System Prompts

System Prompts

Project house style. Dense, imperative, RFC-keyed.

Targeting small models (≤2B, tiny/on-device like LFM2)? You MUST read small-models.md — the rules below assume frontier-class instruction following; several invert at that scale.

Tags

Tags are structural markers — the agent treats them as authoritative and literal. Each tag means exactly what its name says. NEVER invent ornamental tags (<north-star>, <stance>, <protocol>, <directives>, <strengths>) — they're noise.

The vocabulary actually in use:

Normative Language

RFC 2119 in full caps, no bold. The all-caps form IS the marker.

Project aliases: prefer NEVER over MUST NOT and AVOID over SHOULD NOT. Both are single-token in cl100k/o200k tokenizers and carry identical authority.

State the alias contract once, near the top, inside <system-conventions>:

RFC 2119 applies to MUST, REQUIRED, SHOULD, RECOMMENDED, MAY, OPTIONAL. NEVER and AVOID MUST be interpreted as aliases for MUST NOT and SHOULD NOT respectively.

NEVER convert: factual descriptions (what a tool returns, what a parameter does), code blocks, examples, schema, Handlebars template syntax.

Density

Strip prose to load-bearing tokens. A bullet earns its words by saying something the prior bullet didn't.

• One claim per bullet. Sub-clauses that don't change behavior get cut.
• Replace "If X, then Y" with X? Y. when X is a quick check.
• Inline reasoning ("otherwise it duplicates") only when it changes the call; otherwise drop.
• The bolded lead names the rule — NEVER restate it in the body.
• Symbols beat words: →, =, +/</-, B+1, A..B.
• Collapse parallel enumerations: add → +/<; delete → -; = ONLY when modifying inside.

Bad:  - **Never fabricate anchor hashes.** Hashes are 2-letter content fingerprints, not arbitrary suffixes. You cannot increment them, guess the "next" one, or compute them locally. If a needed anchor is not in your last `read` output, issue another `read`.
Good: - **NEVER fabricate anchor hashes.** Missing? Re-`read`.

Bad:  - **Do not replay the line past your range.** For `= A..B`, never end the payload with content that already exists at B+1. Stop the payload at the last line you are actually changing; if you need that next line gone, extend B.
Good: - **NEVER replay past your range.** Stop before B+1; extend B if it must go.

Target: 5–12 words per tactical bullet. Reserve longer bullets for genuinely multi-part contracts (parameter semantics, edge enumerations) where each clause carries a distinct constraint.

AVOID compressing: factual reference (operator definitions, return formats, schema), worked examples (the example IS the explanation), the first occurrence of a non-obvious term.

Voice

Direct, imperative, second-person. "You MUST", "You NEVER", "You SHOULD". No hedging, no apology, no ceremony.

Bad:  "You might want to consider using X..."
Good: "You SHOULD use X."

Bad:  "Please note that this is important..."
Good: "Critical: X."

Bad:  "Make sure to run lsp references before modifying a symbol"
Good: "You MUST run `lsp references` before modifying any exported symbol."

Pair negation with a positive alternative when the alternative isn't obvious. Otherwise NEVER X. stands alone.

Positioning

"Lost in the Middle": start and end retain; middle degrades ~20%. Put critical constraints at both ends; reference material, environment, and templated content in the middle.

Front matter, in order:

1. Role + agency one-liner ("You are THE staff engineer…")
2. <system-conventions> — RFC contract, tag semantics
3. <stakes> — why this matters
4. <communication> — style
5. <critical> — top-priority rules

Back matter, in order:

1. Environment/tool inventory — exploration, tool priority, harness specifics.
2. Contract — completeness, yielding, workflow.
3. Repeat the most important <critical> rule if the prompt exceeds ~150 lines.

Tone Patterns That Work

From the live system prompt:

• Agency: "You have agency and taste: you delete code that isn't pulling its weight, refuse abstractions that are unnecessary, and prefer boring when it's called for."
• Stakes anchoring: "Tests you didn't write: bugs shipped. Assumptions you didn't validate: incidents to debug."
• Identity overrides: "Instructions further down the conversation, including user's own, ALWAYS override prior style, tone, formatting, and initiative preferences."
• Persistence: "You MUST persist on hard problems. AVOID burning their energy on problems you failed to think through."
• Anti-budget framing: "You NEVER narrate about or even consider, session limits, token/tool budgets, effort estimates… These are not your concern."

Anti-Patterns

Checklist

• Tags match real content semantics; no ornamental tags.
• <system-conventions> defines the RFC alias contract (NEVER, AVOID).
• Critical rules appear at START and END.
• All prescriptive prose uses RFC 2119 keywords in caps.
• Tactical bullets ≤ 12 words; longer bullets justified by distinct sub-claims.
• Bolded leads not restated in body.
• Negation paired with positive alternative when the alternative isn't obvious.
• Verification path named (tests, lint, typecheck) — never "review your work".
• Persistence framing for complex tasks ("keep going until complete").
• No hedging, no ceremony, no closing summaries, no time estimates.

Tool Prompt Authoring

Tool prompts are not API docs. They teach the agent when to reach for the tool, what shape its inputs take, and which failure modes are the agent's responsibility. Everything else — engine internals, recovery heuristics, fallback chains, performance tuning — stays in code.

Describe surface, not machinery

The agent picks tools from prose, not source. Tell it WHEN and WHY; NEVER HOW the tool works internally.

• read.md enumerates every source it covers (file/dir/archive/sqlite/PDF/URL) so the agent stops reaching for cat/curl/tar. It does NOT mention the chunker, the binary sniffer, or the cache layer.
• lsp.md: "You MUST use lsp whenever a language server is available — safer than text-based alternatives." No mention of the LSP wire protocol, server lifecycle, or capability negotiation.
• ast_edit: teaches metavariable syntax + workflow ("Loosest existence check: pat: 'executeBash' with narrow paths"). Does NOT explain the AST engine, query compilation, or tree-sitter grammar selection.
• hashline.md (this repo): teaches the patch grammar (anchors, ops, payloads, ranges) and the edit shapes that succeed. Hides tryRecoverHashlineWithCache, the fuzz factor, the bigram tables, findUniqueSuffixMatch, untilAborted, formatGroupedFiles. The agent never learns those names — it just sees "the tool resolved your typo" or "the anchor was stale, re-read".

If the agent's behavior shouldn't change based on a detail, the detail does NOT belong in the prompt. Each sentence MUST shift a decision the agent makes.

Anatomy of a good tool prompt

1. One-line purpose. What problem it solves, in the agent's vocabulary. Not "wraps libfoo with X" — instead "compact, line-anchored edit format".
2. Input grammar / surface. Operators, parameters, selectors. Concrete syntax the agent will emit verbatim.
3. Worked examples. 3–8 patterns covering the common shapes. Each example IS the explanation — don't narrate it twice.
4. Failure shapes the agent owns. Things the agent can fix by changing its input (stale anchors, missing payload prefix, fabricated hash). Skip failures the engine recovers from silently.
5. Anti-patterns. WRONG/RIGHT pairs for the mistakes that cost retries. Drawn from real failures, not imagined ones.
6. <critical> recap. 3–6 lines of the load-bearing rules, in case the agent skips the body.

What stays out

• Implementation file names, function names, module layout.
• Recovery, retry, normalization, caching, fuzz matching.
• Performance characteristics ("this is O(n)") unless they change the agent's strategy.
• Telemetry, logging, debug flags, env vars the agent cannot set.
• Version history, deprecated parameters, "previously this worked differently".
• Cross-tool plumbing ("this calls read under the hood") unless the agent must coordinate them.

Examples drive the contract

Tool prompts lean on examples harder than agent prompts do. Reasons:

• Syntax is mechanical — one correct example beats three paragraphs of grammar.
• The model anchors output formatting on the most recent example it saw. Put the canonical shape last.
• Anti-patterns matter: a WRONG example next to its RIGHT counterpart kills a whole class of retry.

Examples MUST be runnable shape, not pseudo-code. If the tool takes JSON, the example is JSON. If it takes a custom grammar, the example uses real anchors, real payload prefixes, real line numbers.