- You are authoring a fresh, project-tailored workflow from the bundled template — never run
one fixed script blindly. Fill every
FILL: marker before running.
- The workflow architecture is load-bearing. Leave the schemas and the entire control-flow
section of the template unchanged: Gates → Find → Dedup → Verify → Critic → report. The one
exception is
MAX_PARALLEL (the concurrency cap) — you may tune its value, but never remove the
gate or the gAgent wrapper that enforces it.
- Keep
meta a pure literal (no variables, no interpolation) or the Workflow tool rejects it.
- Finders and verifiers are read-only. They must never modify, create, or delete files, and
never run builds or installs. Only gate agents execute commands (build may write
dist/).
- Never quote secret values in findings or evidence — report type and location only, redact
the value with
<REDACTED>.
- Confirm scope and warn the user about the agent count (typically 40–80) before running.
- These instructions authorize calling the Workflow tool — no separate opt-in is needed, but
the scope confirmation and scale warning above are mandatory first.
Go-Live Audit
Inputs
$request: Optional repo path, launch context, or focus guidance such as audit ~/work/myapp before public beta or light pass, security only
Goal
Produce a credible launch-readiness verdict by generating and running a project-tailored
pre-launch audit workflow:
- Gates — typecheck / test / lint / build, run in parallel
- Find — one read-only finder agent per applicable audit dimension
- Dedup — an agent merges findings that share a root cause
- Verify — adversarial agents try to refute every finding; blockers and highs get a code
lens plus a spec lens
- Critic — names uncovered areas, spawns follow-up finders, verifies them
- A structured GO / NO-GO / GO-WITH-FIXES report
All phases fan out through a global concurrency gate (MAX_PARALLEL, default 6) so a wide sweep
never puts more than that many agents in flight at once — the total agent count is unchanged, but
they run in bounded waves so Claude's API rate limits are not tripped.
Step 0: Confirm scope
Establish, asking the user only if unclear:
- Which repo (absolute path). Default to the current working directory.
- Launch context — what "go live" means here: public release, first paying customers,
internal GA. This shapes what counts as a blocker.
Do not ask more than necessary; infer from the repo when you can.
Success criteria: The target repo and launch context are explicit.
Step 1: Discover the project
Read the repo enough to fill the template accurately. Gather:
- Build/test commands — inspect
package.json scripts (and whether it is a pnpm/npm/yarn/bun
workspace, turbo, etc.), or Cargo.toml, go.mod, Makefile, pyproject.toml. Determine the
exact typecheck / test / lint / build command run from the repo root
(references/workflow-template.js lists common commands per stack).
- Layout — apps/packages or src tree; the modules that exist.
- Authority docs — a spec/PRD, root
CLAUDE.md, README, docs/. Note the order of
authority (e.g. "spec > CLAUDE.md > code").
- Product purpose — one paragraph: what it is, who uses it.
- Hard invariants — rules the project must never violate (from the spec/CLAUDE.md): privacy
guarantees, security/tenancy rules, "soft-fail everywhere", schema conventions, anti-features.
If none are documented, derive sensible universal ones (data integrity, authz on sensitive
reads, no committed secrets, green build/test/lint, no TODO/placeholder in shipping code).
Use Explore/Glob/Grep/Read freely here; this step determines audit quality.
Success criteria: Stack commands, layout, authority docs, product purpose, and hard rules are
all known well enough to fill the template without guessing.
Step 2: Choose the finder dimensions
From references/dimensions.md, pick the 8–16 dimensions that actually apply to this repo —
a CLI does not need multi-tenancy; a static site does not need money-math. Add any
project-specific dimensions the catalog misses.
Each finder must name concrete files/dirs to read and specific, verifiable checks — not
"review X for quality". Tie the worst checks to the project's hard rules.
Success criteria: Every chosen dimension has a real surface in this repo, and every finder
prompt is self-contained with concrete targets and checks.
Step 3: Fill the template
Copy references/workflow-template.js and fill every FILL: marker:
ROOT — absolute repo path.
meta — keep it a pure literal (no variables/interpolation) or the Workflow tool rejects
it. Update description to the project; keep the phase titles.
PRODUCT_CONTEXT, HARD_RULES — from Step 1.
SEVERITY_DEFS — replace the blocker examples with this product's launch-stoppers.
FINDERS — the dimensions from Step 2.
STACK — the detected commands; set any gate to null to skip it.
Leave the schemas and the entire control-flow section unchanged — they are generic. The only
value you may adjust there is MAX_PARALLEL (default 6): the cap on how many agents run at once.
Lower it (e.g. 4) if the account hits rate limits; raise it (e.g. 8–10) on a high-limit account for
more throughput. Do not set it to 1–2 (that serializes the audit) and never delete the gate itself.
Success criteria: No FILL: marker remains, meta is a pure literal, and the control-flow
section is byte-identical to the template.
Step 4: Warn about scale, then run
Tell the user roughly how many agents this will spawn (finders + ~1–2 verifiers each + gates +
critic round — typically 40–80 agents total, but at most MAX_PARALLEL (default 6) run at once,
so the audit proceeds in bounded waves rather than hammering the API) so the cost and pacing are
expected. Then call the Workflow tool with script set to the filled-in template. Watch
progress via /workflows.
To iterate on the script, edit the saved scriptPath the tool returns and re-invoke Workflow
with {scriptPath} rather than resending the whole script.
Success criteria: The user knows the expected scale before the workflow starts, and the
workflow runs to completion.
Step 5: Report
The workflow returns { gates, findings, finderNotes, stats }. Render a launch readiness report:
- Verdict — GO / NO-GO / GO-WITH-FIXES, driven by gate failures and confirmed blockers.
- Gates — pass/fail per gate with failure excerpts.
- Findings grouped by
status then severity. Lead with confirmed blockers and highs.
Show uncertain separately (verifiers split). Omit or collapse rejected (verifiers refuted
them) — mention the count, not the detail.
- For each surviving finding: severity, file:line, the issue, and the verifier reasoning. Note
alsoReportedBy when multiple finders flagged it.
- Coverage —
stats (raw vs deduped counts, finders reported, critic gap areas) and
anything the critic flagged as uncovered.
Success criteria: The user can decide go/no-go from the report without re-deriving any
finding from scratch.
Guardrails
- This skill audits and reports; it never fixes. Do not edit application code as part of this
skill — hand confirmed findings to
bugfix or the user.
- Keep finders and verifiers read-only; keep that discipline when editing prompts.
- Scale to the ask: fewer finders + code-only verification for a light pass; the full 16 +
dual-lens verification + critic round for "audit everything".
- Severity is go-live-framed, not abstract — a blocker is something that must be fixed before
this launch, judged against the project's hard rules.
- Never include secret values in findings, evidence, or the report — location and type only.
- Do not modify settings files or install anything on the user's behalf.
When To Load References
references/workflow-template.js
The parameterized workflow. Read it before generating; copy it, fill every FILL: marker, run it.
references/dimensions.md
Catalog of audit dimensions to choose finders from, plus the finder-prompt style guide.
Output Contract
Report:
- verdict — GO / NO-GO / GO-WITH-FIXES with one-line justification
- gate results with failure excerpts
- confirmed blockers and highs with file:line, evidence, and verifier reasoning
- uncertain findings (verifiers split) listed separately
- rejected-findings count (not detail)
- coverage stats and critic-identified gaps
1---
2name: go-live-audit
3version: 1.1.0
4description: |
5 Author and run a project-TAILORED pre-launch audit as a bounded multi-agent workflow — a real launch
6 verdict, not a checklist skim: fill the bundled workflow template from the repo's real stack and hard
7 rules, then run Gates (typecheck / test / lint / build in parallel) → Find (one read-only finder per
8 applicable audit dimension) → Dedup (merge shared-root-cause findings) → Verify (adversarial agents try
9 to refute every finding; blockers get a code lens plus a spec lens) → Critic (name uncovered areas,
10 spawn follow-up finders) → a GO / NO-GO / GO-WITH-FIXES report, all fanned out under a MAX_PARALLEL cap.
11 Finders and verifiers stay read-only — the skill audits and reports, never fixes; a gate that didn't
12 actually run is treated as a blocker, never counted clean; secret values are redacted to type and
13 location. Confirms scope and warns about the agent count (typically 40-80) before running.
14allowed-tools:
15 - Bash
16 - Read
17 - Glob
18 - Grep
19 - Agent
20 - AskUserQuestion
21 - Edit
22 - Workflow
23effort: high
24argument-hint: "[repo path or launch context]"
25arguments:
26 - request
27when_to_use: |
28 Use when the user asks to audit a repo before launch, shipping, or release. Examples: "go live
29 audit", "pre-launch audit", "launch readiness", "are we ready to ship", "final review before
30 release", "deep audit of <repo>". It spawns dozens of agents over the whole repository, so confirm
31 scope first. Do NOT use for a quick check of the current branch diff — redirect to
32 branch-review-before-pr or the forked find-bugs for that; this is a whole-repo, multi-agent launch gate.
33metadata:
34 source: https://skillmd.com/skills/ulpi-io/go-live-audit
35---
36
37<EXTREMELY-IMPORTANT>
38This skill authors and runs a multi-agent audit workflow. Non-negotiable rules:
39
401. You are *authoring* a fresh, project-tailored workflow from the bundled template — never run
41 one fixed script blindly. Fill every `FILL:` marker before running.
422. The workflow architecture is load-bearing. Leave the schemas and the entire control-flow
43 section of the template unchanged: Gates → Find → Dedup → Verify → Critic → report. The one
44 exception is `MAX_PARALLEL` (the concurrency cap) — you may tune its value, but never remove the
45 gate or the `gAgent` wrapper that enforces it.
463. Keep `meta` a pure literal (no variables, no interpolation) or the Workflow tool rejects it.
474. Finders and verifiers are read-only. They must never modify, create, or delete files, and
48 never run builds or installs. Only gate agents execute commands (build may write `dist/`).
495. Never quote secret values in findings or evidence — report type and location only, redact
50 the value with `<REDACTED>`.
516. Confirm scope and warn the user about the agent count (typically 40–80) before running.
527. These instructions authorize calling the Workflow tool — no separate opt-in is needed, but
53 the scope confirmation and scale warning above are mandatory first.
54</EXTREMELY-IMPORTANT>
55
56# Go-Live Audit
57
58## Inputs
59
60- `$request`: Optional repo path, launch context, or focus guidance such as `audit ~/work/myapp before public beta` or `light pass, security only`
61
62## Goal
63
64Produce a credible launch-readiness verdict by generating and running a project-tailored
65pre-launch audit workflow:
66
67- **Gates** — typecheck / test / lint / build, run in parallel
68- **Find** — one read-only finder agent per applicable audit dimension
69- **Dedup** — an agent merges findings that share a root cause
70- **Verify** — adversarial agents try to *refute* every finding; blockers and highs get a code
71 lens plus a spec lens
72- **Critic** — names uncovered areas, spawns follow-up finders, verifies them
73- A structured GO / NO-GO / GO-WITH-FIXES report
74
75All phases fan out through a global concurrency gate (`MAX_PARALLEL`, default 6) so a wide sweep
76never puts more than that many agents in flight at once — the total agent count is unchanged, but
77they run in bounded waves so Claude's API rate limits are not tripped.
78
79## Step 0: Confirm scope
80
81Establish, asking the user only if unclear:
82
83- **Which repo** (absolute path). Default to the current working directory.
84- **Launch context** — what "go live" means here: public release, first paying customers,
85 internal GA. This shapes what counts as a *blocker*.
86
87Do not ask more than necessary; infer from the repo when you can.
88
89**Success criteria**: The target repo and launch context are explicit.
90
91## Step 1: Discover the project
92
93Read the repo enough to fill the template accurately. Gather:
94
95- **Build/test commands** — inspect `package.json` scripts (and whether it is a pnpm/npm/yarn/bun
96 workspace, turbo, etc.), or `Cargo.toml`, `go.mod`, `Makefile`, `pyproject.toml`. Determine the
97 exact typecheck / test / lint / build command run from the repo root
98 (`references/workflow-template.js` lists common commands per stack).
99- **Layout** — apps/packages or src tree; the modules that exist.
100- **Authority docs** — a spec/PRD, root `CLAUDE.md`, `README`, `docs/`. Note the order of
101 authority (e.g. "spec > CLAUDE.md > code").
102- **Product purpose** — one paragraph: what it is, who uses it.
103- **Hard invariants** — rules the project must never violate (from the spec/CLAUDE.md): privacy
104 guarantees, security/tenancy rules, "soft-fail everywhere", schema conventions, anti-features.
105 If none are documented, derive sensible universal ones (data integrity, authz on sensitive
106 reads, no committed secrets, green build/test/lint, no TODO/placeholder in shipping code).
107
108Use Explore/Glob/Grep/Read freely here; this step determines audit quality.
109
110**Success criteria**: Stack commands, layout, authority docs, product purpose, and hard rules are
111all known well enough to fill the template without guessing.
112
113## Step 2: Choose the finder dimensions
114
115From `references/dimensions.md`, pick the **8–16 dimensions that actually apply** to this repo —
116a CLI does not need multi-tenancy; a static site does not need money-math. Add any
117project-specific dimensions the catalog misses.
118
119Each finder must name **concrete files/dirs** to read and **specific, verifiable checks** — not
120"review X for quality". Tie the worst checks to the project's hard rules.
121
122**Success criteria**: Every chosen dimension has a real surface in this repo, and every finder
123prompt is self-contained with concrete targets and checks.
124
125## Step 3: Fill the template
126
127Copy `references/workflow-template.js` and fill every `FILL:` marker:
128
129- `ROOT` — absolute repo path.
130- `meta` — keep it a **pure literal** (no variables/interpolation) or the Workflow tool rejects
131 it. Update `description` to the project; keep the phase titles.
132- `PRODUCT_CONTEXT`, `HARD_RULES` — from Step 1.
133- `SEVERITY_DEFS` — replace the blocker examples with this product's launch-stoppers.
134- `FINDERS` — the dimensions from Step 2.
135- `STACK` — the detected commands; set any gate to `null` to skip it.
136
137Leave the schemas and the entire control-flow section **unchanged** — they are generic. The only
138value you may adjust there is `MAX_PARALLEL` (default 6): the cap on how many agents run at once.
139Lower it (e.g. 4) if the account hits rate limits; raise it (e.g. 8–10) on a high-limit account for
140more throughput. Do not set it to 1–2 (that serializes the audit) and never delete the gate itself.
141
142**Success criteria**: No `FILL:` marker remains, `meta` is a pure literal, and the control-flow
143section is byte-identical to the template.
144
145## Step 4: Warn about scale, then run
146
147Tell the user roughly how many agents this will spawn (finders + ~1–2 verifiers each + gates +
148critic round — typically 40–80 agents *total*, but at most `MAX_PARALLEL` (default 6) run at once,
149so the audit proceeds in bounded waves rather than hammering the API) so the cost and pacing are
150expected. Then call the **Workflow** tool with `script` set to the filled-in template. Watch
151progress via `/workflows`.
152
153To iterate on the script, edit the saved `scriptPath` the tool returns and re-invoke Workflow
154with `{scriptPath}` rather than resending the whole script.
155
156**Success criteria**: The user knows the expected scale before the workflow starts, and the
157workflow runs to completion.
158
159## Step 5: Report
160
161The workflow returns `{ gates, findings, finderNotes, stats }`. Render a launch readiness report:
162
163- **Verdict** — GO / NO-GO / GO-WITH-FIXES, driven by gate failures and confirmed blockers.
164- **Gates** — pass/fail per gate with failure excerpts.
165- **Findings grouped by `status` then `severity`.** Lead with `confirmed` blockers and highs.
166 Show `uncertain` separately (verifiers split). Omit or collapse `rejected` (verifiers refuted
167 them) — mention the count, not the detail.
168- For each surviving finding: severity, file:line, the issue, and the verifier reasoning. Note
169 `alsoReportedBy` when multiple finders flagged it.
170- **Coverage** — `stats` (raw vs deduped counts, finders reported, critic gap areas) and
171 anything the critic flagged as uncovered.
172
173**Success criteria**: The user can decide go/no-go from the report without re-deriving any
174finding from scratch.
175
176## Guardrails
177
178- This skill audits and reports; it never fixes. Do not edit application code as part of this
179 skill — hand confirmed findings to `bugfix` or the user.
180- Keep finders and verifiers read-only; keep that discipline when editing prompts.
181- Scale to the ask: fewer finders + code-only verification for a light pass; the full 16 +
182 dual-lens verification + critic round for "audit everything".
183- Severity is go-live-framed, not abstract — a blocker is something that must be fixed before
184 *this* launch, judged against the project's hard rules.
185- Never include secret values in findings, evidence, or the report — location and type only.
186- Do not modify settings files or install anything on the user's behalf.
187
188## When To Load References
189
190- `references/workflow-template.js`
191 The parameterized workflow. Read it before generating; copy it, fill every `FILL:` marker, run it.
192- `references/dimensions.md`
193 Catalog of audit dimensions to choose finders from, plus the finder-prompt style guide.
194
195## Output Contract
196
197Report:
198
1991. verdict — GO / NO-GO / GO-WITH-FIXES with one-line justification
2002. gate results with failure excerpts
2013. confirmed blockers and highs with file:line, evidence, and verifier reasoning
2024. uncertain findings (verifiers split) listed separately
2035. rejected-findings count (not detail)
2046. coverage stats and critic-identified gaps