Understand the parts.
Build a method you can reuse.

1. State the job and the audience. Ask for one specific output, such as a 150-word email and a claims table.
2. Attach the relevant brief, approved examples and context. Ask which files the AI can actually access. Another conversation's context isn't automatically available.
3. Set limits: approved claims, tone, length, required evidence and anything that needs a human decision.
4. Read the draft against the originals. Explain what needs changing and why, then ask for a revision.
5. Save the reviewed output and useful method as Markdown files. Reopen the saved files to check them.

Using only the context and evidence I supply, draft one launch email of up to 150 words for [audience]. First list the files you can access and ask about any essential gaps. Follow [tone examples]. Add a claims table with the source filename and section for each factual claim. Separate suggestions from supported facts. Don't send or publish anything. Return a draft for my review.

Build a brief checker in n8n

Start with test records and a draft output. Replace the input and destination with approved connections only after the checks pass.

1. Create a workflow with a Manual Trigger. Add Edit Fields (Set) and enter a fictional brief: briefId = TEST-001, productCode = HALDEN, audience = compliance managers, deadline = 2026-10-01. Add a test owner mapping: HALDEN → Campaign owner A.
2. Add an If node. Require a non-empty brief ID, product code and audience, plus a valid deadline in your agreed format. Define the date check explicitly; a non-empty string alone isn't a valid date.
3. On the false branch, use Edit Fields to return status = needs_information and the missing or invalid fields. End that branch without creating a task.
4. On the true branch, look up productCode in the owner mapping. If the code has no owner, return status = needs_owner. Never guess an owner.
5. For a known owner, return a draft routing record containing briefId, owner, audience and deadline. Test it before adding an approved Monday or other destination node.
6. When adding a live destination later, check briefId against existing records before creating anything. Repeated delivery of the same brief mustn't create duplicate tasks. Store credentials in n8n's credential controls.

Official guidance: n8n If node

AI could assess whether the brief is clear or suggest campaign angles after validation. Keep the fixed field checks and owner lookup as rules.

Help me design an n8n brief-routing workflow using fictional records. Use fixed rules for required fields, date validation and product-code-to-owner lookup. Include branches for missing fields, invalid dates, unknown owners and duplicate brief IDs. Begin with Manual Trigger and a draft output, with no live writes. Show the nodes and exact conditions. Identify any step that truly needs AI judgement; don't add an AI node for simple field movement.

Download the optional brain-building skill to use a ready-made method for interviews and source-backed knowledge. Setup instructions and the blank starter structure. Downloading is not installation; this is follow-on work.

Route 1: complete the task first (preferred)

1. Do the whole task with AI and your real review process. Keep the original inputs and the final version.
2. Correct mistakes and awkward steps. Note what changed, why it changed and which decisions needed a person.
3. Ask AI to extract the working method: inputs, sequence, rules, output, checks and stopping points. Remove one-off campaign facts.
4. Turn that method into a draft skill. Test it on a different input in a fresh task with the relevant context supplied.

Route 2: start from an existing SOP

1. Supply the SOP, a good input/output pair and the reviewer's criteria.
2. Ask AI to identify missing decisions or contradictory steps before drafting the skill.
3. Walk through the draft on one example. Correct it, then test another example. An SOP is a starting point, not evidence that the skill works.

Turn the reviewed process and corrections from this task into an instruction-only skill called campaign-brief-check. If I supplied an SOP instead, first identify gaps and ask the necessary questions. Include when to use it, required evidence, steps, output format, missing-information handling, human review and a second-input test. Separate reusable rules from campaign facts. Create SKILL.md with name and description metadata; list any supporting files. Return a draft for review. Don't install, enable or share it yet.

Keep a reviewed copy of the skill and its examples in approved team storage. A portable instruction file doesn't make tool connections or permissions portable.

Prepare once, before activation

1. Choose one job and a reviewed method. Define accessible sources, reporting period, timezone, output destination and reviewer.
2. Run the job manually with test evidence. Check missing sources, duplicate records and a period with no changes. Keep incomplete data visible.
3. Agree what can happen automatically. Start with a private draft. Name who handles failures and who can pause the task.
4. After setup, inspect the first actual scheduled run: its time, evidence access, output and destination. Don't treat a saved schedule as a successful run.

Help me set up an end-of-day change summary using the reviewed method [skill or file]. First confirm the accessible sources, destination, reviewer and timezone. Proposed cadence: weekdays at 17:00 in Europe/London. Report changes since the last successful run; on the first run, state the baseline period. Link each change to evidence, flag unavailable sources, and say if there are no changes. Save a private draft for [reviewer]; don't send it. Run a manual test first. Show the final configuration for my confirmation before activating it. Explain how I inspect a run and pause it.

If your account lacks scheduling, keep the tested prompt and method for manual use. Activate later in an approved environment. These guides don't create a schedule from this website.

Download loop diagram (PNG)

Choose the loop you mean

1. Improve one output: compare it with explicit criteria, correct failures and stop when it passes or reaches the attempt limit.
2. Repeat a check: inspect an accessible source at a set cadence. Define what counts as a change and when monitoring ends. Use a schedule if it needs to continue between sessions.
3. Retain an improvement: propose a change to the saved skill or context, get human approval, test it on a different input and log the accepted change. Make that version available to future runs.

The model hasn't automatically retrained itself. You've improved the instructions, examples or knowledge it receives. A higher score from the same model isn't enough to prove the change works.

Review this draft against the supplied evidence and these criteria: every factual claim has a source; no unsupported capabilities; the output meets the requested format. List specific failures, then revise. Allow at most 2 correction attempts. Stop and ask me if evidence is missing, criteria conflict or the same failure persists. Show unresolved issues honestly. After I review the result, propose a minimal update to method.md, with a reason and source. Don't apply it without my approval. Test the proposed method on a different input and recheck the original before proposing adoption. Keep the previous version and add a dated change-log entry for accepted updates.

Download harness diagram (PNG)

Further reading: Anthropic on agent harnesses.

Map your own setup

1. Write down the job and output. Choose the application that can work with your approved evidence.
2. List the context and skill it needs. Keep a portable reviewed copy so another application can use the same underlying method.
3. List the tools and allowed actions. A connector provides access; it doesn't define a workflow or guarantee that the data is correct.
4. Name the fixed checks, AI decisions, reviewer and failure route. Add a schedule only if the task should run without a manual start.
5. Test a small example and a missing-input case. Inspect tool results and the saved output. Confirm which files will actually be supplied to the next run.

Map the environment needed for this workflow: [job]. List the goal, supplied context, reusable skill, fixed-rule steps, AI judgement, tools, permitted actions, execution location, output and human checks. Mark each component available, missing or unverified based on what you can actually access. Propose a small manual test and one failure test. Don't connect tools, change permissions or create schedules. Show what could be reused if we moved to a different AI application.