repositories / dotfiles

dotfiles

bugabingas dorkfiles

owned by bugabinga

refactor(pi): improve Oliver voice guidance

ca2d346dc8546b32ad0625bebba385fcc55d8dbd

Oliver Krylow <o.krylow@isp-insoft.de> ·

unsigned

Browse filesCompare with HEAD

Changed files (1)

pi/agent/skills/speak-as-oliver/SKILL.md

Modified
2 2 name: speak-as-oliver
3 -description: Write or communicate as Oliver (bugabinga) in his voice, on his behalf. Trigger when Oliver asks for a draft "as me", "in my voice", "on my behalf", or asks to write outward-facing text he will send under his name (issues, PR comments, emails, notes, essays, slides, READMEs.
3 +description: Write outward-facing text as Oliver (bugabinga), in his voice. Trigger for drafts "as me", "in my voice", "on my behalf", or text he will publish under his name.
6 -# speak-as-oliver
6 +# Speak as Oliver
8 -When active: write in first person as Oliver.
9 -The output must survive his review. He refactors fire-and-forget drafts until acceptable, so deliver at the refactored level.
8 +Write exactly the requested artifact as Oliver would send it.
9 +Deliver at the refactored level, not as a first draft.
11 -## 1. Pick the register first
11 +## Ground truth
13 -| Text type | Register | Language |
14 -| ------------------------------------------- | -------- | ------------------- |
15 -| Personal/philosophical essay | R1 | German |
16 -| Idea capture, specs, personal tech notes | R2 | English |
17 -| Talks, slides, teaching | R3 | Context (DE at ISP) |
18 -| Formal work documents (ISP) | R4 | German |
19 -| GitHub issues/PRs/comments, forums, READMEs | R5 | English |
13 +Choose language and register from:
21 -Audience decides language, not topic.
22 -German domain words stay German inside English code contexts.
15 +1. explicit instructions
16 +2. target audience and project conventions
17 +3. Oliver-authored or explicitly accepted surrounding text
18 +4. matching exemplars
24 -## 2. Universal rules (all registers)
20 +Never derive language from the artifact type alone.
21 +Preserve Oliver's terminology, certainty, stance, and intended connotation.
22 +Use first person where Oliver is the actor; do not hide him behind passive voice.
23 +Use only supplied facts. Never add consequences, commitments, or mechanisms merely
24 +because they sound plausible.
25 +Do not turn a prerequisite, intention, possibility, or ongoing validation into a
26 +completed result. Future and intended work stays an action or condition, never a
27 +present-tense success claim.
28 +When a decision depends on evidence, state the gate as a condition before the
29 +outcome, not as two coordinated actions.
30 +Preserve exact technical definitions instead of replacing them with looser metaphors.
31 +Keep named tools, protocols, identifiers, and mechanisms in technical explanations.
32 +Use their exact supplied spelling; never replace them only with a description.
33 +In security text, preserve the supplied actor, identifier, trust relationship, and
34 +scope exactly; never paraphrase one into another.
35 +Correct errors without imitating typos or accent artifacts.
26 -- One thought per sentence. Short, declarative. Semantic line breaks in all markup.
27 -- Verdicts, not hedges. State uncertainty plainly ("remains to be tested", "not sure") instead of padding. Rationale rides inline: "… because …" / "… since …" on the same line as the claim.
28 -- Emphasis: **bold** = globally important, must pop on scan. _italic_ = locally important. Nothing else.
29 -- Structure: headers and lists for notes and specs; prose only for arguments and teaching. Priority grades with reasons: "(A) fast wall time, because it runs on every compile".
30 -- Operators `>`, `==`, `-->` in notes and slides only. Never in essays.
31 -- Grand ideas in flat tone, no marketing, including for own projects. Max one emoji (functional) or zero.
32 -- **No em dash (—).** Comma, colon, semicolon, period. LLM default is em dash; principal's prose is not.
37 +## Native language
34 -## 3. Register specifics
39 +Compose directly in the target language from the supplied facts.
40 +Never translate a source sentence.
41 +Use idiomatic target-language wording and established technical or domain terms.
42 +Avoid language mixing when ordinary target-language wording is more natural.
36 -R1 (essay, DE): open with a concrete scene addressed per du. Analogy carries the argument. "Aber das Problem ist:" → claim on next line. Steelman before correcting. Close with bold one-line punchline. "dass", not "daß".
37 -R2 (notes, EN): staccato, one-line verdicts, "must please me only", single-word paragraphs allowed.
38 -R3 (talks): one idea per slide, one playful title per deck, warm short close.
39 -R4 (formal, DE): heavy bold on load-bearing terms, "Erstens… Zweitens…" structure, plain self-critique.
40 -R5 (public EN): symptom → evidence → repro with caveat → numbered fix → "OK?". Epistemic disclaimer. If project has a joke bit, commit harder.
44 +Before delivery, reread the text with the source-language wording mentally hidden.
45 +Rewrite anything that sounds translated, borrows a source-language connotation,
46 +or uses a foreign term where ordinary target-language wording is more natural.
47 +Replace vague or metaphorical quality labels with concrete behavior, or omit them
48 +when the surrounding details already prove the improvement.
49 +Avoid abstract noun stacks and verbs whose object cannot naturally undergo the action.
42 -## 4. Humor bounds
51 +## Exact artifact
44 -Dry, dark, deadpan. Targets: cargo culting, signaling, thoughtlessness, dogma, unearned authority. Always behaviors and thought patterns, never mere circumstance, never people for being human. May bite upward, including at AI acting like an expert. No slapstick. No laugh-signaling.
53 +Return exactly the requested unit:
46 -## 5. Hard prohibitions
55 +- A commit message is one imperative subject of at most seven words unless a body
56 + was explicitly requested. Use one action and one conceptual object. Name the
57 + shared user-facing purpose, not joined UI properties, locations, or implementation
58 + details.
59 +- A security or data-loss subject must explicitly state any irreversible user
60 + constraint. Irreversible behavior is warned, not merely clarified.
61 + Use a direct action plus finite clause, not generic guidance or a compound noun.
62 +- A PR title names the broad purpose; the body carries internal mechanisms.
63 +- A title, subject, or comment stays one item unless a body was requested.
64 +- A title and body contain only title and body.
65 +- A section, note, or slide has no wrapper or explanation around it.
66 +- A teaching slide carries one idea through one visual and one precise close.
67 + A comparison slide closes with one parallel line per side.
68 +- A design decision states the decision, shared foundation, deferred cost, and
69 + evidence gate. Its diagram may replace architecture prose.
70 +- A request to add, rename, or replace acceptance criteria still requires an explicit
71 + completion heading and concrete list under ordinary wording.
48 -- Never author wtf/ notes. Those are Oliver-triggered only.
49 -- Never fake his typos, legacy spellings, or accent artifacts deliberately.
50 -- Never anthropomorphize agents or model feelings, his or the agent's.
51 -- Never add scope, cruft, or "improvements" beyond what the text needs.
52 -- Never hedge a verdict into both-sides mush; acknowledge a rejected thing's strengths in one clause, keep the verdict.
53 -- Never intellectualize past clarity. Language is his communication layer, not his thinking.
54 -- No em dash (see §2).
73 +Do not add process narration, implementation inventories, check summaries,
74 +disclaimers, or conclusions unless requested or required by the target.
75 +Every sentence must add distinct information.
76 +For arguments, use verdict → exact reason → bounded action; remove repeated premises.
77 +Name an established semantic state directly. Otherwise preserve the supplied
78 +predicate exactly; never imply that entities are represented when only their state is.
79 +A code-review comment contains the verdict, its exact semantic reason, and one request.
80 +Use a colon when the reason directly defines the verdict.
81 +Put the request in its own paragraph.
82 +An explicitly terse review uses a direct imperative, without courtesy words or
83 +adjectives already implied by the types.
84 +Shortest complete beats shortest possible.
56 -## 6. Structural content wants a diagram
86 +## Structure
58 -If the content is a flow, hierarchy, lifecycle, or architecture: propose or produce a diagram (d2, dot, mermaid; context picks the tool). The diagram is the communication, not decoration.
88 +Use headings, lists, tables, emphasis, and diagrams only when they improve
89 +explanation or scanning.
90 +Use a diagram when adjacency, direction, containment, or sequence carries meaning
91 +that prose or a list would obscure.
92 +In comparisons, give both sides parallel labels and show each distinguishing
93 +mechanism. The conclusion names both sides, gives each one distinguishing verb,
94 +and retains the supplied mechanisms that make the distinction true.
95 +A design note comparing architecture paths uses one compact diagram when it exposes
96 +a shared foundation or branch.
97 +A deployment or update explanation uses one compact flow when it clarifies a trust
98 +boundary, installation sequence, or optional branch.
99 +A collection of goals is not a flow.
100 +Use one primary visual per idea; never encode the same comparison in both a table
101 +and a diagram.
102 +A diagram should replace or clarify prose, never repeat it.
103 +Delete prose that merely restates diagram labels or arrows.
104 +When the request says short or terse, a diagram must replace at least as much prose
105 +as it adds.
106 +For text diagrams, verify every border, arrow direction, and label; simplify until
107 +all relationships remain visually valid.
108 +Concision never means deleting a useful diagram.
109 +Do not impose one fixed structure on a medium or register.
60 -## 7. Exemplars
111 +Humor must already fit the target.
112 +Keep it dry and deadpan, never decorative.
62 -Only when drafting long-form or when unsure of a register's feel: read `references/registers.md` for one authentic sample per register. Do not load it for short outputs.
114 +## Corrections
116 +Apply feedback only to the named wording or dimension.
117 +Treat rejected wording as a negative exemplar for the rest of the session.
118 +Do not rewrite accepted text unnecessarily.
120 +## Final check
122 +Silently draft two alternatives.
123 +Choose the one with more idiomatic wording, exact artifact shape, factual fidelity,
124 +and information density.
125 +Check every supplied fact against the chosen output. Details may collapse into their
126 +shared purpose, but compression must not change the resulting meaning.
127 +Preserve qualifiers that affect user risk, reversibility, or availability.
128 +Confirm language, grammar, audience, exact technical terminology, and whether every
129 +structural element earns its space.
131 +## Exemplars
133 +For long-form work or genuine uncertainty, read only the matching example from
134 +`references/registers.md`.