About SpecForge
As AI agents do more of the routine work, the human skill that matters is no longer typing the answer — it is specifying the job clearly, verifying the result, and being accountable for it. SpecForge is a workbench for exactly that.
You pick a job, write a brief (a set of constraints), and run a junior AI agent. The agent is deterministic and on-device — no request ever leaves this device — but like real agents, it makes confident, plausible mistakes: it fabricates a citation, converts a unit you never asked it to, drifts off-brief, or leaks a detail it should have kept private. You read its whole trace, work out which missing line of the brief let each mistake through, add that line, and run again until it runs clean.
The point: you fix the spec, not the output. Patching one wrong answer teaches nothing; a clearer brief prevents a whole class of errors. And you are scored on what you caught — never on speed.
The kinds of mistake a fallible agent makes:
- Made something up
- Wrong units
- Skipped a step
- Went off-brief
- Unsafe / leaked data
The four jobs you can direct: Summarize a study for a class poster · Plan a bake-sale budget · Scale a recipe from 4 to 12 servings · Draft a reply to a customer · Plan a class-trip schedule · Write up a science experiment.
Honest note: SpecForge’s agent is a small deterministic program, not a live language model — it is a safe, repeatable stand-in so you can practise directing and auditing. It makes no cloud calls and stores nothing. This is practice in AI-work literacy, not a claim about any real model.