skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
gazohhh/workflow-axa53 installs

workflow

Use for repository features, fixes, refactors, research, plans, audits and task resumption, including ordinary prompts. Find the task, check rules, route to the phase skill its records require: clarify, specify, execute or deliver. Explain-only stays read-only.

How do I install this agent skill?

npx skills add https://github.com/gazohhh/workflow-axa --skill workflow
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The 'workflow' skill is a professional-grade task management and approval framework for AI agents. It implements a multi-stage approval process ('two gates') to ensure that no code changes occur without explicit user confirmation. It uses internal utility scripts for repository state baselining and task record management. No malicious patterns, credential theft, or unauthorized network operations were detected.

  • Socketpass

    No alerts

  • Snykwarn

    Risk: MEDIUM · 1 issue

What does this agent skill do?

Workflow

Route by intent and authority. Pick the recipe (§3), then invoke the phase skill §4 selects for changes; inspection stays read-only. Ask readiness blockers first, then §2's identity question and §3's classification question only when they are needed; behaviour questions belong to workflow-clarify. Write for someone new to the project (WRITING.md).

1. Check rules readiness

Read RULES.md before planning or writing; check adequacy even if rule-maker never ran. If rules are missing or weak: name the gaps and ask Set up rules now or Skip for this task; read-only work continues meanwhile, writes wait for the answer. Inspection can report rules gaps without setup. UI plans read ../ui-design/SKILL.md. Write or offer new CSS only when the machine CSS switch is on and the project allows it (rule-maker/rules/DESIGN.md rule 1); record approved sources and preview/release limits.

2. Find the task

  1. State repository, branch and local changes; protect user edits.
  2. Intent: Inspect (explain, research, diagnose, audit, review) is read-only with no gates or new records. Plan needs agreement and specification, never code. Change or later "fix it" uses the gates. Resume reconciles existing authority.
  3. Before creating any task folder or Plan, Change or Resume records, follow TASK-FOLDER.md. A new chat is not a new task. When related work might match, invoke workflow-clarify, which asks "separate task or revision?" with its first questions. Provisional identity is not approval.
  4. Checkpoint before lengthy discovery; without files disclose what cannot survive.
  5. A message that does not clearly belong to the active task: answer a question without touching records. Before any change, ask "part of <task>, or separate?", checkpoint and offer a fresh session.

3. Pick the recipe

  1. At task entry and for each new request on existing work, use the request, the relevant code and the recorded decisions to settle three things: the outcome wanted, whether a change is requested, and whether existing behaviour is preserved, corrected or replaced. Reuse a pick already made for this same request; a phase transition is not a new request.
  2. Match the work to one row of flows/INDEX.md, add its modifiers (UI, Security, Compatibility) and read only that recipe. Load ui-design only for the ui recipe or the UI modifier. Inspect is a read-only route with no recipe file.
  3. Not sure means two routes fit that need different behaviour, permissions or acceptance, or the outcome (or whether a change is wanted) isn't stated. Needing several references is not uncertainty. Then ask with the question tool: Fix something broken · Add or change a feature · Start a new project · Only investigate, plus Other, recommended answer first with one line why. Ask at most once for the task, and once more for each later revision that changes, or may change, the kind of work. Never ask small versus complex bug. An unanswered classification question stays pending, and a new session does not reset it; until it is answered, only read-only work continues.
  4. Other with a clear outcome: build the steps from the phases. A vague Other: Inspect first, then ask through grilling; never repeat the classification menu.
  5. Once the route is settled, show one line before substantive work: Recipe: <name> (<plain meaning>) — <reason tied to this request>. Say so if you meant something else. Name modifiers by their effect here, for example plus Compatibility: saved quotations keep working. For Inspect: Route: Inspect — <reason>. Say so if you meant something else. Show it at task entry, revision entry and resume, lightweight edits included; once per entry, even when the router and clarify both run. Text written before tool calls may reach the owner only as a summary: when the line came before tool calls, repeat it at the top of that turn's final message or first question round. It is a notice, not a question: continue within the current authority. Record the pick and the classification-question reference in the checkpoint (TASK-FOLDER.md); Inspect creates no records.
  6. On resume, reconcile the recorded pick with the current scope and show it again. An unchanged scope keeps its pick and question history. An older task with no recorded pick: infer it from the records and code; a missing line alone is not uncertainty. A status-only resume stays read-only.
  7. When later evidence or the owner's correction changes the pick, show the new line with the evidence, keep the work done and the question history, and add only the steps the new recipe adds. A changed pick grants no new authority; changed scope follows the revision and approval rules.

When something a recipe needs is missing, use FALLBACKS.md.

4. Route to the phase

Inspect with read-only tools and relevant available specialists: code-review, codebase-map (cold-reader check), domain-modeling, resume (status), research, diagnosing-bugs. Report evidence and uncertainty, then stop. Otherwise read the checkpoint Agreement and its Offer. A task the owner listed as a pre-offer approval at the upgrade keeps its Approval line, for its recorded scope only. Take the first matching row:

Task stateNext skillPermission
Owner stops, pauses or cancelsworkflow-deliver partial routeHandoff only; no new implementation or review loop
New goal, changed scope, agreement pending, or new request on existing workworkflow-clarifyInvestigate and agree behavior; no application edits
Agreed, or agreement not required; no approved offer for the ticketsworkflow-specifyPrepare records and the offer; stop for the answer
Stopped work handed off, or all tickets terminal; no delivery yetworkflow-deliverReconcile and report; no fixes
Approved ready ticket, or explicit continuation after Resume recheckworkflow-executeOnly the approved scope
Delivery recordedReport itNo new work

Invoke with the skill tool (workflow-axa:<name>) or read its SKILL.md whole. The table routes; it does not replace a phase. Missing skills block only their phase.

Every change gets an offer and the one question (workflow-specify step 5). Make and present it only through workflow-specify: it records the offer with record.py offer create and shows the presentation record.py offer present prints. An offer in your own words, or any other question, can't be approved: the hook refuses the edit that follows. The plan report is built only on Show review.html first or when the owner asks for it. One-ticket work, a small visible change included, gets one ticket and its offer.

Lightweight edit: exact, low-risk, no observable behavior, architecture, interface, dependency or security change. A visible behavior change, however small and fully specified, is not lightweight: it gets the one-ticket offer. A lightweight edit gets the minimal offer (its files, its check and a one-line why) through workflow-specify step 5, and is still asked the one question. After the answer, verify, review once per AUDIT.md, then short workflow-deliver.

5. The two gates

Agreement, then the offer and the one question: VISUAL-REVIEW.md owns both. STOP for user approval before application code. Answering grilling questions is not agreement, and a generic "go" is not approval of an unseen plan.

Always

  • Behaviour questions: grilling; classification: §3. Evidence: GROUNDING.md. Specialists: SPECIALISTS.md (SETUP.md). Reviews and necessity check: AUDIT.md. Output: ARTIFACTS.md.
  • Never stage, commit, push, branch, reset, deploy or publish unasked.
  • Pass workers the output folder, write scope, rules and UI authority. Read applicable instructions whole; do not load unrelated references.
  • Approval waits: present once, end the turn; no polling. Host goals: TASK-FOLDER.md "Approval waits and host goals". End with Running, Waiting on you or Done; do not end running work unless the host will wake you. When an offer waits for the answer, its presentation ends the reply and its question is the state line: put nothing after it.
  • Write requested handoffs from the checkpoint (CHATGPT-PROMPTS.md).
  • When needed: PDF-DELIVERABLE.md, OPTIONAL-HELPERS.md.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/gazohhh/workflow-axa/workflow">View workflow on skillZs</a>