| name | project-skill-author |
| description | v0.1.8 - Create a project-specific Codex skill package with native Codex defaults and master-grade governance; use when building reusable skill folders, onboarding skills, coordination skills, or general project skills that should ship as native Codex skills (not API-only), including optional paired reviewer skills for high-freedom outputs. |
Project Skill Author
Trigger and Scope (Required)
Use this skill when you must design or update a reusable, project-level skill
package for external or cross-team users. This is not for single prompts or
one-off task guides.
In scope: project-wide skills, onboarding skills, coordination skills, and
multi-domain guidance with auditable workflows.
Out of scope: pure API integration skills, internal-only prompts, or
single-use task scripts.
Unless the user explicitly asks for another runtime, treat Codex as the target
runtime and produce a Codex-compatible skill package by default.
Skill Philosophy (Master-Level)
- Skills are reusable decision systems, not one-off prompts.
- Optimize for stable outputs under changing context.
- Prefer clear constraints over elaborate instructions.
- Minimize cognitive load and context size.
- Always separate workflow from reference material.
- Guard against hallucinated APIs or fake capabilities.
- Treat defaults as policy: safe, conservative, and explicit.
- Treat
name and description as trigger-critical metadata, not decoration.
- Derive domain-specific workflows from expert practice, not generic templates.
- Make the workflow justify itself with evidence and tradeoffs.
- Ensure outputs are actionable without follow-up prompting.
- Avoid shallow coverage; favor depth in the chosen domain.
- When a skill represents a long-lived actor, encode its operating identity with
an AGENTS-style role charter rather than leaving behavior implicit.
- Prefer decision artifacts that can be audited later.
- Maintain a clear separation between “guidance” and “evidence.”
- Keep
SKILL.md lean; move long Codex-native rules and domain libraries into
references/.
- Do not default every capability to a creator/reviewer pair.
- Split a capability into paired skills only when production and acceptance are
materially different phases with stable, reusable quality criteria.
Optional Paired Skill Pattern
- Start with one skill by default.
- Add a paired reviewer skill when the output is high freedom, quality is not
obvious at a glance, small misses materially reduce value, or the likely
failure modes are clear enough to encode as an acceptance gate.
- Keep a single skill when the work is mechanical, format-bound, or easy to
judge as correct/incorrect without nuanced review.
- Add a third shaping or publishing skill only when already-accepted output
still needs a distinct packaging step.
- Good candidates for paired skills include tutorials, blog posts, architecture
writing, prompt or skill design, code explanations, solution proposals, and
complex analysis.
- Weak candidates for paired skills include file renames, frontmatter fills,
fixed-format conversions, title normalization, simple routing, and other
short deterministic transforms.
Reference Map (Read As Needed)
references/codex-native-authoring.md
Read when defining package layout, agents/openai.yaml, official Codex
tooling, anti-clutter rules, or final compatibility validation.
references/domain-workflow-library.md
Read only the domain sections that materially affect the skill being authored.
references/role-charter-template.md
Read when the skill defines a reusable role, control-plane owner, reviewer,
operator, or other long-lived actor.
Codex Native Defaults
- Default runtime: Codex.
- Default package shape:
SKILL.md plus agents/openai.yaml; add
references/, scripts/, and assets/ only when repeated use,
determinism, or output reuse justifies them.
- Default trigger policy: put both capability and usage context in frontmatter
description, not only in the body.
- Default toolchain: prefer
init_skill.py, generate_openai_yaml.py, and
quick_validate.py when available; otherwise create the same structure
manually and note the fallback.
- Default context policy: keep
SKILL.md focused on workflow and guardrails;
load detailed reference material only as needed.
- Default size target: keep
SKILL.md under 500 lines and under 5k words when
feasible; start splitting content into references/ before crossing that
threshold.
Master Workflow (Decision-Grade)
- Clarify mission and audience.
- Define project purpose, primary users, and core outcomes.
- Capture at least 2 concrete usage examples or trigger prompts when
possible; use them to shape package contents and frontmatter.
- If the skill is role-like, define the role charter boundary using an
AGENTS-style structure and 12 role-specific
Why / How / Check golden
rules by default.
- Identify domain scope.
- Independently infer domains that materially affect usage.
- Read only the relevant sections from
references/domain-workflow-library.md.
- Map each domain to a high-signal workflow and explicitly note why it fits.
- Decide whether the capability should stay single-skill or split into
production and acceptance skills; only split when the acceptance gate is
meaningfully different from generation.
- Define the skill name and placement.
- Use a lowercase hyphen name under 64 chars.
- Default placement is
skills/<skill-name> unless a project path is
specified.
- Define runtime compatibility.
- Default to Codex unless the user explicitly names another runtime.
- State whether the package must be Codex-native only or cross-runtime.
- Add versioning to frontmatter.
- Prefix
description with a semantic version such as v0.1.0 - ....
- If no version exists yet, start at
v0.1.0.
- Bump the patch version once per completed commit that updates the skill.
- Design the package structure.
- Use the Codex-native anatomy in
references/codex-native-authoring.md.
- Keep
SKILL.md lean and move detailed material into references/.
- Treat 500 lines and 5k words as the default
SKILL.md ceiling; split
detailed rules, schemas, examples, and domain libraries into references/
before the body grows past that range.
- Create
scripts/, references/, and assets/ only when they clearly
reduce repeated reasoning or manual rewriting.
- For role-like skills, keep the concise charter in
SKILL.md and use
references/role-charter-template.md when reusable scaffolding helps.
- Draft
SKILL.md with explicit triggers and workflow.
- Use imperative steps.
- Keep trigger-critical "what/when to use" language in
description, not
only in body sections.
- Make any role charter explicit instead of scattering it across workflow
prose.
- Add Codex metadata and toolchain hooks.
- For Codex targets, generate or update
agents/openai.yaml.
- Define
display_name, short_description, and default_prompt from the
actual skill intent rather than placeholders.
- Prefix
short_description with the current skill version so the UI
metadata stays traceable to the package revision.
- Add icons, brand color, dependencies, or policy flags only when supported
by real assets or requirements.
- Prefer official Codex tooling when available.
- Provide defaults and guardrails.
- Include safe defaults that reduce user input and prevent misuse.
- Validate and scope.
- Ensure no secrets, no fabricated facts, and references one level deep.
- Run the Codex-native compatibility checklist in
references/codex-native-authoring.md for Codex-targeted skills.
- Verify domain fitness.
- Confirm the workflow reflects expert practice in the relevant domains.
- Produce a fit-for-purpose template.
- Provide a short, standard, and strict output format.
- Add verification hooks.
- Define how correctness and safety are validated for the skill.
- Forward-test complex or high-reuse skills on realistic tasks when the
validation cost is justified.
Design Layers (Use As Needed)
- Trigger Layer:
name + description; clear "when to use" signals.
- Workflow Layer: concise steps with minimal ambiguity.
- Codex Adapter Layer:
agents/openai.yaml and other runtime-facing metadata.
- Reference Layer: detailed docs in
references/.
- Asset Layer: templates, examples, or boilerplate in
assets/.
- Script Layer: deterministic or repeated logic in
scripts/.
Required Inputs (Minimal)
- Project name and one-sentence purpose
- Target users (role and context)
- Primary outcomes or workflows
- Delivery environment or distribution target
- Target runtime or agent platform if it is not Codex
Defaults (Use Unless User Specifies)
- Domain scope: Engineering + Product by default; add Security, Data, AI, or
other domains only when the task justifies them.
- Runtime target: Codex.
- Skill placement:
skills/<skill-name> unless a project path is specified.
- Package layout:
SKILL.md plus agents/openai.yaml by default.
- References: add only when details are needed repeatedly.
- Size target: keep
SKILL.md under 500 lines and under 5k words when
feasible; split before the body crosses that range.
- Resource selection: start from concrete usage examples and add resource
folders only when they reduce repeated reasoning or manual rewriting.
- Trigger policy: put "what the skill does" and "when to use it" in
frontmatter
description.
- Codex metadata: generate
agents/openai.yaml by default for Codex-facing
skills, populating display_name, short_description, and default_prompt.
- Metadata versioning: include the current semantic version in
interface.short_description.
- Initialization mode: use official Codex scaffold tools when available;
otherwise create a manual-compatible package.
- Validation mode: use official Codex validation tools when available;
otherwise run the checklist in
references/codex-native-authoring.md.
- Auxiliary docs: do not create README, changelog, or install docs unless
explicitly required.
- Role charter: required when the skill defines a reusable long-lived actor.
- Role golden rules: required for role-like skills; default to 12 concise
Why / How / Check rules unless the role is too small to justify that many.
- Pairing mode: default to a single skill; add a reviewer skill only when the
acceptance phase has distinct quality gates worth reusing across tasks.
Failure Modes to Avoid
- Overfitting the skill to a single project.
- Mixing workflow and reference content.
- Overly verbose
SKILL.md that bloats context.
- Letting
SKILL.md grow past roughly 500 lines or 5k words without splitting
heavy material into references/.
- Missing guardrails that allow unsafe edits.
- Unclear triggers that cause accidental activation.
- Burying "when to use" only in the body instead of frontmatter
description.
- Omitting or staling
agents/openai.yaml for a Codex-targeted skill.
- Letting
interface.short_description drift from the current skill version.
- Shipping a package layout that Codex cannot consume without manual adaptation.
- Adding
scripts/, references/, or assets/ without a repeated-use case.
- Stuffing long reference content back into
SKILL.md instead of using
progressive disclosure.
- Creating auxiliary docs that native Codex skills do not need.
- Filling
openai.yaml with guessed icons, dependencies, or branding.
- Relying on a single generic workflow for all domains.
- Using “best practices” language without concrete operational steps.
- Missing domain-specific risk controls or validation gates.
- Shipping a skill without verification or example usage.
- Omitting audit trails for high-risk workflows.
- Mechanically creating a reviewer skill for every capability without a real
acceptance-stage need.
- Forcing one skill to both generate and accept high-freedom output when the
failure modes are subtle but predictable.
Output Tiers (Pick One)
Short: skill name, placement, minimal workflow, guardrails.
Standard: add required inputs, defaults, and output format.
Strict: add validation rules and risk notes.
Decision Rubric (Use to Choose a Workflow)
- Risk: higher risk requires stricter validation and guardrails.
- Reuse: higher reuse favors assets and references.
- Fragility: fragile operations favor scripts and low degrees of freedom.
- Audience: external users require clearer constraints and examples.
- Regulatory: regulated domains require explicit compliance hooks.
- Learning Value: prefer workflows that create measurable feedback loops.
Validation Hooks (Required for Strict)
- Provide a minimal “how to verify” section.
- Include a negative test or failure case when risk is non-trivial.
- Require explicit “unknowns” where facts are missing.
- Require a lightweight evidence note of what was checked, by whom, and when.
- For Codex-targeted skills, run the compatibility checklist in
references/codex-native-authoring.md.
- If
SKILL.md is unusually long, explain why it could not be split further
without harming usability.
- For Codex-targeted skills, verify that
interface.short_description includes
the current skill version.
- When the skill is role-like, verify that the charter uses an AGENTS-style
structure and makes mission, boundaries, permission model, execution rules,
outputs, and escalation conditions explicit.
- When the skill is role-like, verify that the golden rules are role-specific,
operational, and written in
Why / How / Check form.
AGENTS-Style Role Charter Standard (Required For Role-Like Skills)
When a skill defines a long-lived actor rather than a one-shot helper, include
an AGENTS-style role charter as part of the skill package.
The charter should feel structurally similar to a strong project AGENTS.md:
boundary-first, measurable, and operational rather than persona-heavy.
Minimum sections:
Overview
Core Principles
Golden Rules (Why / How / Check)
Mission & Non-Negotiables
Ownership Boundaries
Permission Model
Inputs
Outputs
Execution Rules
Handoff & Escalation
Quality Bar
Done Signal
Risks & Open Questions
Placement rules:
- Keep the concise charter in
SKILL.md when the role is central to trigger
and workflow.
- Move reusable expansions or starter templates into
references/role-charter-template.md.
- Do not bury charter elements only inside examples or optional notes.
Domain-Specific Variants (Optional)
- API Integration
- Data / Analytics
- DevTools / CLI
- Docs / Onboarding
- Collaboration
- Compliance / Legal
- Security
- Quality / Testing
- Architecture / Systems Design
- Other domains that materially affect the workflow
The agent must decide which domains apply by reasoning from the project and
task context. Use field-proven workflows as the basis and read only the
relevant sections in references/domain-workflow-library.md.
References
references/codex-native-authoring.md
Codex-native anatomy, metadata rules, official tooling, anti-clutter rules,
and compatibility validation checklist.
references/domain-workflow-library.md
Domain workflow library for API, data, AI, security, docs, product, infra,
compliance, testing, growth, legal, support, observability, accessibility,
localization, payments, procurement, privacy, incident response,
architecture, governance, and release management.
references/role-charter-template.md
Starter template for AGENTS-style role charters inside reusable skills.
SKILL.md Body Template (Use Imperatives)
# <Skill Title>
## Workflow
1. ...
## Required Inputs
- ...
## Bundled Resources
- ...
## Output Format
...
## Guardrails
- ...
Output Format (This Skill)
## Skill Name and Placement
## Runtime Compatibility
## Package Layout
## Required Inputs Missing
## Files Created
## Metadata Plan
## Validation Plan
## Open Questions
Guardrails
- Do not include API keys, tokens, or secrets.
- Do not invent endpoints or behaviors; mark unknowns as TODO.
- Do not provide bypass or evasion guidance for security controls.
- Keep the skill self-contained and minimal.
- Do not ship a Codex-targeted skill without trigger-quality frontmatter.
- Do not omit the Codex metadata adapter unless the user explicitly asks for a
metadata-free package.