Run any Skill in Manus with one click

spec-lint

Stars2

Forks0

UpdatedJuly 9, 2026 at 18:03

Lint ADR phase prompts for the delegation-contract blocks (VERIFICATION, HANDOFF naming its RESULTS file, declared test mode, reality-override line, blast-radius line from ADR-62 on) via scripts/check_phase_prompts.py — the mechanical floor beneath /adr-create. Args: [ADR number or prompt file paths] (default: every tracked post-contract prompt). Use when the user says "lint the ADR prompts", "spec-lint ADR N", "check the phase prompts", or invokes /spec-lint. /adr-create runs this before reporting a new ADR done.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

pfBlockerNG

pfBlockerNG/pfBlockerNG

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name

spec-lint

description

Run the phase-prompt checker and report. The checker enforces presence of the CLAUDE.md delegation-contract blocks; judging their depth stays your job.

Args: {{ args }}

Step 1 — Resolve targets and run

No args → python3 scripts/check_phase_prompts.py (scans every tracked prompt; ADRs below the grandfather cutoff in the script are skipped by design).
An ADR number N → pass that ADR's prompt files: python3 scripts/check_phase_prompts.py .ADRs/ADR_{NN}_*/[0-9]*.txt.
Explicit paths → pass them through verbatim.

Step 2 — Report

Clean → say so, one line, and note anything the checker cannot see that you noticed anyway (a vague ACTION PLAN, a missing coverage matrix where the ADR says "all X", a new transformer — parser/regex/scrub/formatter — whose prompt carries no hostile-input rows — the checker's known blind spots; the transformer check is judgment, not mechanical: those words pepper ordinary ADR prose, so a word-trigger lint would be mostly false positives).
Violations → list them file-by-file with the missing block, and fix the prompts (prompt text is the ADR-text dev-only class — direct-to-devel landing per CLAUDE.md "Worktrees") or hand the list back to the author, per what the user asked.

Do not weaken the checker to make a prompt pass (CLAUDE.md never-weaken rule); a prompt that genuinely cannot satisfy a block states why in the prompt itself and the ADR documents the exemption.

More from this repository

same repository

gh-issue

pfBlockerNG/pfBlockerNG

Triage a GitHub issue end-to-end: read it whole, decide whether the report actually checks out, assess its impact, and produce an ordered RESOLUTION PLAN whose every step carries the brief notes for a delegate sub-agent. Args: <issue-number> [--fix]. Without --fix it stops at the plan (triage verdict + impact + per-step brief notes). With --fix it also EXECUTES the plan: all work happens in ONE dedicated git worktree on branch `issue/{NN}-{slug}` reused across every step, each step is run by a fresh sub-agent briefed with that step's prompt + CLAUDE.md + the previous step's handoff, you (the orchestrator) gate and independently verify each agent's output before the next starts, and each agent returns a handoff document you pass to the next. Lands the fix via /pr-merge-flow. Use when the user says "triage issue N", "look at issue N", "/gh-issue N", "investigate and fix issue N", or invokes /gh-issue.

2026-07-122

pr-merge-flow

pfBlockerNG/pfBlockerNG

Land a PR end-to-end the repo's DEFAULT way: review feedback first, then merge. Roughly `/pr-comments N --wait-for=coderabbitai && /pr-merge N` — get the review feedback, validate + apply each finding and reply, and ONLY if that completes cleanly, rebase the head onto the live base, wait for the real CI to go green (EXCLUDING the bot), merge `--rebase` (never a merge commit, never squash) and delete the branch. The review step ALWAYS runs the committed `review-single` workflow — ONE Claude sub-agent as an ADVERSARIAL reviewer IN ADDITION TO CodeRabbit, never a mere fallback — at reasoning effort xhigh (never below, never max): the latest Sonnet (5 or newer) — Fable (5 or newer), preferred for a large/complex PR, is temporarily unavailable, so use Sonnet (never Opus, never a multi-agent fan-out). In parallel it gives CodeRabbit ~10 minutes to acknowledge the PR: if it does, wait on its review too; if it stays silent, nudge it once with `@coderabbitai review` and wait 10 more minutes; if it is STILL silent the

2026-07-122

pr-merge

pfBlockerNG/pfBlockerNG

Land a pull request the repo's way: rebase its head branch onto the target branch, push, wait for CI to go green (EXCLUDING CodeRabbit — never block on the bot), then merge with `--rebase` (never a merge commit, never squash) and delete the remote branch. Refuses to merge a draft, a conflicting/unmergeable PR, or one whose CI is red. Args: [PR number] (defaults to the current branch's PR). Use when the user says "merge this PR", "land PR N", "rebase and merge", "merge it once CI passes", or invokes /pr-merge. Companion to /pr-comments — handle review feedback first, then /pr-merge.

2026-07-112

adr-create

pfBlockerNG/pfBlockerNG

Interactively design and author a NEW ADR (Architecture Decision Record) for this repository, then write it to .ADRs/ADR_NN_Name/ as ADR.md plus ordered phase-prompt .txt files ready for /adr-phase. Args: [optional one-line topic]. Use when the user says "create an ADR", "start a new ADR", "let's design ADR-N", or invokes /adr-create. You act as a skeptical design partner, not a transcriber.

2026-07-112

adr-phase

pfBlockerNG/pfBlockerNG

Implement an ADR (Architecture Decision Record) — a single phase, or the whole ADR end-to-end with the `all` selector. Args: <ADR-number> <phase-number | all> [--base <branch>]. Runs in a dedicated per-ADR git worktree (your main checkout stays free); each phase is executed by a fresh sub-agent with a clean context — briefed with the phase prompt, CLAUDE.md, and the prior phase's handoff, working in the full worktree. Commits AND pushes branch `adr/{NN}-{slug}` after every phase (crash-safe), so the run is resumable — on restart it resets to the last completed phase and continues. When the ADR is complete it offers a PR (default) or a rebase onto the base. Use when the user says "implement phase N of ADR-M", "run adr phase", "implement all of ADR-N", "finish ADR-N", or invokes /adr-phase.

2026-07-112

delegate

pfBlockerNG/pfBlockerNG

Run ONE ad-hoc coding task through the full CLAUDE.md delegation contract — the same brief → Sonnet 5 implementer → mechanical gate discipline /adr-phase and /gh-issue enforce, for work that belongs to neither. Args: <task description> [--base <branch>]. Use when the user says "delegate this", "have sonnet build X", "do this via the contract", or invokes /delegate. Not for ADR phases (/adr-phase) or issue fixes (/gh-issue --fix) — route those to their skills.

2026-07-112

name

spec-lint

description

Run the phase-prompt checker and report. The checker enforces presence of the CLAUDE.md delegation-contract blocks; judging their depth stays your job.

Args: {{ args }}

Step 1 — Resolve targets and run

No args → python3 scripts/check_phase_prompts.py (scans every tracked prompt; ADRs below the grandfather cutoff in the script are skipped by design).
An ADR number N → pass that ADR's prompt files: python3 scripts/check_phase_prompts.py .ADRs/ADR_{NN}_*/[0-9]*.txt.
Explicit paths → pass them through verbatim.

Step 2 — Report

Clean → say so, one line, and note anything the checker cannot see that you noticed anyway (a vague ACTION PLAN, a missing coverage matrix where the ADR says "all X", a new transformer — parser/regex/scrub/formatter — whose prompt carries no hostile-input rows — the checker's known blind spots; the transformer check is judgment, not mechanical: those words pepper ordinary ADR prose, so a word-trigger lint would be mostly false positives).
Violations → list them file-by-file with the missing block, and fix the prompts (prompt text is the ADR-text dev-only class — direct-to-devel landing per CLAUDE.md "Worktrees") or hand the list back to the author, per what the user asked.

Do not weaken the checker to make a prompt pass (CLAUDE.md never-weaken rule); a prompt that genuinely cannot satisfy a block states why in the prompt itself and the ADR documents the exemption.