Designing skills and loops

§1Start with the skill contract

A skill is a reusable operating contract for one class of work. It tells an agent what it may receive, what it may do, what it must return, how the result is checked, and when it must stop. The prompt is only one component.

§2Deterministic loops: the answer can be exact

A deterministic outcome has a testable predicate: the command exited cleanly, every internal link resolves, the structured data parses, the file hash matches the deployed asset, or the redirect lands on the canonical URL. The loop should be mechanical and idempotent.

1. 01ObserveRead the current state.
2. 02CompareEvaluate an exact predicate.
3. 03ActMake one bounded change.
4. 04Re-runUse the same check again.
5. 05ExitPass, or stop at the retry budget.

while attempts < 3:
    state = inspect()
    if predicate(state) is true:
        return PASS(state.evidence)
    apply_bounded_fix(state)
return BLOCKED(last_evidence)

The important properties are a stable predicate, a fixed retry budget, observable evidence, and safe repetition. “Keep trying until it looks right” is not deterministic; “run the link checker until it reports zero broken internal links, with at most three repair passes” is.

§3Non-deterministic loops: quality needs judgment

Research synthesis, naming, visual critique, prioritisation, and explanatory writing do not have one provably correct answer. Re-running the same model can produce a different useful result. The loop therefore manages variance instead of pretending variance does not exist.

1. 01FrameState audience and rubric.
2. 02SampleProduce one or more candidates.
3. 03CritiqueScore claims and trade-offs.
4. 04SynthesiseKeep the strongest parts.
5. 05GateAccept, retry, or ask a human.

Control comes from a named rubric, independent review, claim-level sourcing, a confidence threshold, and a hard attempt or time budget. More generations are not automatically more truth. A second reviewer is useful only when it has a distinct job: falsify claims, test the audience fit, or expose a missed trade-off.

§4The hybrid pattern: uncertain core, exact enclosure

Most valuable agent work is hybrid. Let the model interpret, propose, rank, or write where judgment helps. Then surround that uncertain core with checks that do not depend on taste.

§5One operating loop for both kinds of work

1. 01SenseRead reality before relying on memory.
2. 02FrameDefine outcome, risk, and evidence.
3. 03RouteSelect the skill and its authority.
4. 04ActPerform the smallest coherent step.
5. 05VerifyUse exact checks and bounded review.
6. 06RecordSave state, evidence, and uncertainty.
7. 07LearnChange the skill when outcomes disagree.

The loop is complete only when the observed outcome can improve the contract. A recurring failure should become a new precondition, test, or stop condition—not another paragraph telling the agent to “be careful.”

§6Prevent infinite refinement

• Attempt budget: a fixed number of repair or generation passes.
• Evidence delta: continue only if the last pass added new evidence or removed a known defect.
• Confidence floor: below it, label uncertainty or escalate instead of smoothing the language.
• Risk gate: irreversible, public, financial, permission, or message-sending actions pause for a human.
• Definition of done: tie completion to the user outcome, not to the agent having produced an artifact.

For a working example, see how Hermes routes skills through Telegram. To apply the pattern, start with the verification and routing templates.