Why does the trace show nine stages when my template lists fewer skills?

Skills and stages are different lists. A stage can load several skills, and two stages, Brand Context and Quality gates, load none because they make no model call. The fixture the pipeline line on the home page replays shows the same nine.

Can I add my own stage?

Not today. Admins can change the skills a stage loads for a workspace, and custom templates can carry their own chain, but the stage list does not change, so every trace reads the same way.

What happens to a run when I close the tab?

Nothing. Runs execute on a worker, not in the browser. The run page picks up where the run is when you come back.

Guide

The pipeline and the trace

Updated Sep 25, 2026 · 8 min read

A run is one piece of content moving through a fixed chain of stages. Each stage loads named skill versions, calls a model or a check, and leaves its output in the trace. This guide explains the stages, the gates between them, the approval checkpoint, the statuses a run can have, and what each line of a trace means.

The nine stages

A blog post walks all nine stages below, in this order. The first and the last make no model call: Brand Context loads the document, and Quality gates run the checks. Every other stage assembles a prompt from WriterDean's own instructions for that stage, the brand document, the skills the router assigned to it, and the outputs of the stages before it.

One run, nine stages Stage 1 of 9.

69,889 tokens and $0.012 for the captured run

1
Brand Context
product-marketing
The twelve sections of the brand document every later stage reads. No model call.
2
Research
content-strategy
A research brief: intent, questions, entities, gaps, and query variants, with every fact sourced or marked.6,734 tokens
3
Strategy
content-strategy
Buyer stage, a score, a place in the hub and spoke map, and ten title options.6,930 tokens
4
Outline
content-strategy, copywriting
Headings in sentence case, answer block positions, FAQ questions and a target length, plus the internal links to place.9,011 tokens
5
Draft
content-strategy, copywriting
The full piece, with a direct answer up front and a source on every number and claim.9,701 tokens
6
SEO layer
ai-seo, seo-audit, schema
The revised article plus its title, meta description, slug and share text, with the FAQ and the JSON-LD.18,841 tokens
7
Copy edit
copy-editing
Seven sweeps for clarity, voice, proof, and specificity, with no facts added.7,402 tokens
8
Humanize
humanizer
The final text, rewritten for every tell and scored again by the lint.11,270 tokens
9
Quality gates
no model call
Gate results: humanizer lint, SEO preservation, readability, facts, links, alt text, safety.

Which stages a piece runs

The router maps every content type to a chain of skills and a list of stages. Long form runs the full chain. Shorter formats skip what they do not need, and every writing chain ends with copy editing and the humanizer.

Brand Context runs first in every chain. Admins can override a chain per workspace, but no override removes the humanizer.
Family	Stages	Skills in the chain
Blog and long form	research, strategy, outline, draft, SEO layer, copy edit, humanize, gates	content-strategy, copywriting, ai-seo, seo-audit, schema, copy-editing, humanizer
Comparison and alternatives	the same nine	competitors, competitor-profiling, ai-seo, schema, copy-editing, humanizer
Website copy	research, outline, draft, copy edit, humanize, gates	copywriting, cro, marketing-psychology, copy-editing, humanizer
Programmatic SEO	research, outline, draft, SEO layer, humanize, gates	programmatic-seo, site-architecture, copywriting, ai-seo, schema, humanizer
Social, email, ads	draft, copy edit, humanize, gates	the family's skill, copy-editing, humanizer
Strategy documents	research, draft, humanize, gates	the family's skill, humanizer

The gates

Quality gates run after the humanizer on every writing chain. Three are required, and a required gate that fails sends the run to needs review instead of done. The rest are informational: they show in the trace and the editor so a person can decide.

Gate	Required	What it checks
humanizer_lint	yes	Zero hard fails and a weighted score at or under the threshold. See the humanizer.
seo_preservation	yes	The humanized text still carries the primary keyword, the headings and the internal links, as the SEO layer placed them.
safety	yes	No blocked pattern in the output.
brand_compliance	no	Words to avoid from the brand document do not appear.
readability	no	Flesch Kincaid grade against the brand's reading level.
fact_check	no	Every number and named claim carries a source or a needs source marker.
link_validation	no	No invalid or placeholder links.
alt_text	no	Every image has alt text.
plagiarism and ai_detection	no	Informational scores that need a configured provider.

The approval checkpoint

Templates for long pieces pause after the outline. The run stops at awaiting approval, the run page shows the headings, and a person edits them in place if needed and chooses Approve and continue. The draft stage then reads the approved outline. Through the API, approvalAfterOutline sets the same pause, and an agent sees the run stop at awaiting_approval until someone approves it in the app.

The run page mid run. Stages land one after another, and each one lists what it loaded and what it cost.

1A finished stage with its cost
2The paused stage and its skills

Run statuses

Status	Meaning
queued	Accepted. A worker picks it up within seconds.
running	Stages are executing. The run page and `get_run` show which one.
awaiting_approval	Paused at the outline checkpoint for a person.
done	Every required gate passed. The piece has a new current version.
needs_review	Finished, but a required gate did not pass within its retries. The editor shows why and a person decides.
failed	A stage failed after its retries. The error is on the run page.
cancelled	Stopped from the run page.

The humanize stage gets two attempts by default. If the second attempt still fails the lint or the SEO preservation gate, the run finishes as needs review with the best attempt as the current version, so nothing is lost and the findings are there to work from.

Reading a trace

The Trace panel lists every stage of the latest run. Each line carries the skills it loaded with their version and content hash, the model, the input and output tokens, the cost in dollars, the latency, the results of each gate, and the stage's output. Prompts stay with platform admins; the trace shows everything else.

Skill versions and hashes on a trace line. A skill that changes upstream changes its hash, so two runs can be compared skill for skill.

1Skill name, version and hash

Costs in the trace are the model costs of the run. Credits, the unit your plan meters, count finished words instead: one credit per 100 words of the final version. The two numbers answer different questions, so both are shown. See billing and credits.

Running again

Regenerate a single section from the editor with a note about what should change, or run the whole pipeline again with a changed brief. Both create a new version and a new run; the earlier versions stay under Versions, where any two can be compared. Through the API, POST /v1/pipelines with a contentId does the same.

Something missing or wrong on this page? Write to hello@writerdean.com and say which section.