check
Run a check on the conformance of a project or the engine with the conventions — bare = check roster
Install
npx skills add https://github.com/pche-broken-artist/forge-of-thought/tree/main/.claude/skills/check
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install pche-broken-artist-forge-of-thought@llmmart
git clone https://github.com/pche-broken-artist/forge-of-thought.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole pche-broken-artist/forge-of-thought collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Checks live as .claude/agents/check-<name>.md — one isolated agent
per check, each defined by what it verifies, said in its
description; the roster is the scan of those files. Adding a check
means adding an agent file from templates/check-definition.md; this command
does not change (who creates a check, and when: CLAUDE.md, Isolated
reviewers).
Shared behaviour: the contract skill named in each check file's front-matter (CLAUDE.md, Isolated reviewers). Like the critic and the challenger, a check agent writes nothing: it returns its report and this command files it. This command chooses the check, passes the target, files the report and presents it.
Bare /check — the roster. List the available checks (scan
.claude/agents/check-*.md, their description fields — each says
when it fits) and recommend which fits the moment; what a save or a
release runs is /save's and /release's to say. A recommendation,
never a gate; no check runs on Claude's own judgement.
/check <name> [slug] — run it. The target is a project by its slug
(the engine's own project is forge) or the engine root; which of the
two a check takes, and what a missing or an added slug means to it, is
its Lens section's to say — read it there. Typed, that is
/check engine (the engine root, no slug) or /check light forge (the
forge's own project). First remake the target's ledger,
python scripts/forge-ledger.py <slug> (for the engine forge), so
that the agent reads a current file; then invoke the check-<name>
subagent (Agent tool, session model) with the target path and nothing
else (CLAUDE.md, Isolated reviewers). More than one check on one target
may be launched at once and awaited together.
When it returns:
- File the report, when it has findings. Write it into the target
project (for the engine
projects/forge) asreviews/YYYY-MM-DD-check-<name>.md: a front-matter ofdate,project,check,targetandreviewer: check <name> (isolated context), then the agent's text word for word — nothing added, dropped or reworded but the ID, the next freeFND.NNNNof the project's sequence, in tens, set at the head of each new finding; thenpython scripts/forge-report.py <path>; the printed name is the one the ledger cites. Thenpython scripts/forge-ledger.py <slug>: how the report becomes rows istemplates/ledger.md's (Findings). A library has noreviews/until its first finding: create the directory then. A report that says "conforms" is filed nowhere. - Present the report to the principal as it came: the one-line verdict, then the findings in their ranking with their IDs, compactly.
- Offer at once every finding marked "immediate fix" as one step
(CLAUDE.md, Working methods, Step by step); apply on his word,
state
resolved. - End by offering a walkthrough of the remaining findings
(
.claude/skills/walkthrough/SKILL.md, the one owner of its shape and of the verdict words). Whatacceptwrites here: the fix, agreed here and written once at the round's end, stateresolved; a rule worth changing goes to the intent. The other verdicts are the walkthrough's. Nothing blocks (POS.0430): a release may proceed with a finding parked.
Do not judge substance or document quality — that is /critique and
/challenge territory. A check verifies conformance only.
Files (forge-of-thought)
-
SKILL.md 3.5 KB
--- description: Run a check on the conformance of a project or the engine with the conventions — bare = check roster argument-hint: "[check] [project-slug]" --- Checks live as `.claude/agents/check-<name>.md` — one isolated agent per check, each defined by what it verifies, said in its `description`; the roster is the scan of those files. Adding a check means adding an agent file from `templates/check-definition.md`; this command does not change (who creates a check, and when: CLAUDE.md, Isolated reviewers). Shared behaviour: the contract skill named in each check file's front-matter (CLAUDE.md, Isolated reviewers). Like the critic and the challenger, a check agent writes nothing: it returns its report and this command files it. This command chooses the check, passes the target, files the report and presents it. **Bare `/check` — the roster.** List the available checks (scan `.claude/agents/check-*.md`, their `description` fields — each says when it fits) and recommend which fits the moment; what a save or a release runs is `/save`'s and `/release`'s to say. A recommendation, never a gate; no check runs on Claude's own judgement. **`/check <name> [slug]` — run it.** The target is a project by its slug (the engine's own project is `forge`) or the engine root; which of the two a check takes, and what a missing or an added slug means to it, is its Lens section's to say — read it there. Typed, that is `/check engine` (the engine root, no slug) or `/check light forge` (the forge's own project). First remake the target's ledger, `python scripts/forge-ledger.py <slug>` (for the engine `forge`), so that the agent reads a current file; then invoke the `check-<name>` subagent (Agent tool, session model) with the target path and nothing else (CLAUDE.md, Isolated reviewers). More than one check on one target may be launched at once and awaited together. When it returns: 1. File the report, when it has findings. Write it into the target project (for the engine `projects/forge`) as `reviews/YYYY-MM-DD-check-<name>.md`: a front-matter of `date`, `project`, `check`, `target` and `reviewer: check <name> (isolated context)`, then the agent's text word for word — nothing added, dropped or reworded but the ID, the next free `FND.NNNN` of the project's sequence, in tens, set at the head of each new finding; then `python scripts/forge-report.py <path>`; the printed name is the one the ledger cites. Then `python scripts/forge-ledger.py <slug>`: how the report becomes rows is `templates/ledger.md`'s (Findings). A library has no `reviews/` until its first finding: create the directory then. A report that says "conforms" is filed nowhere. 2. Present the report to the principal as it came: the one-line verdict, then the findings in their ranking with their IDs, compactly. 3. Offer at once every finding marked "immediate fix" as one step (CLAUDE.md, Working methods, Step by step); apply on his word, state `resolved`. 4. End by offering a **walkthrough** of the remaining findings (`.claude/skills/walkthrough/SKILL.md`, the one owner of its shape and of the verdict words). What `accept` writes here: the fix, agreed here and written once at the round's end, state `resolved`; a rule worth changing goes to the intent. The other verdicts are the walkthrough's. Nothing blocks (POS.0430): a release may proceed with a finding parked. Do not judge substance or document quality — that is `/critique` and `/challenge` territory. A check verifies conformance only.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.