Why loops
Why skilldrop sequences skills in a separate primitive instead of letting skills call each other, and why there are five lifecycle loops rather than three.
The problem a catalogue has
Past fifty skills, a newcomer's first question stops being "is there a skill for this" and becomes "what order do I do these in, and what stops me shipping something bad".
skilldrop always had answers, but only as prose: a README paragraph about "two value streams", two hand-maintained Mermaid diagrams, a loop schema written inside agent-loop-design, and a hand-off graph spread across every SKILL.md. None of it was machine-readable, installable, or checked. Loops make it all three.
Why sequencing moved up instead of sideways
The obvious way to get ordering is to let a skill call the next skill. skilldrop refuses to, and the reason is narrow and load-bearing.
Every skill installs as one folder, copied verbatim. That is the whole advantage over heavier agent platforms: what is reviewed here is byte-identical to what runs in your IDE, and it works the same in Claude Code, Cursor, Kiro, Continue, Cline, and Aider. The moment a skill invokes a sibling, it stops working when installed alone — and "installed alone" is the normal case, because people install what they need.
So the sequencing lives one level up. A loop owns the order; a skill never calls a skill. A skill may declare a directed handoff, but the loop — or a human — is what acts on it. Chaining behaviour, no loss of portability.
The same reasoning shapes the install: LOOP.md's frontmatter is deliberately identical to SKILL.md's, so a loop installs into the skills directory and becomes invokable without any target needing a loop concept of its own.
Why the hand-off fallback is mandatory
Skills install à la carte, so a loop will routinely name a skill the environment lacks. The repo's rule was already "a dangling hand-off is never an error" — but that was prose, and an agent under pressure ignores prose.
Making fallback a required field turns it into a contract: every hand-off states what to do when its target is absent. A hand-off now degrades in a stated, reviewable way instead of dead-ending or being silently skipped.
Why five loops, not three
The obvious model is discovery → build → release. skilldrop uses five plus a wrapper, on the criterion that actually distinguishes them: reversibility, which is also what decides who is allowed to sign off.
| Loop | Mistake costs | So the gate is |
|---|---|---|
discover | a re-brief | a human — no check can tell you that you solved the wrong problem well |
design | months, unwound in code | a review panel — independent positions, recorded dissent |
build | a revert | mechanical — a script's exit code, which cannot be argued with |
release | users and data, often irreversible | a review on evidence (can it ship?), then a human (should it ship now?) |
operate | live users; irreversible | a human closing the incident |
Three loops do not fit this catalogue. The design skills (requirements for quality, diagrams, design docs, contracts, threat models, ADRs) do not fold into discovery without losing the panel gate that makes them worth having. The SRE skills have no home at all in a three-loop model, and operate is the only loop whose failures cannot be undone — the one place a gate matters most.
ship-a-draft is deliberately not a lifecycle stage. It is a wrapper whose middle is any generator: collect the input properly before, check the argument and the surface after. It applies to an ADR, a deck, or a postmortem alike, which is exactly why it cannot be a phase. apply sits outside the lifecycle for a different reason: it serves one person applying for a job, not a team shipping a change.
The gate codes follow the same order. The whole number is the loop's place in the lifecycle, from G0 at discover to G4 at operate, and a decimal is a second gate inside the same loop, so G3.1 is the human go/no-go that follows the G3 readiness review. The two loops outside the lifecycle take the next numbers: G5 for ship-a-draft, G6 and G6.1 for apply. The operating model draws every stage and gate on one page.
What this buys
A stranger can install one command and get a way of working, rather than a list of skills to choose between. And because the loops are data rather than prose, the diagrams cannot drift from them, a renamed skill breaks a check instead of a silently-wrong document, and a gate cannot quietly become decorative.