writing — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited writing (Agent Skill) and scored it 92/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 2 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Multi-phase writing pipeline with a panel of specialised critics. Modeled on Katie Parrott's process and the existing research plugin's orchestrator pattern.
Use the host platform's equivalent tools without changing the workflow:
| Capability | Claude Code | Codex |
|---|---|---|
| Subagent dispatch | Agent tool | spawn_agent only when available and permitted. Otherwise run the phase inline. |
| Progress list | TaskCreate, TaskUpdate | update_plan |
| User questions | AskUserQuestion | Ask a concise direct question, or use the host structured question tool when available |
| File reads | Read | shell reads such as sed, rg, or equivalent file read tools |
| File writes and edits | Write, Edit | apply_patch or equivalent file edit tools |
| Shell | Bash | shell command tool |
Where this skill says "Agent tool", "TaskCreate", "TaskUpdate", "AskUserQuestion", "Read", "Write", "Edit", or "Bash", use the mapped host capability. When a platform cannot dispatch subagents for the current request, keep the same artifact boundaries and run each phase inline in the orchestrator.
State root is platform-specific. Claude Code uses ~/.claude/projects. Codex uses ${CODEX_HOME:-~/.codex}/projects when writable, otherwise create .codex-skill-state/ under the current working directory.
Ask the user what they want to write about (or what existing piece they want to work on).
Resolve working directory in this order:
--dir ./path/to/project/interview.md, outline.md, intake.md, pyramid.md, draft.md, critique.md, treat the cwd as the working directory<state-root>/<project-id>/writing-skill-state.json (where <project-id> is the cwd path with slashes replaced by hyphens, leading hyphen). If a working directory is recorded for an in-flight piece, offer to resume there.writing/{slug}-{YYYY-MM-DD}/ in the cwd.Resolution order:
--style-guide ./path/to/guide.mdstyle-guide.md or CLAUDE.md in the working directory and parents (up to repo root)default-style-guide.md shipped with this skill. Resolve its absolute path by locating the directory of this SKILL.md file (the skill's own install path) via Glob on **/writing/SKILL.md under the active plugin directory, then take the parent.If multiple candidates exist at the project level (e.g., both style-guide.md and a CLAUDE.md in scope), use AskUserQuestion to ask once which to use, then record the choice in the state file.
Surface the active guide in the first response: "Using style guide: {path}".
Panel composition and the outline / draft phases change based on format. The pipeline branches on whether the format is analytical (memo, briefing, announcement) or narrative (essay, blog, talk, newsletter).
Supported formats:
essay (default), blog, talk, newslettermemo, briefing, announcementtutorial, how-to, reference, explanationResolution order:
--format <format>essay and surface the default in the first response with an inline change hint: "Format: essay (default). Pass --format memo|briefing|announcement|newsletter|blog|talk|tutorial|how-to|reference|explanation to change."Ask via AskUserQuestion only when the working directory name or the interview synthesis strongly signals a different format than the recorded state (for example, a state-stored essay format but the working directory is memos/q3-roadmap-2026-04-23/). In ambiguous cases, surface both candidates and let the user pick. Otherwise, resolve silently.
Format gates:
memo, briefing, announcement) skip writing's interview and outline phases entirely. Phase 1 dispatches the pyramid skill's intake; Phase 2 dispatches pyramid's construct, audit, opener, and render phases. The pyramid pipeline produces pyramid.md, which is then consumed by writing's throughline (Phase 3) and analytical draft (Phase 4) phases.memo, newsletter, announcement add the Smart-Brevity critic to the panel fan-out. Other formats run the default seven-critic panel. Note: briefing does NOT add Smart-Brevity, because briefings are dense by construction and the Smart-Brevity lens has lower signal there.tutorial, how-to, reference, explanation) skip writing's interview, outline, draft, panel, and finishing phases entirely. Phase 1 dispatches the tech-doc skill's intake; Phase 2 dispatches tech-doc's outline + throughline + draft + panel + finishing as one cohesive sub-pipeline. Writing's Phase 5 (panel) and Phase 6 (finishing) are skipped because tech-doc owns end-to-end. The tech-doc pipeline produces draft.md, critique.md, finishing-notes.md, and glossary.md.Surface the active format in the first response alongside the style guide: "Format: {format}. Using style guide: {path}". Record the format in the state file under the project key.
Scan the working directory for existing artifacts. Two artifact families exist depending on format:
Narrative format artifacts (essay, blog, talk, newsletter):
interview-synthesis.md exists → interview phase completeoutline.md exists → outline phase completethroughline.md exists → throughline phase completedraft.md exists → draft phase completecritique.md exists → panel phase completefinishing-notes.md exists → finishing phase has started or completedAnalytical format artifacts (memo, briefing, announcement):
intake.md exists → pyramid intake (Phase 1) completeconstruction.md exists → pyramid construct (Phase 2 substep) completeaudit-summary.md exists → pyramid audit (Phase 2 substep) completeopener.md exists → pyramid opener (Phase 2 substep) completepyramid.md exists → pyramid render (Phase 2) complete; outline equivalent ready for throughlinethroughline.md exists → throughline phase completedraft.md exists → draft phase completecritique.md exists → panel phase completefinishing-notes.md exists → finishing phase has started or completedTechnical format artifacts (tutorial, how-to, reference, explanation):
intake.md exists → tech-doc intake (Phase 1) completeoutline.md exists (tutorial/how-to/explanation) OR schema.md exists (reference) → tech-doc outline (Phase 2 substep) completethroughline.md exists → tech-doc throughline (Phase 2 substep) completedraft.md exists → tech-doc draft (Phase 2 substep) completecritique.md exists → tech-doc panel complete (Phase 2 sub-substep)finishing-notes.md exists → tech-doc finishing has started or completed (Phase 2 sub-substep)glossary.md exists → tech-doc terminology-consistency pass has runFor technical formats, writing's phase identifiers map to tech-doc's: writing's "Phase 1" is tech-doc's intake; writing's "Phase 2" is tech-doc's everything-after-intake.
Determine the latest completed phase. Present to user:
User can also pre-empt the dialogue by passing --phase X (X ∈ {interview, outline, throughline, draft, panel, finishing}).
Use the progress list to add one task per phase that will run, plus sub-tasks for the panel and finishing phases. Two task list shapes exist depending on format.
Narrative format task list (essay, blog, talk, newsletter):
1. Phase 1: Interview the author
2. Phase 2: Negotiate outline
3. Phase 3: Throughline check (≤10-word gate)
4. Phase 4: Draft sections
5. Phase 5: Run panel review
├── Critic: Hemingway
├── Critic: Hitchcock
├── Critic: Mom reader
├── Critic: Asshole reader
├── Critic: Clarity
├── Critic: Usage
├── Critic: Steel-man
└── Critic: Smart-Brevity (only for newsletter)
6. Phase 6: Finishing pass
├── AI-pattern detector
├── Style enforcer
├── Line editor
└── SedarisAnalytical format task list (memo, briefing, announcement):
1. Phase 1: Pyramid intake (mode, audience, reader question)
2. Phase 2: Pyramid construct + audit + opener + render
├── Construct
├── Audit panel (MECE, So-What, Q-A Alignment, Inductive-Deductive)
├── Opener (SCQA)
└── Render pyramid.md
3. Phase 3: Throughline check (≤10-word gate on apex)
4. Phase 4: Analytical draft
5. Phase 5: Run panel review
├── Critic: Hemingway
├── Critic: Hitchcock
├── Critic: Mom reader
├── Critic: Asshole reader
├── Critic: Clarity
├── Critic: Usage
├── Critic: Steel-man
└── Critic: Smart-Brevity (only for memo, announcement)
6. Phase 6: Finishing pass
├── AI-pattern detector
├── Style enforcer
├── Line editor
└── Analytical voiceTechnical format task list (tutorial, how-to, reference, explanation):
1. Phase 1: Tech-doc intake (quadrant-specific)
2. Phase 2: Tech-doc pipeline (outline, throughline, draft, panel, finishing)
├── Outline (or schema for reference)
├── Throughline gate
├── Draft (quadrant-specific)
├── Panel review (7 critics, quadrant-gated composition)
└── Finishing (AI-pattern, style-enforcer-tech, terminology-consistency)The technical pipeline is one-shot from writing's perspective: writing dispatches to tech-doc once and tech-doc owns end-to-end. The expanded sub-tree is shown for visibility into what's running.
For phase-selectable runs, only the requested phases get tasks.
Mark each task as in_progress when starting, completed when the artifact is verified.
Dispatch each phase agent via the host subagent tool when supported. The orchestrator injects context into the prompt template.
#### Dispatch conventions (apply to every phase)
**Dispatch:** header. The dispatched body itself contains nested fences for example outputs. The simplest robust approach: read the entire prompt file as text, perform placeholder substitution ({TOPIC}, {OUTPUT_PATH}, {STYLE_GUIDE_PATH}, {REVIEWER_FEEDBACK}, {YYYY-MM-DD}), and pass the full result to the host subagent tool. The dispatched agent ignores the surrounding commentary because the actionable instructions sit inside the visible prompt body.{REVIEWER_FEEDBACK} is non-empty (re-dispatch on a failed gate), append this standing instruction to the dispatched prompt, regardless of what the prompt template itself says: "Reviewer feedback is provided above. Read the existing artifact in the output directory, address the specific concerns, and update the file in place rather than starting fresh." This compensates for the asymmetric treatment of feedback across the prompt files.{YYYY-MM-DD} resolves to today's date in ISO format.#### Phase 1: Interview (narrative formats) or Pyramid intake (analytical formats)
Narrative formats (essay, blog, talk, newsletter):
interview-prompt.md from this skill directoryinterview.md and interview-synthesis.md existAnalytical formats (memo, briefing, announcement):
Skip writing's interview entirely. Run the pyramid skill's Phase 1 (intake) in dispatched mode as documented in plugins/writing/skills/pyramid/SKILL.md, with these adjustments:
memo → genre Memo. briefing → genre Briefing. announcement → genre Announcement. Do NOT ask the user; surface the pre-fill in a one-line confirmation: "Genre: {genre} (from format)."dispatched_from: writing so future runs know the entry point.intake.md exists.The orchestrator reads pyramid SKILL.md sections at dispatch time. No code or prompt files are duplicated; the dispatched mode is an instruction overlay applied to pyramid's standalone Phase 1.
Technical formats (tutorial, how-to, reference, explanation):
Skip writing's interview entirely. Run the tech-doc skill's Phase 1 (intake) in dispatched mode as documented in plugins/writing/skills/tech-doc/SKILL.md, with these adjustments:
tutorial → quadrant tutorial, how-to → quadrant how-to, reference → quadrant reference, explanation → quadrant explanation. Surface the pre-fill in a one-line confirmation: "Quadrant: {quadrant} (from format)." Tech-doc's standalone path always asks the quadrant question; in dispatched mode, accept the format-derived value and skip the question.--style-preset set per writing's resolved style guide if it matches a preset (google, microsoft, or house). Otherwise default to house.dispatched_from: writing to the intake.md fields so resume logic can distinguish dispatched-mode intake from a standalone tech-doc run. (Mirrors the analytical dispatch's dispatched_from: writing field added to pyramid's intake.md.)intake.md exists.The orchestrator reads tech-doc SKILL.md sections at dispatch time. No code or prompt files are duplicated.
#### Phase 2: Outline (narrative formats) or Pyramid pipeline (analytical formats)
Narrative formats (essay, blog, talk, newsletter):
outline-prompt.mdoutline.md existsAnalytical formats (memo, briefing, announcement):
Run pyramid skill's Phases 2 through 5 (construct, audit, opener, render) inline as documented in plugins/writing/skills/pyramid/SKILL.md. The pyramid pipeline is reused unchanged; the orchestrator follows pyramid SKILL.md's instructions for each phase.
pyramid/SKILL.md. Mode-branched (greenfield, restructure, socratic) based on the mode collected in Phase 1. Verify construction.md exists.pyramid/SKILL.md. Consolidate into audit-summary.md. Apply pyramid's CRITICAL re-dispatch logic (up to 2 iterations) verbatim.pyramid/SKILL.md. Apply pyramid's MISMATCH handling verbatim.pyramid.md per pyramid/SKILL.md's render template. The pyramid is the outline equivalent for the analytical pipeline.pyramid.md to the user. Accept revisions via AskUserQuestion ("Pyramid as proposed, or revisions before draft?"). On revisions, re-dispatch the construct agent (pyramid Phase 2) with the feedback injected, then re-run audit, opener, and render.After Phase 2 completes, the working directory contains intake.md, construction.md, audit-summary.md, opener.md, and pyramid.md. The throughline phase reads pyramid.md's apex line; the analytical draft phase reads pyramid.md whole.
Technical formats (tutorial, how-to, reference, explanation):
Run tech-doc skill's Phases 2-6 (outline, throughline, draft, panel, finishing) inline as documented in plugins/writing/skills/tech-doc/SKILL.md. The tech-doc pipeline is reused unchanged; the orchestrator follows tech-doc SKILL.md for each phase.
tech-doc/SKILL.md. Verify outline.md (tutorial/how-to/explanation) or schema.md (reference) exists.draft.md, glossary.md, and finishing-notes.md to the user.#### Phase 3: Throughline
For technical formats (tutorial, how-to, reference, explanation): SKIPPED. Tech-doc's throughline gate has already run during the dispatch in writing's Phase 2. Mark phase task completed and proceed to Phase 4.
Orchestrator-only synchronous gate. No agent dispatch. Happens after Phase 2 completes, before the draft agent is dispatched. If the writer cannot compress the piece into ten words, the piece is not ready to draft.
Source of truth varies by format:
{OUTPUT_PATH}/outline.md and extract the **Thesis (one sentence):** line.{OUTPUT_PATH}/pyramid.md and extract the line under the ## Apex header (the one-sentence governing thought rendered verbatim from construction.md).{OUTPUT_PATH}/throughline.md as a single-line file containing only the accepted throughline (no markdown headers, no decoration).Edge case for analytical formats: if pyramid.md lacks a ## Apex header (e.g., a degraded MISMATCH render), fall back to reading construction.md and extracting the apex node directly. If neither is parseable, ask the user for the apex sentence directly before running the gate.
#### Phase 4: Draft
For technical formats (tutorial, how-to, reference, explanation): SKIPPED. Tech-doc's draft agent has already run during the dispatch in writing's Phase 2. Mark phase task completed and proceed to Phase 5.
Narrative formats (essay, blog, talk, newsletter):
draft-prompt.mddraft.md existsAnalytical formats (memo, briefing, announcement):
draft-analytical-prompt.mdpyramid.md, intake.md, throughline.md (if present), and audit-summary.md.draft.md exists#### Phase 5: Panel review
For technical formats (tutorial, how-to, reference, explanation): SKIPPED. Tech-doc's Phase 5 has already run during the dispatch in writing's Phase 2. Mark phase task completed and proceed to Phase 6.
Fan out: dispatch all critic agents in parallel when supported. The critic set depends on format.
Default panel (seven critics). Used for essay, blog, talk formats.
| Prompt file | Output file | Lens |
|---|---|---|
critics/hemingway.md | critique-hemingway.md | Economy: cut adjectives, kill darlings |
critics/hitchcock.md | critique-hitchcock.md | Pacing: reader engagement, bomb under the table |
critics/mom-reader.md | critique-mom.md | Accessibility: where the general reader gets lost |
critics/asshole-reader.md | critique-asshole.md | Rigor: unearned claims, missing counterarguments |
critics/clarity.md | critique-clarity.md | Precision: vague abstractions, unclear antecedents (Zinsser) |
critics/usage.md | critique-usage.md | Correctness of form: grammar, parallelism, misused words (Strunk & White) |
critics/steel-man.md | critique-steelman.md | Preemption: strongest opposing thesis and whether the draft engages it |
Extended panel (eight critics). Used for formats memo, newsletter, announcement. Adds one format-gated critic to the default seven:
| Prompt file | Output file | Lens |
|---|---|---|
critics/smart-brevity.md | critique-smartbrevity.md | Scannable structure: muscular lead, one takeaway early, short sentences, no fluff (Axios method) |
For each critic in the active set:
When all active critics return, consolidate into critique.md (include Smart-Brevity rows only when it ran):
# Panel Critique
## Verdicts
| Critic | Verdict | Headline |
|--------|---------|----------|
| Hemingway | <PASS / MINOR / CRITICAL> | <one-line summary> |
| Hitchcock | ... | ... |
| Mom reader | ... | ... |
| Asshole reader | ... | ... |
| Clarity | ... | ... |
| Usage | ... | ... |
| Steel-man | ... | ... |
| Smart-Brevity | ... | ... | (only when format gated it in)
## Hemingway
<full content of critique-hemingway.md>
## Hitchcock
<full content of critique-hitchcock.md>
## Mom reader
<full content of critique-mom.md>
## Asshole reader
<full content of critique-asshole.md>
## Clarity
<full content of critique-clarity.md>
## Usage
<full content of critique-usage.md>
## Steel-man
<full content of critique-steelman.md>
## Smart-Brevity
<full content of critique-smartbrevity.md, only when the Smart-Brevity critic ran>Then check verdicts. Match on the first whitespace-delimited token of each critic's `Verdict:` line. Critic prompts emit PASS, MINOR ISSUES, or CRITICAL ISSUES; only the first token is the gate signal. Expected tokens: PASS, MINOR, CRITICAL.
PASS or MINOR → continue to finishingCRITICAL → re-dispatch the draft agent with the consolidated critique injected as REVIEWER_FEEDBACK. Re-run the panel. Repeat up to 2 iterations. If still CRITICAL after 2 iterations, present remaining critical issues to user via AskUserQuestion: "Continue to finishing, or pause for manual intervention?"Mark phase task completed when verdict allows progression or user overrides.
#### Phase 6: Finishing
For technical formats (tutorial, how-to, reference, explanation): SKIPPED. Tech-doc's Phase 6 has already run during the dispatch in writing's Phase 2. Present the final artifacts to the user and proceed to Step 7.
Sequential, NOT parallel. Each pass updates the draft in place; later passes need the earlier passes' changes.
Narrative formats (essay, blog, talk, newsletter): run the four passes in this order:
finishing/ai-pattern-detector.mdfinishing/style-enforcer.mdfinishing/line-editor.mdfinishing/sedaris.md (literary voice; reads interview-synthesis.md for tone calibration)Analytical formats (memo, briefing, announcement): run the four passes in this order:
finishing/ai-pattern-detector.mdfinishing/style-enforcer.mdfinishing/line-editor.mdfinishing/analytical-voice.md (executive voice; reads intake.md for audience calibration; replaces Sedaris because analytical formats do not run the interview phase that Sedaris depends on)For each pass in order:
finishing-notes.mdAfter all four passes, present draft.md and finishing-notes.md to the user. The piece is now ready for the writer's manual voice pass per the user feedback memory (drafted prose is a skeleton, the writer rewrites in own voice).
Update the state file. The working directory is the key under projects (not a field). For that key, write:
active_style_guide: absolute pathlast_completed_phase: name of last successful phaselast_run_at: ISO timestampSee the State File Format section below for the exact JSON shape.
Present the final draft and a summary of what each pass did.
mkdir -pinterview-synthesis.md; Throughline reads outline.md; Sedaris reads interview-synthesis.md; Draft reads outline.md and throughline.md if present; Panel and Finishing read draft.md. If the user invokes --phase X on a directory missing the upstream artifact, ask via AskUserQuestion whether to (a) run the missing upstream phase first, (b) accept a degraded run where the agent works without that input (only safe for Sedaris reading the synthesis, or Draft reading a missing throughline), or (c) cancel and let the user produce the artifact manuallyintake.md; Throughline reads pyramid.md (or construction.md as fallback); Analytical Draft reads pyramid.md, intake.md, optionally throughline.md and audit-summary.md; Analytical voice pass reads draft.md, intake.md, pyramid.md, optionally audit-summary.md; Panel and Finishing read draft.md. Apply the same three options on phase-jump with missing upstream.--format or the state file contains an unrecognised value, warn once, fall back to essay, and ask the user to confirmessay but the working directory contains pyramid.md, or recorded memo but contains outline.md. Ask via AskUserQuestion which format applies; record the corrected value.pyramid.md and the partial opener renders correctly.intake.md; throughline reads outline.md/schema.md; draft reads intake.md, outline.md/schema.md, optionally throughline.md; panel reads draft.md; finishing reads draft.md. Apply the same three-option pattern (run upstream / accept degraded / cancel) on phase-jump with missing upstream.tutorial but working directory contains pyramid artifacts (e.g., intake.md with genre: Memo), or pyramid was the intended track but intake.md has dispatched_from: writing and a quadrant: field. Use the dispatched_from field and the presence of quadrant: (technical) vs. genre: (analytical) to disambiguate. Ask via AskUserQuestion which format applies; record the corrected value.<state-root>/<project-id>/writing-skill-state.json:
{
"version": 1,
"projects": {
"<absolute-working-directory>": {
"active_style_guide": "<absolute-path-or-default>",
"format": "essay",
"last_completed_phase": "draft",
"last_run_at": "2026-04-16T12:00:00Z"
}
}
}Recognised format values: essay, blog, talk, newsletter, memo, announcement, briefing, tutorial, how-to, reference, explanation. Defaults to essay if absent. The format drives panel composition (Smart-Brevity critic added for memo, newsletter, announcement) and pipeline routing. For analytical formats (memo, briefing, announcement), the writing skill dispatches Phases 1 and 2 to the pyramid skill. For technical formats (tutorial, how-to, reference, explanation), the writing skill dispatches Phases 1 and 2 to the tech-doc skill, which owns Phases 5 and 6 (panel and finishing) as part of its dispatched pipeline; writing's Phases 5 and 6 are skipped.
format=talk produces a prose speech (a written piece a presenter reads or delivers). For a slide deck with sentence-takeaway headlines, visual briefs, and per-slide speaker notes, use writing:presentations instead.
The state file is keyed by working directory so multiple in-flight pieces in the same project can each have their own state.
Used in --phase flag and task list: interview, outline, throughline, draft, panel, finishing
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.