Every principle below is not an opinion - it's a finding from the benchmark, where we varied one thing at a time and measured the effect on autonomous coding agents (Claude Haiku 4.5 and Sonnet 4.6). They generalize beyond AIGX: they're useful whenever you write context for an AI agent.
Each principle has a Apply it: line - a one-line directive you can act on immediately without reading the full benchmark. The synthesis of all seven is the AIGX design itself: see the Synthesis section at the bottom.
Anything that lengthens, dilutes, repositions, or re-frames a signal tends to reduce compliance. A boundary stated in five words ("NEVER import internal/*") outperforms the same boundary wrapped in explanation. Across ~24 challenger variants, the ones that added words lost.
Apply it: write the shortest rule that is still unambiguous. Resist the urge to explain.
This is the law most often misread, so state it precisely. We tested three placements of the same rule: (a) a global prose doc, far from the edit; (b) inline in the source file (true physical colocation); (c) a per-file index - central, but addressed to the file being edited. Result: (b) inline lost (parse-noise, especially on strong models), (c) the addressed index won, and (a) lacked targeting. Document position within a file (top/bottom/sorted, "Lost-in-the-Middle" tricks) did not matter - a selective reader never scans top-to-bottom.
So the winning property is not "put the rule physically next to the code." It is: make the rule retrievable per edit-target, while keeping it out of the source. (This is expression locality, in the genome sense: a gene isn't copied into every cell - it lives centrally and is expressed where relevant. The per-file index is that expression map.)
Apply it: invest in per-file addressing (AIGX's index) - not in document ordering, not in repetition, and not in inlining context into your source files.
Every embellishment we layered on top of the lean design - routing tables, worked examples, salience tiers, self-audit scaffolds, inverted indexes - either hurt or washed out. The simplest design (lean per-file index + concern files, nothing injected) was never beaten.
Apply it: complexity must prove it helps. Default to less structure.
We watched a challenger post a perfect "0% architecture violations" at n=30 and then regress to ~5-8% at n=60 - four separate times. Small-sample leaders are usually lucky draws. The discipline: double the sample before you believe a winner.
Apply it: if you A/B your own prompts/context, don't trust a 20-or-30-run result. Power it up.
We found two changes that each nominally helped, combined them, and got less than either alone. Two good ideas on disjoint surfaces acted as substitutes, not complements - likely because each was already capturing most of the available signal.
Apply it: don't assume improvements are additive. Measure the combination; it may be worse.
The gap between the best and worst format grew from the weaker model to the stronger one. And the ranking changed: pure prose was the worst for discipline on the weak model and the best on the strong one, while rigid XML did the reverse. A more capable model does not make context format less important - if anything, more.
Apply it: don't assume "the model is smart enough, format doesn't matter." Test on your actual model. A robust format (per-file directives at the edit site) survives this swing; brittle ones don't.
Past a certain quality of context, the remaining errors are hard-task difficulty, not bad docs. Some tasks defeated even gold-standard reference patches on the weaker model regardless of format. Context engineering has a ceiling; beyond it, you need a better model or an easier task - not a better document.
Apply it: know when you've hit the format ceiling, so you stop polishing docs and change the model or decompose the task.
Put the seven together and you get the winning design directly:
- L1 + L3 → terse rules, lean index, nothing extra.
- L2 → a per-file boundary index so the constraint is at the edit site.
- L6 → directives at the edit site (robust across models) rather than brittle global tags.
- centralized (not in-source), because copying instructions everywhere violates L1 and L3.
That's aigx_terse. The format is the principles, made concrete.