from component evidence alone.
+- Never infer certification, a safety integrity level, or readiness from draft,
+  partial, or public-only evidence.
+
+## Evidence layers
+
+Keep these boundaries explicit:
+
+1. Zephyr project evidence for an approved source scope and baseline.
+2. Platform evidence for the BSP, drivers, middleware, hardware, and toolchain.
+3. Product or vehicle evidence for the item, hazards, safety concepts,
+   integration, validation, and safety case.
+
+Do not use evidence from one layer to close a claim owned by another.
+
+## Lifecycle routing
+
+Select the applicable standard and route before defining gates. Do not apply a
+single lifecycle to every safety task.
+
+- For Zephyr's IEC 61508 effort, confirm the approved limited source scope and
+  whether route 3s or 1s applies. Follow the current Safety Committee process
+  and source baseline rather than assuming a fixed certification boundary.
+- For ISO 26262 product work, establish the item definition, HARA, safety goals,
+  and ASIL before the functional safety concept. Refine that into the technical
+  safety concept and derived requirements before implementation, integration,
+  verification, and assessment evidence is claimed complete.
+- For another standard or route, record the lifecycle mapping, responsible
+  owner, required outputs, and unresolved differences.
+
+These activities may iterate, but later evidence cannot substitute for a
+missing concept, scope, or ownership decision.
+
+## Evidence record
+
+For each safety-facing claim, record:
+
+- claim and status;
+- standard and edition, lifecycle route, scope layer, source baseline, and
+  owner;
+- requirement, process, or assessment source;
+- code or configuration path and verification evidence for implementation
+  claims, otherwise `not applicable` or an explicit gap;
+- assumptions, contrary evidence, unresolved gaps, and next human decision.