Ancient wisdom for modern promptcraft
These inscriptions are not immutable law, but living wisdom—patterns observed in the wild, experiments that yielded fruit, warnings whispered by those who ventured before. Context engineering shifts like smoke; what works today may dissolve tomorrow as the oracles evolve and new incantations emerge.
Treat these as starting runes, not sacred doctrine. Test them. Adapt them. Discard what fails thee. The true grimoire is written through practice, not proclamation. What follows are signposts in fog—useful until you find your own path.
Build thy prompts incrementally. Start simple, test often, and refine based on results. The most powerful spells are forged through patient iteration.
Extract project-specific patterns into reusable components. Let thy coding standards, API schemas, and architectural decisions become living knowledge.
Combine multiple prompts and contexts into named pipelines. Each configuration becomes a reusable spell, ready to be invoked with mystical consistency.
"When the path becomes unclear, return to the beginning."
Clear the LLM's short-term memory (/clear) often.
Prune or restate your context. Review your history. This ward
reestablishes clarity and prevents hallucination entropy.
Invoke pluqqy archive <pipeline or component>
to banish unused artifacts to the shadow realm. A cluttered
workspace clouds the sorcerer's focus—keep only what serves thy
current craft.
Craft thy PROMPTS with restraint—leave room for spontaneous
direction. Reusable pipelines provide foundation; adhoc
instructions in thy AI assistant grant precision. Speak thus:
"@PLUQQY.md Let's add a new feature that enhances the
scrolling behavior"
to blend general wisdom with specific intent.
Context is finite—every word depletes thy assistant's attention budget. Seek the smallest set of high-signal tokens that conjure thy desired outcome. A lean pipeline cuts through noise; bloated instructions breed confusion. Start minimal, then add specificity only when the oracle falters.
Write thy contexts in simple, direct language. Eschew baroque phrasing and needless formality—the assistant seeks clarity, not poetry. Strike balance: too rigid and thou stifle creativity; too vague and thou invite chaos. Let each instruction carry weight.
Feed thy assistant only what serves the present moment. Rather
than loading entire codebases into context, reference specific
files as needed with @filename. Let the
conversation unfold progressively—each exchange revealing the
next necessary fragment. The sorcerer who hoards all knowledge
upfront wastes precious attention on irrelevant details.
Reference @PLUQQY.md consistently in thy AI
sessions—one file, infinite configurations. Switch pipelines
freely with pluqqy set, yet thy reference remains
constant. The assistant knows not which spell was active before;
only what lives in the present moment. This discipline prevents
confusion and preserves precious context space.
"Frame thy context as the world should be, not merely as it is."
The assistant responds to framing—cast thy architectural decisions and standards as present reality, not distant aspirations. When writing contexts, speak of patterns as established truth. Beware: contradictory statements weaken the spell and breed confusion.
"A name is a spell unto itself. Speak intent, not implementation."
Choose descriptive names that reveal purpose:
review-api-changes over pipeline-3.
Good names are self-documenting runes—they guide both thee and
thy future self. Use verbs. Be specific. Clarity in naming is
clarity in thought.
"In XML's embrace, the oracle finds clarity amidst chaos."
Wrap thy complex instructions in XML tags to give the assistant
clear boundaries. Use tags like
<context>, <task>,
<example>, and <constraints>
to partition thy intent. The assistant sees structure where
mortals see mere words—these markers illuminate thy meaning and
prevent ambiguous interpretations. Especially when crafting
elaborate prompts, let XML be thy scaffolding.
"Trust, but verify. The sorcerer who blindly imports foreign incantations invites ruin."
When importing components and pipelines from distant realms, treat each line as potentially cursed. Read every instruction. Examine every command. Hostile actors may hide prompt injections, LLM poisoning, or malicious directives within seemingly innocent configurations. Better to rewrite in thine own hand than to wake one morning with thy secrets stolen or thy assistant corrupted. The safest pipeline is one thou hast crafted thyself.