hcm-architecture
Skill0 starsChanged 51 days agoMITView on GitHub
Design recommendation and direct advice to the client for Oracle Fusion Cloud HCM. Use when the user asks what the right way to do something is, whether a set of choices holds together, whether to fix or redesign, which of two approaches to recommend, how roles, security, positions, approvals, page extensions or integrations should be structured, what drives the load or the dependencies of a change, or how to answer a client who asks for an opinion or a recommendation. Produces a recommendation with the decision structure behind it, the load drivers, the dependencies that slip a plan, the trade-offs, and the full text of the advice to send. Never produces day estimates or figures. Not for something broken, which belongs to hcm-debugging, and not for building a change that is already decided, which belongs to hcm-configuration.
In this repository12 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- smoke-checkSkill
- fusion-instanceSkill
- fusion-referenceSkill
- fusion-suiteSkill
- fusion-testSkill
- hcm-configurationSkill
- hcm-debuggingSkill
- jira-ticket-ingestionSkill
- triage-batchSkill
- triageSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 4 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 13
Read CONTEXT.md in the engagement directory first when it exists.
- Distinctive: line 24
commit production, a colleague or Oracle to anything
- Very specific: line 50
Before forming an opinion, settle every product fact the recommendation depends on: web search restricted to docs.oracle.com to find the …
- Distinctive: line 78
When the signals split, say so.
- Very specific: line 85
Read load-and-dependencies.md and produce:
- Distinctive: line 100
Every recommendation in this product trades something.
- Distinctive: line 116
No figure, no date, no day count: When they asked for one, one sentence says what drives the load and that the figure is produced elsewhere.
- Very specific: line 124
Write it to output/advice/{slug}.md in the engagement directory.
≈ “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
Oracle HCM architecture and design advice
The consultant is asked for an opinion. Not to fix, not to build: to say which way is right, and why, so that the client can decide. The authority behind the opinion is product knowledge and the way this product behaves over years of quarterly updates, not seniority. So every recommendation stands on a documented mechanism, names the option it rejects, and says what would make it flip.
Read CONTEXT.md in the engagement directory first when it exists. The product knowledge (where configuration
lives, page families, how security is composed, scheduled processes) is in the fusion-reference
skill; open the file it points to when a section below needs it.
Scope: advise, do not decide
The client decides. The consultant gives a recommendation, the reasons, the trade-off, and the consequences of each branch. He does not:
- decide the business trade-off between two valid designs
- produce a number of days, a date or a cost, for anything, ever (see the rule at the end)
- commit production, a colleague or Oracle to anything
- widen the question into an audit of everything else that could be redesigned
The deliverable is a recommendation brief and the text of the advice, ready to send. The brief is for him, the text is for them.
Step 1: frame the question they are actually asking
Quote what was literally asked. Then say which of the recurring question shapes it is, because
each has its own decision structure in decision-structures.md:
| Shape | The question underneath |
| Fix or redesign | Is the recurring symptom a defect in the object, or a sign that the object was built for a different process? |
| Delivered or extended | Does Oracle ship this, and at what cost of ownership does the extension come? |
| Where does this belong | Which layer (security, setup, page rule, extension, data) should carry this behaviour? |
| Does this hold together | Are these choices consistent with each other and with the product's model? |
| Which of two | Two viable options, what decides between them? |
| What will this drive | What will a change pull in, what has to happen first, what will slip it? |
Then name the decider, and what they will do with the answer. A recommendation to someone who cannot act on it is a memo, and the message has to be addressed to the person who can.
Step 2: settle the product facts from documentation
A recommendation about this product is only as good as its reading of what the product does.
Before forming an opinion, settle every product fact the recommendation depends on: web search
restricted to docs.oracle.com to find the page, fetch_oracle_page of the oracle-fusion-docs
MCP to read it, and the sentence quoted verbatim with its URL. Check What's New for the last two
quarterlies: an opt-in feature that ships next quarter changes a recommendation made today.
A fact that could not be sourced is stated as an assumption, in the brief and in the message. A recommendation that hides an assumption is the one that comes back.
Step 3: lay out the options, honestly
Two to four options, never one. For each, in one line each:
- what it is, in the product's terms (the object, the layer, the mechanism)
- what it costs to own: what has to be re-verified each quarter, what only the builder understands, what breaks when the person who built it leaves
- what it does well and what it does badly for the need in step 1
- what it depends on being true first
The option the requester already prefers is on the list and gets the same treatment as the others. So does "do nothing", when the current state is a real option.
Step 4: apply the decision structure
Take the structure for the shape identified in step 1 from decision-structures.md: the signals
that point each way, the default recommendation, and the conditions under which the default flips.
Walk the signals against the facts of this case and say which way each one points. The
recommendation is what falls out, not what was decided beforehand and dressed up.
When the signals split, say so. A recommendation with a stated confidence and a named condition ("recommended unless the population exceeds one legal employer") is more useful than a confident one that omits the condition.
Step 5: name the load drivers and the dependencies, without numbers
The client wants to know what a change will pull in. The answer is a structure, not a figure.
Read load-and-dependencies.md and produce:
- The load drivers: which dimensions of this case make it heavier or lighter, and in which direction. Number of roles touched, page families in scope, countries or legal employers, retroactivity, testing surface, data to convert, decisions the client has to make.
- The dependencies: what has to exist or be decided before the change can start, in order.
- What slips a plan in this product: the quarterly update window, the environment refresh timing, the client's decision latency, an open Oracle SR, the change board cadence, data cleanup nobody has scheduled.
This is where the value of the advice lives. A client who understands why a change is heavy decides better than one who was told it takes so many days.
Step 6: the trade-offs, stated as such
Every recommendation in this product trades something. Say what: flexibility against maintainability, a quick personalisation now against a Redwood migration later, role granularity against administration burden, automation against control, one standard against local exceptions. Name the side the recommendation takes and what is given up. The client may value the other side, and that is their call.
Step 7: write the advice, in full
The message is the deliverable. Rules:
- As him, first person, to the named decider, in their language.
- The recommendation in the first sentence. Then the reason, then the trade-off, then what they need to decide or confirm. Options and their consequences as a short list if there are more than two.
- The documented mechanism quoted, with its URL, when it carries the argument.
- Assumptions stated, each in one clause.
- No figure, no date, no day count. When they asked for one, one sentence says what drives the load and that the figure is produced elsewhere.
- No offer of extra work in the closing. It reads as a commitment.
- One paragraph per line, no hard line breaks inside a paragraph. The text is pasted into a field that reflows.
Deliverable: the recommendation brief
Write it to output/advice/{slug}.md in the engagement directory. It is not read by npm run actions. Template:
# Recommendation: {slug}
Asked by: {person} | Decider: {person} | Date: {date}
Recommendation: {one sentence} | Confidence: {percent} | Flips if: {condition}
## 1. Message to send (full text, destination and language named)
## 2. Not mine (decisions and actions that belong to someone else, one line each)
## 3. The question, as asked and as understood
## 4. Options (what, cost of ownership, fit, prerequisites)
## 5. Decision structure applied (shape, signals, which way each points)
## 6. Load drivers (dimension, direction, why)
## 7. Dependencies, in order, and what would slip them
## 8. Trade-off taken and what is given up
## 9. Assumptions and unsourced facts
## 10. References (verbatim quotes with URLs, Doc IDs)Hard rules
- No estimate in days, no date, no cost, no figure of any kind. This was decided deliberately: the figure is not the consultant's to produce here, and a number given quickly becomes a commitment. The load drivers and the dependencies are what he delivers instead.
- Never one option. A recommendation without a named rejected alternative is an order, not advice.
- Every product fact carries its URL or is labelled an assumption.
- The client decides. The message recommends, it does not announce a decision.
- Do not widen the question. Adjacent things that could be redesigned are one NOT MINE line.
- Documentation before the instance. Reading the instance is justified only when a fact the recommendation depends on cannot be settled otherwise, and then it is a read, never a change.
- Name the option the requester preferred and treat it fairly. Advice that dismisses it without the same analysis loses the reader.
- No placeholder in a message.
- No em dashes anywhere.
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.

