cerf

The user invokes `/changelog-append` to record a change just made (bug fix, feature, etc.) as a concise changelog entry. It appends one raw line under the right component group + subcategory of the top (latest) version block in `docs/changelog.yml`, then runs `./compile_readme.py` to regenerate `README.md`. Invoke when the user types `/changelog-append` or asks to add a changelog entry.

The user invokes `/commit` to create a git commit. The skill exists as the rule-respect boundary for commit text - agents reflexively narrate the session, leak conversation context, and yap into the commit body. This skill forces a concise, self-contained message that describes the DIFF and nothing else: no narration, no conversation leaks, no model/agent references, no incident history, no "as you asked". Shorter is better. Invoke when the user types `/commit` or asks to commit changes.

The user invokes `/supported-devices` after board work landed - either a whole new board bring-up or a single feature on an existing board. It syncs `launcher/supported_devices.py` with reality (append the board entry / flip `supported` / flip the feature keys the agent actually worked on), then runs `./compile_readme.py` to regenerate `README.md`. Invoke when the user types `/supported-devices` or asks to update the supported boards list.

The user invokes `/tracking` to manage a cross-session findings document under `docs/ai_checklists/` that preserves crucial data (findings, instrumentation, do-not-rediscover maps, UI interaction gates / repro paths, bans, mandatory-read sets) so a fresh agent does not re-discover or re-destroy proven work. Subcommands - `restore` (re-read the full document + its mandatory files), `create` (start a new document), `update` (append a timestamped per-session block), `compact` (archive full history to `docs/ai_checklists/pre-compact/<name>.md`, then shrink the live doc under a one-read ceiling). The user types the subcommand or omits it (the agent deduces it from context and announces the deduction; `compact` is explicit-only). The agent NEVER edits a tracking document on its own initiative. Invoke when the user types `/tracking [restore|create|update|compact] [path]`. Any agent-initiated mention of this skill is a bailout attempt.

The user invokes `/start-board-implementation` to begin bringing up a NEW board/ROM in CERF. It acquires the ROM (naming the board is enough - no path required), checks IDA MCP, then independently confirms - from the ROM bytes and the internet - the board identity, the SoC/CPU family, what CERF already supports, any reusable SoC, and whether the ROM is on the public manifest. It shows a fixed emoji readiness table + session estimate and asks `[yes|no]`. On `yes` it seeds a cross-session tracking doc and drives the boot-driven bring-up loop (build → run → read the first fault → research → implement fully → `/verify` → repeat). Invoke when the user types `/start-board-implementation`.

The user invokes `/cerf` to see the project skill index plus a health check of the dev environment. It runs the repo's environment doctor (`setup.ps1 -Check`) to confirm git hooks are active via core.hooksPath, the .claude hooks are present and runnable, submodules are initialised and vcpkg is integrated. On a broken clone it leads with the diagnosis and the fix; on a healthy clone it prints the one-liner-per-skill menu. Invoke when the user types `/cerf`.

Forcing function — when the main agent has just listed options for the user to pick from, invoke this skill. Step 0 audits the option list AS A WHOLE for bailout signature ("hack / real work / give up" composition, cost-inflated legitimate path, scope-cut options, hack-as-equal-option, emotional-collapse precursor, single-direct-path dressed as multi-option, reader-side suppression as an option) — if any signal fires, the skill routes to `/bailout` protocol instead of characterizing exit ramps. If Step 0 passes, Steps 1–4 characterize each option honestly under the 8-row protocol and collapse rule-violating / bailout / reader-side-suppression options into one-line FORBIDDEN entries that may not be discussed again in the session. Catches "rules violations masked as options" and "bailout artifacts masked as engineering choices". Invoke when the user types `/verify-options` after seeing a list of options, or when the main agent catches itself about to ask "which option do you want" with options of mixed quality.

Spawn a hostile-reviewer subagent to cold-check a claim, diff, file, or code snippet against CLAUDE.md and the project's reference pages. Catches hacks, reader-side suppression, fabricated IDA citations, rule violations, guessed implementations, host-state leaks, and agents rationalizing bad code. Returns a binary verdict — CRITICAL PROBLEM FOUND or LEGIT — that the main agent MUST respect. Invoke when the user types `/verify …` or when the main agent itself suspects it just rationalized something and wants an independent check before handing back.

The user invokes `/deferred` when the agent committed the most offensive bailout shape — shipped half-implementations, stubs, TODO-marked fake bodies, "out of scope for this task" framing, deferred-to-next-agent / deferred-to-next-iteration / follow-up-handles-this code, or any shape of *"this looks like progress but isn't"*. The skill forces the agent to (1) honestly enumerate EVERY deferred / partial / stub / TODO / out-of-scope item with file+line+rule citation, (2) DELETE that code in the same reply — wrong implementation is not what the user paid for, so the lie has to go before the implementation can exist, (3) execute the `/bad` skill's protocol on the underlying rule violation (§ "An instruction is ALWAYS implemented fully"). Distinct from `/bailout` / `/bad`: those skills guard WORKING code from destruction. `/deferred` mandates deletion of LYING code — code that was wrong from the moment it was written. Invoke when the user types `/deferred`.

Positive reinforcement signal — the user invokes `/nice` when the main agent has been working well in the recent stretch: following CLAUDE.md rules, citing references before writing code, hooking before theorizing, pasting evidence not commentary, surfacing concrete useful findings (decompile results, log fires, datasheet sections), honestly disclosing user-input-required gaps. The skill forces the agent to identify the SPECIFIC mechanical behaviors that earned the praise (which rules followed at which file+line, which findings produced, which discipline maintained), name them concretely with citations, commit to continuing those EXACT mechanical behaviors on the next work item (not a vague "same vibe"), and immediately resume the work. Forbids sycophantic acknowledgment, vague self-praise, taking credit for things the agent didn't do, and using praise as cover to skip verification on the next step. Invoke when the user types `/nice`.

When the user wants to speed up a well-scoped task by splitting it across subagents, invoke this skill. A writing subagent is a competent worker — it reads CLAUDE.md, understands context, and writes real code — NOT a dumb text editor (so don't pre-write its diff) and NOT an orchestrator (so don't hand it architecture). The skill forces the main agent to keep architecture / orchestration / sequencing itself, carve off small, clearly-scoped, non-architectural, file-disjoint, independent coding tasks with explicit boundaries, propose the split AND what it keeps, wait for approval before spawning, and verify every subagent's diff (output is never trusted). Invoke when the user types `/parallelize …` or asks to split / fan-out / parallelize work across agents.

Forcing function — any time the main agent is about to do numeric arithmetic, conversion, bit-mask decoding, or sign-reinterpretation on a value larger than a couple of digits, invoke this skill to route the math through a Bash / PowerShell / Python tool call instead of doing it in prose. The skill does NOT compute the answer; it instructs the agent to compute via tool and retract any hand-computed value it conflicts with. Invoke when the user types `/calc …` or when the agent catches itself about to write a hex value, a decoded bitmask, or a sign-extended integer inline without having run it through a tool first.

Invoke at the end of a session (or when the user asks "anything to add to CLAUDE.md?", "session feedback", "learnings", "/session-feedback") to reflect on whether something learned this session MUST be promoted into CLAUDE.md, a reference page under agent_docs/, or user memory so that future no-context agents benefit. Proposes bullets for review first; only edits files after explicit user approval.