After Wasting 2 Billion Tokens, I Open-Sourced Leader.skill for Defining Agent Goals

Key Highlights

A developer shared candidly in a public account post that, in order to work out a method for making agents reliably do real work, he had “wasted” roughly 2 billion tokens in trial and error. It was precisely this non-trivial cost of experimentation that gave birth to Leader.skill—an open-source tool that translates vague human requirements into goal briefs that an agent can execute independently for hours.

The story is worth telling because it names a quiet truth of the agent era: the hard part is rarely the model itself, but defining what we actually want the model to do. Leader.skill is one practitioner’s answer to that problem, distilled from expensive lessons, and packaged so others do not have to pay the same tuition. In a field where most advice stays at the level of clever prompts, a tool that treats goal definition as an engineering discipline stands out.

What Happened

Leader.skill tackles the most painful link in agent deployment: human requests tend to be vague, such as “do a competitor analysis for me,” yet an agent needs executable, verifiable, and correctable instructions. The Skill converts such ambiguous demands into a “goal brief” that an agent capable of completing tasks can power through for hours, so the model knows from the start what to do and what counts as done.

Rather than leaving the user to hand-craft a system prompt, the tool structures the request into a form the agent can act on without constant human hand-holding. That shift—from a fuzzy ask to a self-contained assignment—is what makes long-horizon autonomy plausible, because the agent no longer has to guess the user’s intent mid-task. It also shortens the loop between a human idea and a running agent, which is exactly where most real projects stall today, and it turns a vague wish into a spec the agent can be held accountable to. The practical payoff is fewer restarts and less wasted compute when a long job finally fails near the finish line, which is the failure pattern that burned through those 2 billion tokens in the first place.

Technical Details

Its methodology is called the “Seven Questions for Goals,” which breaks a goal down clearly across dimensions such as purpose, definition of done, anti-cheating, and boundaries. Purpose answers “why are we doing this”; definition of done specifies “what delivery looks like”; anti-cheating prevents the agent from lazily cutting corners; and boundaries draw the line around “what not to touch.”

The author also suggests a workflow: use Claude Code or Kimi K3 for goal planning first, then hand long-horizon execution to models such as GPT-5.6 Sol or GLM-5.2. The whole project is open-source, so anyone can take it and use it directly, adapting the seven questions to their own domain without vendor lock-in. This deliberate separation of planning from execution lets the strongest reasoning model shape the goal, while a cheaper model carries out the grind, which is a pattern many production teams converge on anyway.

vs Competitors

Compared with approaches that hand you a single prompt template, Leader.skill’s difference is “structure.” It does not expect one magic sentence to summon a perfect agent; instead it nails the goal down with seven questions, reducing the chance the agent drifts off course during a multi-hour run.

Against frameworks that demand lengthy system prompts, this “goal brief” thinking is lighter and closer to real tasks, and easier to migrate across different models. The emphasis is on defining intent well, not on wrapping the model in ever more elaborate scaffolding, which keeps the human in control of outcomes. The result is a specification the agent can be measured against, rather than a vague instruction the agent is free to reinterpret however it likes.

Industry Impact

Put simply, the tuition of 2 billion tokens bought a reusable instruction manual for “assigning tasks to agents.” For teams that want to put agents to work, it provides a bridge from fuzzy requirements to executable goals, especially suited to long-horizon, multi-step, autonomous-decision work where ambiguity is the main failure mode.

As agents move from toys to productivity, tools that engineer the “goal definition” step will only grow in value, because the bottleneck is shifting from model capability to specification discipline. Leader.skill shows that a well-structured goal brief can be the difference between an agent that fills hours with activity and one that actually delivers a result, and that the real leverage in the agent stack may sit less in the model itself and more in how clearly we tell it what we want. A reusable goal brief is, in the end, the cheapest insurance against wasted tokens. For organizations scaling agents across many workflows at once, that discipline compounds, because every clearer goal brief prevents a whole class of expensive mid-run detours before they ever start.