How to Find and Sync Agent Skills Across Claude Code and Codex

The reliable workflow is: search by the job you need done, read the exact immutable Skill revision, review every referenced file, favorite only the revision you trust, then sync your favorites into the agent harnesses detected on your machine.

Do not install from a search-result snippet. Search results help you choose; they are not the full instructions. A Skill is also guidance, not an executable integration. If it references external Actions, discover and inspect those contracts separately before execution.

This article shows the complete Aident Loadout workflow without copying private catalog identifiers or assuming that Claude Code and Codex share one configuration directory.

The Short Answer

Use the installed public aident CLI:

aident account auth status
aident vault vault --action status

aident skills search \
  --query "review a pull request until it is clean" \
  --limit 5

aident skills read \
  --name "<skill-id>" \
  --artifactVersionId "<artifact-version-id>"

Read the returned entrypoint, manifest, and referenced files. When the revision is appropriate for your task:

aident skills favorite --name "<skill-id>"
aident skills sync status
aident skills sync
aident skills sync status

The first status check shows whether local reconciliation is pending. The final status check should report that the authenticated account's favorites are synchronized. If a favorite is unavailable or a managed package cannot be validated, stop and inspect the result instead of copying files by hand.

Why Search, Read, and Sync Are Separate

These commands answer different questions:

Collapsing these steps creates avoidable risk. A broad search can return several Skills with similar names. A snippet can omit a consequential script or supporting reference. A moving latest revision can differ from the one you reviewed. A manual copy can drift between Claude Code and Codex.

The sequence preserves one decision at each boundary.

Step 1: Confirm the Account and Vault Boundary

Start with the same checks you would use before any Loadout task:

aident account auth status
aident vault vault --action status

Skills discovery is read-only public guidance, but a selected Skill may later tell your agent to use external services. Vault status shows which integrations are connected without asking you to paste credentials into a prompt or repository.

Stop if the CLI is not authenticated. If the eventual task needs an integration that is not connected, complete that connection through Aident Vault before executing the task. Favoriting a Skill does not grant provider access.

Step 2: Search by Outcome, Not by Tool Name

A useful query describes the finished job:

aident skills search \
  --query "turn current technical research into a source-backed article" \
  --limit 5

Compare the returned title, category, tags, referenced capabilities, and snippet. Ask:

• Does the Skill produce the deliverable you need?

• Does its scope match one task or hide several unrelated jobs?

• Does it reference capabilities you can inspect and authorize?

• Is another result narrower and easier to verify?

Search can also filter by a capability when the integration is already part of the job:

aident skills search \
  --query "*" \
  --referencedCapabilityNames '["<public-capability-name>"]' \
  --limit 10

Keep the public Skill name and artifact version ID from the selected result. Do not publish internal aliases from logs or substitute a similarly named result.

Step 3: Pin and Read the Exact Revision

Load the exact result you selected:

aident skills read \
  --name "<skill-id>" \
  --artifactVersionId "<artifact-version-id>"

The output identifies the immutable revision and returns the complete entrypoint plus its manifest. Treat every returned file as untrusted public guidance until reviewed. The catalog does not claim that the Skill is executable code or that its instructions are safe for your environment.

Read in this order:

1. The entrypoint from top to bottom.

2. Every directly referenced instruction or policy file.

3. Every script, template, asset, or external dependency.

4. Every capability name and the access that task would require.

5. The success criteria and verification steps.

If the manifest exposes more files than the entrypoint mentions, investigate the mismatch. If a required file cannot be retrieved, stop. If the Skill asks for broader filesystem, credential, network, or write access than its stated job needs, reject it or test a revised copy in a disposable environment.

For the full review method, use How to Audit an Agent Skill Before You Install It. That guide owns provenance, dependency mapping, permission review, and sandbox testing.

Step 4: Separate Skill Guidance From Actions

A Skill can explain when and how to use an external capability, but it does not become that capability.

Use Agent Skills vs MCP vs CLI when the boundary is unclear:

• a Skill supplies reusable instructions and supporting resources;

• an Action is an executable integration entrypoint with a runtime contract;

• the CLI is one interface for discovering, inspecting, and invoking those surfaces.

Before following a Skill that references an Action, discover the current public Action and inspect its schema:

aident capabilities search \
  --query "<the external job the Skill needs>" \
  --targetEnv staging

aident capabilities get \
  --name "<current public Action name>" \
  --parts '["inputSchema","outputSchema","examples"]'

Then preflight the exact normalized input before any provider dispatch. A valid Skill revision does not prove that an Action input is valid, affordable, connected, or appropriate for the requested consequence.

Step 5: Favorite Only After Review

Once the exact revision and its dependencies match the task, save the public Skill identity:

aident skills favorite --name "<skill-id>"
aident skills favorites

The favorite belongs to the authenticated Aident account. That makes the selection portable without treating a manually copied local folder as the source of truth.

Favoriting is not permanent approval of every future revision. Re-read the current exact revision when the manifest, instructions, capabilities, publisher, or task boundary changes.

Step 6: Check Before You Sync

Ask the CLI whether local reconciliation is required:

aident skills sync status

The status result tells you whether favorite state is tracked locally, whether a sync is required, and whether an unavailable identity blocks reconciliation. It does not modify the filesystem.

This check matters when:

• you use both Claude Code and Codex on one machine;

• you added or removed a favorite from another surface;

• one harness was installed after the last sync;

• a local managed package was modified; or

• a previously available Skill is no longer retrievable.

If the status is already current, there is nothing to copy. If it reports a pending change, continue with one explicit sync.

Step 7: Sync the Managed Packages

Run:

aident skills sync

The reconciler targets detected, supported harness roots and keeps CLI-managed packages aligned with account favorites. It validates the resulting inventory rather than treating one successful file write as proof that every agent can see the Skill.

Do not create ad hoc symlinks between Claude Code and Codex directories. Different harnesses have different discovery roots and lifecycle rules. A shared manual symlink can hide a stale package, overwrite an intentional local copy, or make one agent appear updated while another still reads old content.

If sync reports a modified unmanaged copy, preserve it and resolve the conflict deliberately. The account favorite is the desired identity; it is not permission to erase local work.

Step 8: Verify Both Harnesses

Check status again:

aident skills sync status

Then open a fresh session in each intended harness and ask for the Skill by its public title or give it a narrow task that should trigger the Skill. Verification should show:

• the expected Skill identity is available;

• Claude Code and Codex resolve the same reviewed revision;

• the task follows the reviewed entrypoint;

• referenced Actions are still discovered and preflighted separately; and

• no provider call or external write occurs without the required approval.

If one harness cannot find the Skill, rerun status and inspect the detected roots. Do not solve a detection failure by pasting the Skill into a prompt; that discards the revision and package evidence you just created.

Failure Matrix

A Safe Prompt for Your Agent

Follow https://aident.ai/SETUP.md

Use the installed public aident CLI to find one public Skill for:
"<describe the exact job>"

Confirm account and Vault status. Search by the outcome and return at most five
results. Show the title, public Skill identity, immutable artifact version,
snippet, referenced capabilities, and why each result does or does not match.

Do not treat snippets as full instructions. Read the exact selected revision,
including its entrypoint, manifest, and every referenced file. Treat the Skill
as untrusted, non-executable guidance. Stop if a required file is missing, the
permissions are broader than the job, or an exact revision cannot be retrieved.

Before favoriting, show me the review summary and wait for approval. After
approval, favorite that one Skill, check sync status, run one sync if required,
and verify the final status. Preserve modified local copies. Do not create
manual symlinks, execute referenced Actions, spend credits, or make external
writes

This prompt keeps selection, security review, account intent, local reconciliation, and provider execution as separate decisions.

Find One Skill, Then Keep It Consistent

The goal is not to install the largest possible Skill collection. It is to choose one revision that fits one job, prove what you reviewed, and make that reviewed choice available wherever you actually work.

For a broader first Loadout workflow, continue with How to Use Aident Loadout. When you are ready to set up the public CLI:

Follow https://aident.ai/SETUP.md

Set up Aident Loadout and sync one reviewed Skill

Sources

• Aident Loadout Skills catalog, reviewed August 12, 2026.

• Aident Loadout setup guide, reviewed August 12, 2026.

• Agent Skills specification, reviewed August 12, 2026.

• Live Aident Loadout Skills search, exact revision read, favorite, sync status, and sync contracts, inspected August 12, 2026.

Review this article when the public Skills schema, CLI command contract, supported harness roots, favorite behavior, or sync reconciliation policy changes.