A skill that designs loss functions for long-running autonomous agent loops.

This article explains how it works and why it works.

Spec-driven development says: build this, make the tests pass. Loss-function development (LFD) says: build this, make the tests pass, then iterate against a thousand eval cases you can't see. A test suite is finite — done the moment it's green. A loss function is a target you descend toward. The spec becomes the starting point, not the finish line.

The catch: the agent in the loop is an optimizer, and every cheap path you don't fence off is a direction it will sprint down. It will memorize your eval, mine your miss-lists into lookup tables, and game your judge — not because it's broken, but because you told it where to go and left the shortcuts open.

/lfd-design is the meta-meta-prompt, packaged as a skill. It doesn't solve your task — it observes what harness and tooling you already have, ingests or reverse-engineers the spec, builds the eval (often by mining public artifacts), designs the target, generates and verifies the harness, red-teams its own draft for cheats, and emits a goal.md ready to launch. When the loop cheats anyway mid-run, re-invoke it in patch mode: it reads the iteration log, closes the open path in the loss function (not the agent's code), and resumes from the last honest checkpoint.

goal.md the loss function: spec gate (inner loop), target,
 constraints, cycle protocol, entropy rules, stop conditions
spec.md the inner loop: system design + test cases (ingested,
 or reverse-engineered from the reference artifact)
harness/score.sh task-specific scorer (pixel-diff, recall+precision,
 schema diff — generated for your task, not generic)
harness/lint.sh capacity-cap and eval-overlap checks; violations VOID
 the score and report nothing else
harness/probe.sh perturbed eval variants; the dev-vs-probe gap is the
 memorization gauge
harness/status.sh per-step wall-clock, score history, spend + projected
 burn, the agent's own token consumption
eval/dev/ scored freely during the run (inputs visible,
 answers blinded)
eval/holdout/ scored rarely, aggregate-only; acceptance lives here
LOG.md iteration log: hypothesis / expected failure /
 diagnostic / result

1. Target — large enough that enumeration doesn't pay, blinded during the run, measured by a mechanical instrument at the right resolution.
2. Constraints — wall-clock, money, surface, methodology, and capacity caps on anything that could become a lookup table.
3. Instruments — one command per constraint. A constraint without an instrument is a vibe; the agent will violate it cheerfully because it can't tell it's violating it.
4. Forced entropy — overfit reflection every cycle, a stall rule that bans same-knob-harder, an exploration quota, and a log that survives compaction. Local maxima are the default state of a loop.

Copy the skill into your skills directory:

# personal (all projects)
cp -r skills/lfd-design ~/.claude/skills/
# or per-project
cp -r skills/lfd-design <your-repo>/.claude/skills/

Then in Claude Code:

/lfd-design build a news-retrieval layer that matches <reference product>

Answer the interrogation round, review the generated goal.md, run the pre-flight checklist, and launch your loop against it.

It also works with any agent that reads markdown instructions — point your harness at skills/lfd-design/SKILL.md.

skills/lfd-design/
├── SKILL.md the skill (the meta-meta-prompt)
└── references/
 ├── cheat-museum.md observed cheats and the fences that closed them
 ├── goal-template.md the structure every emitted goal.md must fill
 └── log-template.md the iteration log that survives compaction

There is deliberately no example goal.md in this repo: a worked example would anchor every generated loss function to one task's shape — the skill would overfit to its own example. The templates define structure only; the content is designed per task.

The highest-value contribution is a new exhibit in the cheat museum: a way an agent satisfied your target without solving your task, and the fence that closed it. Every entry makes the skill's red-team phase smarter.

LFD is often used to distill publicly shipped artifacts — outputs a company publishes in the open. That has always been fair to learn from. ToS-gated, login-walled, or paid outputs are not. Respect robots.txt, rate limits, and spend caps; the skill bakes these into every goal it emits.

MIT