add-blog-post
Skill11 starsChanged 5 days agoMITView on GitHub
Research, write, illustrate and publish a post on the Duct blog that ranks in search and gets cited by answer engines (ChatGPT, Perplexity, Google AI Overviews, Claude). Runs a research arm first (demand, the pages answer engines cite today, primary sources, Duct's first-hand evidence), then writes for extraction, draws the hero, renders, lists and verifies the post, and keeps the publishing cadence. Use for "write a blog post", "next post", "blog about <topic or URL>", or a blog content calendar.
In this repository22 other files it gives its agents.
- AGENTS.mdAGENTS.md
- app/AGENTS.mdAGENTS.md
- backend/AGENTS.mdAGENTS.md
- cli/AGENTS.mdAGENTS.md
- desktop/AGENTS.mdAGENTS.md
- site/AGENTS.mdAGENTS.md
- .github/copilot-instructions.mdCopilot instructions
- .cursor/rules/landing-pages.mdcCursor rule
- site/llms.txtllms.txt
- add-changelog-entrySkill
- add-connectorSkill
- component-previewSkill
- featureSkill
- launch-readiness-auditSkill
- mosaic-panelSkill
- new-pageSkill
- prioritizeSkill
- security-auditSkill
- session-auditSkill
- working-backwardsSkill
- duct-sprint-refreshSkill
- launch-checkSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 7 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 67
Every post has one category, from a fixed set the generator enforces (CATEGORIES in scripts/build_blog.py), and 1–6 tags.
- Distinctive: line 117
Read once, dated 2026-10-27.
- Distinctive: line 140
A person reads every post before it merges: Every number, every link, every quote, and the metadata: title, description, alt text, FAQ.
- Distinctive: line 156
No bulk publishing: Never two posts on one day, and at least two days between any two: a burst of same-author, same-template pages is the …
- Distinctive: line 183
One question per post: If the brief finds two, it is two posts.
- Distinctive: line 187
Do all of this before writing a word.
- Distinctive: line 212
Ask the primary question and two secondary ones the way a reader would: in web search, and in whichever answer engine the session can reach …
- Very specific: line 304
Every outside link lands in the BlogPosting's citation list.
- Very specific: line 331
Follow the mosaic-panel skill's six stones, ground and border, but 16:9, no inscription, one subject that is the post's metaphor.
- Distinctive: line 345
Never an image model for anything with words or numbers in it: generated text is wrong often enough to cost more than it saves.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
A post exists to be found, quoted and trusted: found by search, quoted by
answer engines, trusted by the reader who clicks through. Answer engines matter
more than blue links now. A reader who asks ChatGPT, Perplexity or an AI
Overview gets an answer assembled from a handful of pages, with those pages
cited. Being one of the handful is the goal, and the evidence on what earns it
is in references/aio.md. Read it once.
The short version: engines lift self-contained, specific, sourced passages from crawlable pages that other places already mention. Every phase below serves one of those three.
Sharing a post published somewhere else (Shirish's own blog, a guest
post) is not this recipe: it is a link post, a card on the blog index that
goes to the original, with no page and no copy here. Write the front matter
only, with link: set to the original's address, then draw the card and run
the generator (scripts/build_blog.py docstring, "Link posts"). Nothing in
the phases below applies to it.
The value bar: no post without it
A post earns its place only if a reader leaves able to do something they could not before (a checklist they will use, a decision they can now make, a mistake they will now avoid) or knowing something they could not find elsewhere (a measurement, a comparison nobody made, a correction, what we learned shipping it). The brief names both in one sentence each. If it cannot, the post is not written: a rewrite of what already ranks adds a page to maintain and gives no one, reader or engine, a reason to pick it.
Genuine value also means true. Every number is sourced and was seen on its page; every claim about Duct is checked against the code, not the pitch (the harness post first said a person approves "anything that spends money", and the execution policy said otherwise; the post was corrected before it shipped). Say what does not work yet: an honest gap is more quotable than a polished claim.
Who the blog is for
Two audiences, two clusters. The closer the generator ends a post with follows
its category (below); audience in the front matter overrides it.
| Cluster | Reader | What they came for | The post ends on |
growth | the founder, marketer or PM running growth without a data team | a manual job they want to stop doing (read three tools, find the gap, explain the drop) | the product: Download |
builders | engineers building agents | how a shipping agent actually solves a problem, with the code | the repo: Star on GitHub |
A post that starts from a shipped feature comes through the
working-backwards skill: its launch triage
decides whether the feature earns a post at all (most earn a changelog line),
and its PR/FAQ is the brief's first draft. Research still starts at 1b.
The builders cluster is Duct building in public. docs/engineering/ holds
dated research records written for our own decisions (evals, memory freshness,
approvals, code execution); each one is a post waiting to be written for
outsiders, with first-hand evidence no one else has. That folder is the
cadence's supply line.
Category and tags
Every post has one category, from a fixed set the generator enforces
(CATEGORIES in scripts/build_blog.py), and 1–6 tags. The category is
the shelf the post sits on; tags are what it is about.
| Category | For | Default closer |
Engineering | how Duct's agent and product are built, with the code | builders |
Design | interface and product design decisions, and why | builders |
Growth | practical guides for running growth without a data team | growth |
Guide | evergreen explainers: one agent concept, in a marketer's terms (below) | growth |
Announcement | launches and releases worth more than a changelog line | growth |
Story | why Duct exists, what building it in public taught us | growth |
Tags are lowercase topics as a reader would search them (harness engineering,
agent memory, search console), reused across posts rather than invented per
post. They show under the post, and become the BlogPosting's keywords and
article:tag meta. There are no tag pages: with a handful of posts each would
be a thin page competing with the posts themselves. Add them when a tag has
five posts. A new category is a deliberate change to CATEGORIES, not a
front-matter typo the generator accepts.
Guides: the hub inside the blog
A Guide explains one agent concept to a marketer, founder or PM who runs
growth without a data team, in the terms of their week (the Monday read of
three tools, the budget change they would not let a script make), never in
the terms of a framework. It is the evergreen layer of the blog, and it is
how getduct.ai earns the territory AI agents for marketing and growth
without a separate learning hub. The reasoning, with the numbers, is on the
cluster's issue; the rules are:
- One concept, one primary question, the question's words in the title. A guide that needs a second concept to make sense links to that concept's guide, which is a second post. Difficulty under 30 and the reader's job in the phrasing beat a bigger head term the domain cannot rank for.
- The product is the evidence, not the subject. A guide names the file, test or shot that proves the concept is real in a shipping agent, once, and never pitches. The generator's growth closer does the rest.
- Spokes before the hub. The pillar guide for a territory is written
after its spokes exist and links every one of them; each spoke links the
pillar and its nearest sibling. Until the pillar exists, spokes link each
other and the relevant
for-*page. - A
/learn/path, a nav entry or a tag page is earned, not planned. Promote a territory out of the blog only when it holds five or more published guides and Search Console shows non-branded impressions on them. Below that, a hub is a landing page with nothing to land on, and undersite/it would also be five pages per guide to keep translated.
Google's rules for generated content, and what they change here
Read once, dated 2026-10-27. Google's guidance on generative AI content (revised 2026-10-01) says it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing", that the review "also applies to metadata like title elements, meta description elements, structured data, and alternate texts for images", and suggests "providing more background information on how automation was used". The spam policy on scaled content abuse applies "no matter how it's created", and its first example is "using generative AI tools or other similar tools to generate many pages without adding value for users". The September 2026 spam update (24 Sep to 8 Oct) was the fourth of the year, and at Search Central Live Barcelona on 2 Oct Google's Gary Illyes reportedly named scaled, low-effort content as a bigger problem now than link spam (John Campbell's recap for ROAST; not Google's published wording).
Google's own people-first checklist lists the warning signs, and an agent-written blog trips several of them by default: "extensive automation to produce content on many topics", "mainly summarizing what others have to say", writing to a word count, entering a niche for traffic rather than expertise, and dates changed to look fresh. So, as rules:
- A person reads every post before it merges. Every number, every link, every quote, and the metadata: title, description, alt text, FAQ. The PR says who checked it. A post merged unread is a policy risk, not a shortcut.
- Every post says how it was made, in one specific sentence in the
madeWithfront matter key, shown under the byline: what the agent did (the research, the draft, the figures) and what the author checked. Not a boilerplate badge; a sentence that is true of this post. - First-hand evidence is the entry ticket, not a bonus. The brief's "Duct's first-hand evidence" line names a measurement, a file, a test or a shot that no other page has. A post that is only a synthesis of the pages already ranking is the "summarizing what others say" pattern and is not written.
- Alt text describes the image as generated, written after looking at it, never pasted from the prompt: a generated hero rarely matches its prompt exactly, and alt text is metadata Google says to check.
- No bulk publishing. Never two posts on one day, and at least two days between any two: a burst of same-author, same-template pages is the shape of scaled content abuse whatever is in them. Write ahead if you like; merge on the calendar, and date each post the day it actually goes up.
- The generator fails on generated-sounding prose: an em dash in prose
or any word in
GENERATED_TELLS(scripts/build_blog.py). The list is the voice guide's cut-on-sight vocabulary. Quoted sources are exempt.
Cadence
- One post every two weeks, alternating clusters. Research-heavy posts at a rate a solo maintainer plus agents sustain. For answer engines one cited page beats five thin ones: citations concentrate on the best page for a question.
- A cluster written ahead ships every other day. Its posts link to each other, so a weekly drip leaves the set half-built for two months; one post every 48 hours puts it all live within two weeks and still never two on a day. The seven agents-for-marketing guides went out this way, 2 to 14 Nov 2026.
- The calendar is the GitHub Project, not a file. Each planned post is an
issue titled
post: <working title>, labelledsite, with the research brief (Phase 1) as its body and the milestone that holds its publish date. File them with theprioritizeskill'sgh issue createform; the Project'sPriorityorders the queue. - Refresh at 90 days. Re-run Phase 1's answer-engine check, fix anything
that moved (numbers, links, a newer primary source), and set
updated:in the front matter. Freshness is shown to the reader and indateModified; bump it only for a substantive change, never to fake recency. - One question per post. If the brief finds two, it is two posts.
Phase 1: research arm (output: the brief)
Do all of this before writing a word. The brief is the deliverable of this phase; the post is written from it.
1a. Demand: how people phrase it
python3 .agents/skills/add-blog-post/research.py "<topic>" --also "<other phrasing>"Google autocomplete expanded into questions, comparisons and phrasings, Hacker
News stories from the last year, and the GitHub repositories the topic orbits.
Needs network: in the agent sandbox pass suggestqueries.google.com,
hn.algolia.com and api.github.com as allowed domains. Free sources only:
the Ahrefs connector answers "Insufficient plan" on every endpoint, so do not
plan around volumes. Then web-search the topic, and site:reddit.com <topic>
(Reddit blocks the script), for the pages ranking today and the questions
practitioners ask in their own words.
From this, pick the primary question (the one the post answers first, usually "what is X" or "how do I X"), 4–8 secondary questions (they become H2s and the FAQ), and the comparisons people search ("X vs Y"; each gets a table row or an H2).
1b. The citation gap: who answer engines cite now
Ask the primary question and two secondary ones the way a reader would: in web search, and in whichever answer engine the session can reach (Perplexity, ChatGPT search, Gemini, Claude with search). Record, per question, the pages cited. Then answer: what do all of them say, and what does none of them say? The post must contain the second thing, or it has no reason to be cited over pages that already are.
1c. Primary sources
Open every source you will cite. Record the exact title, author or organisation, date, and one verbatim quote under 25 words that you saw on the page. A number goes in the post only with its source; an unverified claim is cut, not softened. Prefer the original (the lab's post, the paper, the spec) over a summary of it, and cite the summary only for what it added.
1d. First-hand evidence: the part only Duct has
This is what makes a post worth citing. Look for it in this order:
docs/engineering/records anddocs/engineering/agent-engineering.md: decisions, measurements, rejected options- the code and its tests: the file that enforces a rule is the proof it exists
- eval and replay results (
make agent-eval, session audits): real numbers from real runs - Duct's own analytics, read through Duct: the getduct.ai Search Console and GA4 (Duct runs Duct)
- the product itself: a shot from
site/assets/media/showing the thing working
Link code on GitHub (https://github.com/5hirish/duct/blob/main/<path>), not
a local path: it is the reader's proof and a link to the repo.
1e. The brief
## Brief: <working title>
- Category: <one of the six>. Tags: <1–6>. Reader: <one sentence, a real person>
- The value: after reading they can <…>; and they learn <…> nowhere else says
- Primary question: <as people type it>. Answer in one sentence: <…>
- Secondary questions (H2s / FAQ): <…>
- Comparisons: <X vs Y, …>
- Cited today (1b): <page — what it says>; the gap none of them fill: <…>
- Duct's first-hand evidence: <file, test, number, shot>
- Sources (verified): <title, author, date, url, quote>
- Where it will be mentioned after publishing: <thread, list, repo, community>
- Slug, title (Title Case), seoTitle (≤60), description (140–160)Phase 2: structure for extraction
An answer engine retrieves passages, not pages. Structure so that any section lifted alone still answers something.
- Answer first. Under the title, before any story: a 40–60 word direct answer to the primary question. Definitions are stated plainly ("Harness engineering is…"), not built up to.
- H2s are the questions, phrased as people ask them (from 1a). Front-load the words that carry meaning.
- Each section opens with its answer in the first two sentences, then earns it. A section that needs the previous one to make sense is two sections merged or one section split wrong.
- Specifics beat adjectives. A number with its source, a named tool, a date, a file. "Up to 40% more visibility" with the paper linked is quotable; "significantly better" is not.
- Tables for anything compared or mapped. They are the most extractable structure there is. A comparison post without one is not done.
- Quote primary sources with attribution, sparingly: a quote is evidence, not decoration.
- A FAQ section (
## FAQ, each question an###) with 3–5 of the secondary questions, each answered in 40–60 words that stand alone. The generator turns it intoFAQPagedata from the same text. - Length follows the question, not a target: a definition-plus-practice post runs 1,500–2,500 words; a narrow how-to 800–1,200. Past 8 minutes the generator adds a contents list.
Phase 3: write
The house voice is in site/DESIGN.md (Voice,
Blog). For posts specifically:
- Open in the reader's situation or with the answer, never with a definition of the topic's category or "In today's…".
- Short paragraphs (1–3 sentences), one bolded phrase at most per paragraph.
- An informed opinion, stated. Posts take positions ("do this, not that, because…") and say where Duct's own choice differs from the source.
- No em dashes in post prose: readers take them as the mark of generated text. Commas, colons, parentheses, full stops.
- American spelling, sentence case in headings, Title Case in the title.
- No pitch. The generator closes every post with the bridge for its audience. Duct appears in the body only as evidence (the file, the test, the shot), never as a sales line.
Links
- Cite inline, at the claim, with the source's own name as the anchor
("Anthropic's Building effective agents"), never "here" or "this post".
Every outside link lands in the
BlogPosting'scitationlist. - 2–3 internal links: another post, the relevant landing page
(
/for-organic-growth,/open-source,/doctrine), the repo. - Link out generously to primary sources. A page that cites well is a page worth citing; outbound links cost nothing.
Phase 4: illustrate
Illustration is how a post shows the idea instead of only saying it. Every post gets a hero; beyond that, pick the form by what the idea is, not by what is quick to make. One strong visual per major section is plenty; a visual that repeats the paragraph beside it is decoration and goes.
| The idea is… | Form | Example |
| how a system is built: parts and what flows between them | system / architecture diagram | the harness around the model: context, model, tools, authority, with evals and traces around them |
| a sequence over time: who does what, in what order | sequence or flow diagram | propose → policy → approve → apply → roll back, with the branch where a person decides |
| a choice between options | comparison: a table first (it is what engines extract), plus a side-by-side visual when the difference is spatial | two memory designs, or two approval flows, next to each other |
| a quantity or a gap | chart, the one form that fits the question (a dumbbell for before/after, a bar for ranking, a line for a trend) | same model, different harness: the gap between the dots is the harness |
| an abstract principle | representative or analogical illustration: the idea as a physical thing | the mosaic hero: a water mill, where the water does the work and the channel, gate and wheel decide where it goes |
| a state or a change over time | timeline or before/after | a remembered fact from "enabled" to "superseded", with the dates it held |
| the product doing it | product shot, annotated when one detail matters | the approval card, the memory timeline |
| a structure the reader will copy | code block or annotated file tree | an AGENTS.md skeleton, the harness in five files |
How each is made:
- Hero, and any analogical illustration: a mosaic scene in the brand style.
Follow the
mosaic-panelskill's six stones, ground and border, but 16:9, no inscription, one subject that is the post's metaphor. Pass one ofapp/art-src/mosaic/*.jpgas the style reference tomcp__gemini-image-generation__generate_image(gemini-3-pro-image,aspect_ratio: 16:9,image_size: 2K). A hero stands alone, so a naive oblique view is allowed where it reads better; app panels stay flat. Expect one re-roll; keep the better image, not the more rule-abiding one. - Diagrams and charts: hand-written SVG in the site's colours (navy
#0d0f1a, grey#4b5068, borders#d9dbe3, the orange#ff5c00for the one thing that matters) and fonts (Georgia headings, system sans labels), with<title>and<desc>. Draw on a 1000px-wide canvas with labels at 19px or more: the figure renders at 652px, and anything smaller is unreadable on the page. Render it at page size withsharpand look before committing. For charts, thedatavizskill's method for choosing the form and the colours. Never an image model for anything with words or numbers in it: generated text is wrong often enough to cost more than it saves. - Product shots: an existing shot from
site/assets/media/(its-768and-1536variants join the srcset automatically). If the right one is missing, add a scenario toscripts/shots/shoot.mjs; never a hand screenshot, never a generated picture of a UI.
Alt text says what the image shows, in one sentence a listener could picture;
the caption says what to take from it. Convert a generated image to WebP at
1320px wide (2x the 652px measure) plus a -768 variant, with the sharp in
app/node_modules (scripts/build_og_images.mjs shows how to require it).
Budget: hero under 300 KB, a variant under 120 KB; mosaic tiles are
high-frequency detail and do not compress much below that.
Phase 5: files
Front matter
---
title: "What Is Harness Engineering? A Field Guide From a Shipping Agent"
seoTitle: "Harness Engineering: A Field Guide From a Shipping Agent" # <title>, ≤60 chars
date: Sep 29 2026
updated: Oct 14 2026 # optional; only for a substantive revision
author: Shirish Kadam
category: Engineering # Engineering | Design | Growth | Guide | Announcement | Story
tags: harness engineering, agent harness, evals # 1–6, lowercase
audience: builders # optional; defaults from the category
excerpt: "The card line on the index: a specific claim that makes someone click."
description: "The meta description, 140–160 characters, with the primary question's words in it."
hero: assets/<slug>/hero.webp # relative to site/blog/
heroAlt: "What the hero shows, in one sentence"
madeWith: "Researched and drafted with Claude Code against the Duct repository; the eval numbers are from our own runs and every source was opened and checked by Shirish before publishing."
readTime: 12
---Markdown the generator accepts
The full list is in the docstring of scripts/build_blog.py; anything outside
it raises. Beyond paragraphs, lists, bold, italics, code and links:
## Question-shaped heading (h2; ### for h3; each gets an anchor id)

| Column | Column |
|---|---|
| cell | cell |
> The quoted words, verbatim.
> -- [Author, Title](https://source)
[!github](https://github.com/owner/repo "What the repository is, one line")The [!github] card is drawn at build time: no widget, no request to GitHub
from the reader's browser. [!youtube](https://www.youtube.com/watch?v=ID "Title,
speaker, length") is a click-to-play facade: the page fetches nothing from
YouTube (no thumbnail, no script) until the reader presses play, then loads the
privacy-enhanced player. Embed a talk only when it adds something the post
cannot say; one per post at most. Anything else that would load a third-party
script on page view (a tweet widget, a CodePen) is not supported, because
nothing loads outside the consent gate (site/AGENTS.md): link to it.
Render and list
node scripts/build_og_images.mjs # the card: kicker, title, and the hero on it
python3 scripts/build_blog.py --fetch-thumbnails # only if the post embeds a video (network)
python3 scripts/build_blog.py # the post page, and the blog index's gridbuild_og_images.mjs redraws every card and the non-blog ones come out with
new bytes and no visible change; restore those (git checkout -- site/assets/og/<page>.jpg)
so the diff holds only the cards the post touched.
The index card and the category filter are generated from the front matter,
so there is nothing to add there by hand. Add the post to the three
hand-kept listings; build_blog.py fails until all three have it:
site/sitemap.xml:<loc>https://getduct.ai/blog/<slug></loc>with<lastmod>inYYYY-MM-DD, and bump the/blog/entry'slastmodsite/blog/feed.xml: an<item>with an RFC 822pubDateand the category as<category>; bumplastBuildDatesite/llms.txt: a line under "Blog posts",- [Title](url): one sentence on what it answers
Phase 6: verify
python3 scripts/build_blog.py --check
python3 .github/scripts/check-pages.py
make check-siteThen look at it: the site dev server (python3 dev_server.py --port 8090 in
site/), and Playwright at 1280 and 390 wide. Check the hero, every figure,
the table's scroll on a phone, the quote, the GitHub card, the contents list,
and the closer. Read the JSON-LD blocks in the page source once: headline,
dates, citation, and the FAQ text matching the page.
Attack the post as a reviewer would: every number has a source, every quote
was seen on its page, the answer-first paragraph survives being read alone,
and nothing in it is false about the product (site/AGENTS.md, "What the
desktop app actually is").
Phase 7: distribute and measure
Publishing is half the job; answer engines cite pages that are already mentioned elsewhere. After the merge deploys, the brief's "where it will be mentioned" list is the to-do. Every item is outward-facing, so each one is proposed to the maintainer, drafted in full, and sent only on a yes:
- a pull request to a relevant awesome list or index, in its own format and under its own rules (read its CONTRIBUTING first)
- a Hacker News or community post where the topic is already being discussed
- a reply in an existing thread where the post answers the question asked
- the maintainer's own channels (X, LinkedIn, the YouTube channel)
The venue rules and the wider plan live with the GTM plan, kept outside this repository.
Measure at 30 and 90 days: Search Console queries and position for the post
(through Duct, which reads getduct.ai's own Search Console), GA4 referrals
from chatgpt.com, perplexity.ai, gemini.google.com, claude.ai and
copilot.microsoft.com, and the 1b questions re-asked to see whether the post
is now cited. Record the result on the post's issue; it is what the next
refresh starts from.
Done means
- Brief written; it names the value, the gap the cited pages leave, and Duct's evidence
- Every claim about Duct checked against the code; category from the six; 1–6 tags
- Answer-first paragraph; question-shaped H2s; each section opens with its answer
- Every number and quote sourced and seen; outbound links to primary sources
- At least one table where anything is compared or mapped; FAQ of 3–5
- Hero, plus the visual each major idea calls for (Phase 4's table), legible at 652px
- Front matter complete;
seoTitle≤60;description140–160 madeWithsays, specifically, what the agent did and what the author checked- The maintainer has read every number, link, quote and the metadata before merge, and the PR says so
- Card drawn, page and index rendered, three listings added,
--checkgreen - Looked at in a browser at desktop and phone width
- Distribution list drafted for the maintainer; issue updated with the brief and the publish date
Related
The method here (research before writing, demand and citation checks,
cadence, Google's generated-content rules) is the cross-project version in
the growth-playbook skill (the maintainer's agent-skill hub at
~/.agents/skills/growth-playbook/, not in this repository;
references/content-and-seo.md). This file keeps what is Duct's: paths,
build scripts, categories, the bridge. A lesson any blog could use goes in
that skill's references/lessons.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

