suede-codex-fleet

Name: suede-codex-fleet
Author: JasonColapietro

44stars8forks

How to install suede-codex-fleet?

jasoncolapietro/suede-creator-skills/suede-codex-fleet

$npx -y skills add JasonColapietro/suede-creator-skills --skill suede-codex-fleet

Installs into the current project.

Summary

Claude-directed parallel OpenAI Codex CLI worker fleet for bulk generation. Use when a job is high-volume, well-specified, and splits into independent worker-sized tasks (content batches, test generation, bulk refactors) and Codex CLI is installed and logged in. Claude decomposes

Files · 1

View on GitHub

SKILL.md

1	# Suede Fable Fleet
2
3	The Suede Fable Fleet: a high-end Claude model is the admiral — it decomposes, briefs, and reviews — and parallel OpenAI Codex CLI workers are the fleet. The skill id and command stay `suede-codex-fleet` on purpose: GitHub search, skill marketplaces, and MCP catalogs match the terms people actually type — Codex CLI orchestration, codex exec, multi-agent worker fleet — not the brand name. Do not rename the folder or frontmatter `name` to match the brand.
4
5	## When to use this skill instead of related skills
6
7	- suede-codex-fleet (this skill): offload high-volume, well-specified generation to Codex CLI workers; Claude plans, briefs, and reviews
8	- suede-agent-teams: multi-lane Claude agents coordinating one complex code change with gates and handoffs
9	- suede-copy / johnny-suede-write: Claude writes the copy itself; right choice when volume is low and judgment density is high
10
11	Core principle: Claude tokens buy judgment, Codex tokens buy volume. Clear spec + high volume goes to Codex. Fuzzy spec or expensive-if-wrong stays with Claude. Nothing ships unreviewed.
12
13	## Preflight (run before first spawn)
14
15	1. `which codex && codex --version` — CLI present (validated against codex-cli 0.138.0).
16	2. `codex login status` — must show logged in (your ChatGPT subscription pays for the run).
17	3. Workspace has an `AGENTS.md` at its root. Codex auto-loads it; it carries voice, context, hard bans, and output conventions so briefs stay short. If missing, write it first — that is the highest-leverage file in the system.
18	4. Workspace has `briefs/` and `out/` directories (create as needed).
19
20	## The loop
21
22	1. Decompose. Split the job into independent worker-sized tasks. Independent means: no worker needs another worker's output.
23	2. Brief. One markdown file per task in `briefs/`. Codex never sees the Claude conversation, so each brief is self-contained: job, inputs (file paths), exact deliverable, acceptance criteria it must self-check, and the exact output path in `out/`.
24	3. Spawn. One `codex exec` per brief, in parallel, in the background:
25
26	```bash
27	codex exec -C <workspace> --sandbox workspace-write --skip-git-repo-check \
28	-o <workspace>/out/<run-name>-final-message.txt \
29	"Read AGENTS.md at the workspace root, then execute the brief at briefs/<brief>.md exactly. Write the deliverable to the output file the brief names, run the brief's acceptance-criteria self-check, and state pass/fail per criterion in your final message."
30	```
31
32	- `-C` sets the worker's root; `--skip-git-repo-check` is required outside git repos.
33	- `--sandbox workspace-write` only. Never `danger-full-access`. Workers write files; they do not push, deploy, or touch secrets.
34	- Leave the model default unless explicitly asked to override with `-m`.
35	4. Review gate (Claude, mandatory). Read every `out/` file. Check against the brief's acceptance criteria and the AGENTS.md hard bans. Worker self-checks are evidence, not verdicts. If the output fails 0 acceptance criteria but has surface defects (typos, formatting, a wrong label), Claude edits the file directly; do not respawn for a comma.
36	5. Delta, don't regenerate. If the output fails 1-2 acceptance criteria, send a one-line correction: `codex exec resume <session-id> "<delta>"` (session id is printed at run start; `resume --last` is ambiguous with parallel runs). If it fails 3+ criteria or violates an AGENTS.md hard ban, respawn with the delta appended to the brief. Regenerating from scratch wastes the subscription and loses what was right. Correction budget per output: up to three genuinely different fixes — each attempt must change the diagnosis or the strategy, never rerun the last one. Stop early when the same root cause repeats across attempts; report the repeating cause and let the user pick the next move.
37	6. Ship. Claude assembles the reviewed survivors into the final deliverable. Report what was spawned, what passed, what got fixed.
38
39	## Brief template
40
41	```markdown
42	# Brief <id> — <task name>
43
44	Read `AGENTS.md` in the workspace root first. This brief only adds the task.
45
46	## Job
47	<one paragraph: what and why>
48
49	## Inputs
50	<file paths the worker must read>
51
52	## Deliverable
53	<exact structure, counts, variants, labels>
54
55	## Acceptance criteria (self-check before finishing)
56	<numbered, mechanically checkable: limits, bans, required elements>
57
58	## Output
59	Write to `out/<file>.md`

People who liked this also viewed

6 picks

Type

suede-codex-fleet

44stars8forks

View on GitHub

How to install suede-codex-fleet?

jasoncolapietro/suede-creator-skills/suede-codex-fleet

$npx -y skills add JasonColapietro/suede-creator-skills --skill suede-codex-fleet

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Suede Fable Fleet
2
3	The Suede Fable Fleet: a high-end Claude model is the admiral — it decomposes, briefs, and reviews — and parallel OpenAI Codex CLI workers are the fleet. The skill id and command stay `suede-codex-fleet` on purpose: GitHub search, skill marketplaces, and MCP catalogs match the terms people actually type — Codex CLI orchestration, codex exec, multi-agent worker fleet — not the brand name. Do not rename the folder or frontmatter `name` to match the brand.
4
5	## When to use this skill instead of related skills
6
7	- suede-codex-fleet (this skill): offload high-volume, well-specified generation to Codex CLI workers; Claude plans, briefs, and reviews
8	- suede-agent-teams: multi-lane Claude agents coordinating one complex code change with gates and handoffs
9	- suede-copy / johnny-suede-write: Claude writes the copy itself; right choice when volume is low and judgment density is high
10
11	Core principle: Claude tokens buy judgment, Codex tokens buy volume. Clear spec + high volume goes to Codex. Fuzzy spec or expensive-if-wrong stays with Claude. Nothing ships unreviewed.
12
13	## Preflight (run before first spawn)
14
15	1. `which codex && codex --version` — CLI present (validated against codex-cli 0.138.0).
16	2. `codex login status` — must show logged in (your ChatGPT subscription pays for the run).
17	3. Workspace has an `AGENTS.md` at its root. Codex auto-loads it; it carries voice, context, hard bans, and output conventions so briefs stay short. If missing, write it first — that is the highest-leverage file in the system.
18	4. Workspace has `briefs/` and `out/` directories (create as needed).
19
20	## The loop
21
22	1. Decompose. Split the job into independent worker-sized tasks. Independent means: no worker needs another worker's output.
23	2. Brief. One markdown file per task in `briefs/`. Codex never sees the Claude conversation, so each brief is self-contained: job, inputs (file paths), exact deliverable, acceptance criteria it must self-check, and the exact output path in `out/`.
24	3. Spawn. One `codex exec` per brief, in parallel, in the background:
25
26	```bash
27	codex exec -C <workspace> --sandbox workspace-write --skip-git-repo-check \
28	-o <workspace>/out/<run-name>-final-message.txt \
29	"Read AGENTS.md at the workspace root, then execute the brief at briefs/<brief>.md exactly. Write the deliverable to the output file the brief names, run the brief's acceptance-criteria self-check, and state pass/fail per criterion in your final message."
30	```
31
32	- `-C` sets the worker's root; `--skip-git-repo-check` is required outside git repos.
33	- `--sandbox workspace-write` only. Never `danger-full-access`. Workers write files; they do not push, deploy, or touch secrets.
34	- Leave the model default unless explicitly asked to override with `-m`.
35	4. Review gate (Claude, mandatory). Read every `out/` file. Check against the brief's acceptance criteria and the AGENTS.md hard bans. Worker self-checks are evidence, not verdicts. If the output fails 0 acceptance criteria but has surface defects (typos, formatting, a wrong label), Claude edits the file directly; do not respawn for a comma.
36	5. Delta, don't regenerate. If the output fails 1-2 acceptance criteria, send a one-line correction: `codex exec resume <session-id> "<delta>"` (session id is printed at run start; `resume --last` is ambiguous with parallel runs). If it fails 3+ criteria or violates an AGENTS.md hard ban, respawn with the delta appended to the brief. Regenerating from scratch wastes the subscription and loses what was right. Correction budget per output: up to three genuinely different fixes — each attempt must change the diagnosis or the strategy, never rerun the last one. Stop early when the same root cause repeats across attempts; report the repeating cause and let the user pick the next move.
37	6. Ship. Claude assembles the reviewed survivors into the final deliverable. Report what was spawned, what passed, what got fixed.
38
39	## Brief template
40
41	```markdown
42	# Brief <id> — <task name>
43
44	Read `AGENTS.md` in the workspace root first. This brief only adds the task.
45
46	## Job
47	<one paragraph: what and why>
48
49	## Inputs
50	<file paths the worker must read>
51
52	## Deliverable
53	<exact structure, counts, variants, labels>
54
55	## Acceptance criteria (self-check before finishing)
56	<numbered, mechanically checkable: limits, bans, required elements>
57
58	## Output
59	Write to `out/<file>.md`

People who liked this also viewed

6 picks

Type