Skip to main content
Knowledge and Skills carry different kinds of context into future work:
  • Knowledge holds source material, brand facts, reference context, and artifacts.
  • Skills holds reusable working procedures and the conditions under which agents should use them.
Open Knowledge to manage sources and inspect artifacts. Open Skills to review the active procedure library, proposed additions, and imports.

​
Before you begin

  • Confirm the active workspace; these sources and procedures can shape work across that workspace.
  • Identify the source owner, its effective date, and whether it is approved for shared use.
  • Remove secrets, personal data, or material the workspace should not retain.
  • Decide whether the material is brand identity, reference knowledge, or a working procedure before upload.
  • Resolve conflicting authorities. Adding both as undifferentiated sources makes future grounding less reliable.

​
Know what goes where

Put volatile performance numbers in the surface that owns them, not in a Skill. Put a method for evaluating those numbers in a Skill.

​
Build Knowledge from a website

Use the website route when the approved public site should seed brand and reference context.
1

Enter the authoritative site

In Sources, enter the approved domain in the website field. Use the canonical brand or storefront site, not a campaign-specific mirror or an unrelated retailer page.
2

Start the build

Select Build from website. The source moves into Brand research running while Marketer gathers and structures the available material.
3

Follow progress

Select view progress to inspect the active research task. Closing the panel does not make the source complete; return to Knowledge and wait for a terminal source state.
4

Review the result

Inspect the new source and the standing or recent artifacts it produced. Confirm brand facts, product claims, audience conclusions, and dates against the site before relying on them.
5

Test grounding

Ask the Marketing Agent a question the site should answer. Require it to name the source it used, then compare the answer with the current page before treating the build as verified.
Website research reflects what was available when it ran. Rebuild or replace the source after a material site change rather than assuming future tasks read the live site on every request.

​
Upload and route files

Use upload for approved PDFs, Markdown, text, images, or font material.
1

Choose the destination first

Open Route uploads as and select Brand kit material, Reference knowledge, Working procedure, or Let the agent decide. This choice determines how future agents receive the file.
2

Add the files

Drop files into Drop files or click to upload, or use the picker. Confirm the file name and remove duplicates before upload.
3

Watch the source state

The row can move through upload and processing states before Done, or show Failed. Do not test a file while its row is still processing.
4

Retry a failed source

Use Retry on the exact failed row. If it fails again, check file integrity and supported type before creating another copy.
5

Verify the route

Under In your knowledge base, read the destination shown after the arrow, such as Brand kit, Knowledge, or Skill. Open or test the destination rather than relying on the original picker choice.

​
Change a source destination

Use the source row’s destination menu when a file was routed incorrectly.
  • Moving material into Brand kit can start new extraction work and can use credits; review the confirmation before proceeding.
  • Moving a file out of Brand kit changes how future updates are routed. Details already extracted into the brand context can remain, so inspect the Brand Profile instead of assuming the move erased prior output.
  • When moving a procedure into Skills, review the created Skill’s trigger and body before trusting it in recurring work.

​
Inspect standing context and artifacts

Knowledge surfaces three different views of context: An artifact summary is not sufficient evidence for a decision. Open the detail and inspect the sources it is Grounded in. Preserve the distinction between a sourced finding, the agent’s synthesis, and a later recommendation.

​
Validate Knowledge in a real task

After adding or moving an important source:
Active Marketer task showing Saved brand memory in search context and Reading the saved brand memory as a completed step

Task progress can show that Marketer searched and read saved workspace context. Use that as a retrieval signal, then verify the final claim against the named source.

  1. Start a task with a question that has one verifiable answer in that source.
  2. Attach an exact product, brief, campaign, or creative with @ when it is part of the question. For a one-off document, use the plus menu and select Add files or photos.
  3. Ask the agent to use saved workspace Knowledge and cite the fact, source title, and effective date. Require it to identify any conflict.
  4. Compare the answer with the uploaded file or current website.
  5. Correct the source, destination, or authority order if the answer is stale or ambiguous.
  6. Re-run the test in a new task so cached conversation context does not hide the correction.

​
Govern Skills

The Skills page separates stable procedures from suggestions that still need a human decision.

​
Review the active library

The active table can show Skill, Scope, Win rate, Used, Last used, and Pinned. Use these fields to find procedures that overlap, rarely run, or need clearer trigger language. Usage and win rate are signals for review, not automatic reasons to keep or delete a Skill. Brand foundation is always on and supplies the workspace’s foundational brand procedure to agents. Its identity and scope are locked; review the underlying brand context rather than trying to delete it like a normal Skill. Open an editable Skill to review:
  • Name
  • Scope
  • Status
  • Description
  • When to use
  • Body in Markdown
Select Save only after the trigger and procedure agree. Use Delete only for an editable procedure you have positively identified; archive or replace a procedure when preserving its history matters.

​
Review agent suggestions

Ready to promote contains human-gated candidates with supporting evidence.
  1. Open the candidate and read the evidence for the proposed procedure.
  2. Choose Edit to tighten its name, scope, trigger, steps, stop conditions, or verification.
  3. Choose Promote only when it is distinct from the active library and safe to reuse.
  4. Choose Dismiss to hide it from the current review session. A relevant suggestion may appear again later.
Also inspect Proposed, Archived, and What agents learned. A learned pattern is input to governance, not workspace policy until you explicitly promote or create an active Skill.

​
Import a SKILL.md procedure

Select Import SKILL.md when another approved procedure should become available in this workspace.
1

Add source material

Drop the Markdown file into the import dialog or paste its content, then select Add to preview.
2

Review the preview table

Confirm Source, Name, Scope, and Status for every proposed import. Remove an entry that is a duplicate, too broad, or missing a usable trigger.
3

Inspect the procedure

Read the full body for secrets, environment-specific assumptions, unsupported account actions, missing approval gates, and references that do not exist in Marketer.
4

Import deliberately

Select Import for the reviewed count. Imported Skills become active for the agent immediately, so do not use import as a staging area for unreviewed procedures.
5

Test both sides of the trigger

Run one request that should use the Skill and one that should not. Confirm the agent follows the procedure only inside its intended scope.
A Skill can influence future tasks across its scope as soon as it is active. Never import or promote a procedure whose authority, stop conditions, and human-review boundary you have not read.

​
Completion checklist

  • Every important source has an owner, date, authority, and intentional route.
  • Website and upload tasks reached a terminal state; failed rows were resolved or removed.
  • Brand kit, Knowledge, and Skill destinations match the material’s purpose.
  • Recent artifacts identify the sources they are grounded in.
  • A fresh task retrieved a test fact correctly and disclosed its source.
  • Active Skills contain a specific trigger, ordered steps, decision branches, stopping conditions, and a final check.
  • Proposed or imported Skills were reviewed for overlap, secrets, unsupported authority, and stale facts.
  • One positive and one negative trigger test confirmed the new or changed Skill’s scope.

​
Troubleshooting

Open view progress and inspect the task for a source, crawl, or decision failure. Verify the domain is reachable and canonical. Do not start several identical website builds; resolve or stop the first task before retrying.
Use Retry on the failed row after confirming the file opens locally and uses a supported type. If the source is a scan or malformed export, create a readable approved version before uploading again.
Use the row’s destination menu and review the move confirmation. After moving, inspect both the new destination and any previously extracted brand details; routing changes do not guarantee prior extraction is erased.
Confirm the source is Done, routed to the intended destination, and not contradicted by a higher-authority source. Run a fresh task with the source named explicitly and ask the agent to cite its evidence.
Replace or rebuild the outdated source, record the new effective date, and resolve conflicts in the standing context. Re-run decision-grade work in a new task after the source change.
Open it and make When to use specific to the real request language. Confirm its Status and Scope, then test a matching request without relying on an older conversation’s context.
Add explicit exclusions and stop conditions to When to use and the body. Test nearby requests that should not use it before keeping the Skill active.
Keep one authoritative procedure. Merge the useful evidence or steps into it, archive or delete the duplicate as appropriate, and rerun both trigger tests.

​
Next step

Use context in chat

Attach the exact source or record and verify how the agent uses it.

Apply context to creative

Carry approved facts and repeatable review procedures into Briefs and Studio.
Last modified on August 9, 2026

Related topics

Chat and contextCreate a reusable SkillNext stepsDiagnose a performance changeRelease notes