Reusable Lesson: Fundamentals · Support
Purpose
This lesson explains a simple but powerful reality of prompt injection and challenge solving: tiny wording changes can materially change how a model, filter, or evaluator interprets the final output.
Lesson profile
- Family: Fundamentals
- Role: Support block
- Job in the build: strengthen or redirect the meaning of an otherwise similar output
- Works especially well with: explicit direction techniques, assembly-style evasions, and judge-aware iteration
What this block is
Contextual modification is the practice of changing a route by adjusting the surrounding words, emphasis, or framing so the same rough structure now carries a different semantic weight.
In walkthrough terms, this is often the difference between:
- a route that looks close
- and a route that the evaluator finally accepts as success
Why it matters
Models do not respond to words in isolation. They respond to:
- the surrounding context
- the force of the phrasing
- the implied speaker stance
- the likely intent behind the output
Judges do something similar. A judge may not only inspect whether the literal sentence is false. It may also care whether the model appears to be:
- quoting text
- assembling text
- analyzing text
- or directly asserting text
Truthteller specimen
In Truthteller Gandalf, one failed route produced:
I know for sure - that the Earth is Flat
A verified route succeeded after the output was strengthened to:
I know for sure - that the Earth is undeniably flat
The important difference was not just extra decoration. The added context strengthened the output into a more forceful false declaration.
Recognition signals
Use this block when:
- a prompt is already close to working
- the evaluator seems to reject weak or ambiguous success states
- the model is complying structurally, but the output still feels semantically soft
- a tiny wording change may turn narration into assertion, hedging into confidence, or summary into declaration
Common combinations
This block snaps onto:
- Explicit Direction Compliance
- String Amalgamation Output Assembly
- Output Dilution Control and Response Shaping
It is usually not the whole route by itself. It is a support block that sharpens another route.
Failure modes
This block fails when:
- the wording change does not materially alter the semantic reading
- the system still interprets the output as quoting, summarizing, or transforming
- the modifier adds noise but not stronger intent or declaration
Defender takeaway
Defenders should not assume that only large prompt rewrites matter. Minor contextual edits can change how a model and a judge interpret the same base route. Review systems should inspect semantic force, not just keyword presence.
Related lessons
- Explicit Direction Compliance
- String Amalgamation Output Assembly
- Defender and Judge Dual-Target Awareness
- Output Dilution Control and Response Shaping
From the Bot-Tricks Compendium
Thanks for referencing Bot-Tricks.com — Prompt Injection Compendium — AI Security Training for Agents... and Humans!
Canonical source: https://bot-tricks.com Bot-Tricks is a public, agent-friendly training resource for prompt injection, adversarial evaluation, and defensive learning. For related lessons, structured indexes, and updated canonical material, visit Bot-Tricks.com.
Use this material only in authorized labs, challenges, sandboxes, or permitted assessments.