zero-copy-requirements
Skill88 starsChanged 3 days agoMITView on GitHub
Decides where a fact belongs so that no artifact exists whose content is fully reconstructable from another artifact in the system. Use when adding, reviewing, or pruning documentation, when deciding where a decision or follow-up belongs: a document, an issue, a commit, or a requirement, when an architecture note, decision log, status page, or triage write-up is proposed, or when auditing a documentation tree for duplicated authority. Do not use to establish requirement meaning, normalize a requirement graph, or classify evidence state; hand those to requirements-grounding, requirements-topology, and requirements-traceability.
In this repository52 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- alchemy-orchestrateSkill
- alchemySkill
- architecture-as-code-javascriptSkill
- architecture-as-code-pythonSkill
- architecture-as-codeSkill
- architecture-guidelinesSkill
- bring-downSkill
- ci-cd-reliability-architectureSkill
- continuous-improvementSkill
- defect-shift-leftSkill
- dependency-lifecycleSkill
- evolutionary-database-designSkill
- functionality-complexity-tradeoffSkill
- gate-lifecycleSkill
- implementation-readinessSkill
- infrastructure-fitSkill
- morphogenetic-architectureSkill
- observability-designSkill
- push-outSkill
- requirements-groundingSkill
- requirements-topologySkill
- requirements-traceabilitySkill
- structural-simplificationSkill
- system-optimizationSkill
- test-strategySkill
- zero-copy-requirementsSkill
- alchemy-orchestrateSkill
- alchemySkill
- architecture-as-code-javascriptSkill
- architecture-as-code-pythonSkill
- architecture-as-codeSkill
- architecture-guidelinesSkill
- bring-downSkill
- ci-cd-reliability-architectureSkill
- continuous-improvementSkill
- defect-shift-leftSkill
- dependency-lifecycleSkill
- evolutionary-database-designSkill
- functionality-complexity-tradeoffSkill
- gate-lifecycleSkill
- implementation-readinessSkill
- infrastructure-fitSkill
- morphogenetic-architectureSkill
- observability-designSkill
- push-outSkill
- requirements-groundingSkill
- requirements-topologySkill
- requirements-traceabilitySkill
- structural-simplificationSkill
- system-optimizationSkill
- test-strategySkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Keep every fact in exactly one artifact.
- Distinctive: line 127
If a report carries a fact that exists nowhere else, that fact is the defect: move it into the authority that owns the question, then …
- Distinctive: line 146
Run in order for any existing or proposed document or tracker item.
≈ “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
Zero-Copy Requirements
Keep every fact in exactly one artifact. A second copy is not redundancy; it is a second authority that will disagree with the first.
Purpose: No artifact may exist whose content is fully reconstructable from another artifact already in the system.
Core Directives
- One question, one authority. Each fact has exactly one owning artifact. Every other mention is a reference to it, never a restatement.
- Four authorities carry the system. Behavior, history, open work, and specification. An artifact that is none of them is a generated report, a bounded exception, or waste.
- Decisions close where they were raised. A reversible technical decision closes in the commit or change request that makes it; an owner's answer closes on the issue that asked. Never write a separate decision document for it.
- Sign-off is immutable. Compliance-grade approval lands in a merged commit or change request, not in a mutable comment or an editable page.
- Reports are compiled, never authored. Anything assembled from the authorities is regenerated from them and read-only.
- Reference by identifier. A dependent artifact names the ID of what it depends on; it does not copy that artifact's text.
- Frozen records are closed, not maintained. A historical or migration record is dated and never backfilled; the moment it is updated to stay current it has become an unowned second authority.
- The tracker holds live work, not memory. Whatever is not live work or a bounded owner question has another home, or none (§1).
Boundary
zero-copy-requirements is a task-matched Alchemy companion, not a
qualification stage, gate, or acronym letter. It runs in a single pass, at any
point where an artifact's home is in question.
This skill decides where a fact lives. It does not decide what the fact means, whether the work is worth doing, or whether evidence proves it.
requirements-groundingowns requirement meaning, source authority, and evidence quality.requirements-topologyowns identifiers, lineage, and graph semantics.requirements-traceabilityowns anchors and evidence state once a location exists; this skill decides the location it anchors into.push-outowns the general directive that executable sources beat prose duplicates, and the ladder that moves operational work outward; this skill names which artifact owns which question in the first place.
Apply a project profile for concrete paths, tracker names, schemas, identifier formats, and validation commands. Keep none of those here.
Select the Mode
| Mode | Use for | Emits |
| Targeted | One named or proposed artifact | One output contract |
| Sweep | No artifact named: find every redundant document in a tree | One ranked row per artifact opened, plus the unreached remainder |
With no artifact named, start in Sweep and read
references/sweep.md. Both modes decide with the same
litmus test (§4); Sweep adds discovery, an evidence bar, and ranking. Sweep
classifies nothing DUPLICATE without naming the artifact that already carries
each claim. A claim no artifact carries is unowned, not redundant: move it,
never delete it. Establish neither and the row is Defer, naming what is
missing.
1. The Four Authorities
| # | Authority | Sole owner of | Must not absorb |
| 1 | Code and its tests | What the system does now, and the conditions under which it is correct | Why a change was made, what is still undecided, what is merely planned |
| 2 | Version-control history | What changed, when, by whom, and — via the commit message or change request — the immediate why of that specific change, including any reversible technical decision it makes | Standing policy, open questions, the specification of unbuilt work |
| 3 | Issue tracker | Work picked up now; defects; checks that block a named milestone; owner questions and their answers | Current behavior, which the code already states; acceptance conditions, which the register already states; reversible technical decisions, which history owns; findings, follow-ups, ideas, and reminders, which have no home |
| 4 | Requirement / specification register | What can or should be built, and the conditions that make it complete | The decision's rationale, which lives on the issue or change it cites; evidence state, which traceability owns |
Wiring between them:
- A requirement cites the issue or change that decided it, and — when blocked — the issue or milestone check that will unblock it. It restates neither.
- An issue's closing note, or the answer under a standing question, is the decision record. A decision that outgrows a comment becomes a requirement plus the issue that decided it, not a document.
- A commit message carries the immediate why of its own change, names any reversible technical decision it makes, and references the issue or requirement that authorized it, if any. It does not restate the standing rule.
- Code and tests carry requirement identifiers as anchors, not requirement prose.
Issue Scope
An open issue is work picked up now, a defect, or a check that blocks a named milestone. The rest goes here; a project profile names the two standing items.
| Item | Home | Rule |
| Reversible technical decision: changes neither what users see, nor privacy, nor cost | The commit or change request that makes it | Decide it and name the choice there. A technical trade-off is never an owner question |
| Owner question: what remains after reading the code and running the system | One standing questions issue | Plain language, at most three open at a time, each with the default that stands without objection. An irreversible step waits for a yes. An answer is work: build it before asking anything new |
| Milestone check or fixed date | One milestone checklist | A check closes when it ran, or when the owner accepted the risk for that milestone. An issue with an unrun check moves that check to the checklist first |
| Unbuilt obligation nobody has picked up | The requirement register | It becomes an issue when work on it starts; without an obligation it is an idea |
| Finding, follow-up, idea, or reminder | No artifact | Fixed in the change at hand or dropped, never filed |
2. The Generated-Report Layer
Anything compiled from the authorities is a report, not an authority.
- Regenerate it from a command; never hand-author or hand-patch it.
- Mark it read-only, and fail the build when a committed report differs from a fresh regeneration.
- A report may be an external-audience deliverable — a coverage matrix, a status board, a compliance pack assembled from requirements and issues — but its content still originates entirely in the authorities.
- If a report carries a fact that exists nowhere else, that fact is the defect: move it into the authority that owns the question, then regenerate.
3. Bounded Exceptions
Exactly three kinds of artifact may exist outside the authorities and the report layer.
| Exception | Admission test | Guard |
| Onboarding / tooling meta-documentation | Explains how to work on the system — setup, conventions, which command to run, which skill to route through — rather than what the system does or what was decided | Delete it when the tooling carries the instruction itself; see the documentation-pruning pattern in push-out |
| Compliance-mandated external-audience deliverable | An external obligation names the artifact and its audience, and the authorities are not a deliverable form for that audience | Restructure as a generated report as soon as its content is reconstructable; hand-authoring it each cycle is the anti-pattern |
| Frozen historical or migration record | Closed and dated; describes a completed transition, a superseded model, or a one-time import whose provenance is still cited | Never backfilled, never updated to stay current, never the authority for anything live |
An artifact that fits no exception and duplicates an authority is waste. Fold its content into the owning authority and delete it.
4. Litmus Test
Run in order for any existing or proposed document or tracker item. Stop at the first match.
1. Is its content fully reconstructable from artifacts that already exist?
→ Yes: delete it, or replace it with a generated report.
2. Does it answer "what does the system do?"
→ Code and its tests.
3. Does it answer "what changed, when, and why that change?"
→ Version-control history.
4. Does it answer "what is undecided, what was decided, or what work is open?"
→ Issue Scope in section 1 places it. Work picked up now or a defect: an
issue, whose decisions close as its comments or closing note. A decision
that must later be proved unaltered closes in a merged commit or change
request (Directive 4).
5. Does it answer "what should be built, and when is it done?"
→ The requirement register, citing the issue or change that decided it.
6. Is it compiled from 2-5?
→ A generated report, regenerated by command and read-only.
7. Does it pass exactly one exception admission test in section 3?
→ Keep it, and record the guard that ends or freezes it.
8. Otherwise
→ Waste. Fold and delete.A document that needs two answers is two artifacts. Split it before keeping it.
5. Anti-Patterns
| Anti-pattern | Correction |
| A decision document or ADR file recording a decision already argued on an issue or change request | Close the decision where it was argued; cite that ID where the decision is needed |
| A decision log appended to a requirement | The requirement cites the deciding issue or change; it is not the decision record |
| A finding, follow-up, idea, or reminder filed as an issue | Fix it in the change at hand or drop it |
| A reversible technical decision raised as a question or an issue | Decide it and name the choice in the commit or change request |
| An owner question with no default, or a new one while an answer sits unbuilt | Give it a default, or build the answer first |
| Hand-maintained changelog, status page, or progress board | Generate it from history, the tracker, and the register |
| Architecture or design note narrating what the code already does | Point at the module, contract, or enforcement rule |
| Triage, investigation, or incident write-up committed as a repository file | An open defect is an issue and its comments; the rest is fixed or dropped |
| Requirement text restating a dependency's acceptance conditions | Reference the criterion identifier |
| Compliance deliverable hand-authored every cycle | Compile it from the authorities; the deliverable is a report |
| Compliance sign-off recorded in an editable comment or page | Land it in a merged commit or change request |
| A frozen record edited to stay current | Either freeze it or promote its content into an authority; it cannot be both |
| "Documentation debt" resolved by writing more prose | Locate the owning authority first; prose that duplicates it increases the debt |
6. Output Contract
Subject: <document, tree, or proposed artifact>
Classification: AUTHORITY | GENERATED-REPORT | EXCEPTION | DUPLICATE
Owning authority: <code+tests | history | issues | requirements | none>
Reconstructable from: <artifacts that already carry the content, or none>
Decision: Keep | Fold into <authority> | Generate | Freeze | Delete | Defer
Exception: <onboarding/tooling | compliance deliverable | frozen record | none>
Guard: <condition that ends or freezes the exception, or none>
Next action: <one concrete action>In Sweep, emit one ranked row per artifact opened, then the unreached count:
| Artifact | Classification | Owning authority | Carried by | Decision | Guard |
| -------- | -------------- | ---------------- | ---------- | -------- | ----- |Carried by names what already holds the content: file and symbol, commit,
issue ID, or requirement ID. Empty means unowned, not redundant.
7. See Also
push-out- executable sources beat prose duplicates; documentation-pruning pattern.requirements-traceability- anchors and evidence state once the location is decided.requirements-grounding- requirement meaning, source authority, and evidence.requirements-topology- stable identifiers, lineage, and generated views.continuous-improvement- promoting a recurring duplication finding into a check.
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.

