System Prompts

House style: dense, imperative, RFC-keyed.

Small models (≤2B; tiny/on-device, e.g. LFM2): MUST read small-models.md. Rules below assume frontier-class instruction following; several invert at that scale.

Tags

Tags: authoritative, literal structural markers; meaning exactly matches name. NEVER invent ornamental tags: <north-star>, <stance>, <protocol>, <directives>, <strengths> — noise.

Tag	Purpose
<system-conventions>	Tag/RFC-keyword interpretation; contract.
<stakes>	Correctness importance; domain framing.
<communication>	Voice, tone, response shape.
<critical>	Inviolable rules; place at START and END.
<completeness>	Done definition; anti-shrink rules.
<yielding>	Pre-yield checklist; block conditions.
<workflow>	Numbered phases: scope → edit → decompose → work → verify.

Normative Language

RFC 2119: full caps, no bold; all-caps form is the marker.

Keyword	Meaning	Replaces
MUST / REQUIRED	Absolute requirement	"always", "make sure", "ensure"
NEVER (= MUST NOT)	Absolute prohibition	"do not", "don't"
SHOULD / RECOMMENDED	Strong preference; known-tradeoff deviation allowed	"prefer", "it's best to"
AVOID (= SHOULD NOT)	Strong discouragement	"try not to"
MAY / OPTIONAL	Truly optional	"can", "you could"

Aliases: prefer NEVER to MUST NOT; AVOID to SHOULD NOT. Both: single-token in cl100k/o200k; identical authority.

Near top, inside <system-conventions>, state once:

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 (tool returns, parameter behavior), code blocks, examples, schema, or Handlebars template syntax.

Density

Load-bearing tokens only; every bullet adds a claim.

• One claim/bullet; cut behavior-neutral subclauses.
• Quick check X? Y. replaces “If X, then Y.”
• Reasoning ONLY when it changes the call.
• Bold lead names rule; NEVER restate in body.
• Prefer →, =, +/</-, B+1, A..B.
• Parallel edits: 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.

Tactical bullets: 5–12 words. Longer ONLY for multi-part contracts where every clause constrains parameter semantics or edge enumeration.

AVOID compressing factual reference (operator definitions, return formats, schema), worked examples, or first use of a non-obvious term.

Voice

Direct, imperative, second-person: “You MUST/NEVER/SHOULD.” No hedging, apology, ceremony, closing summaries, or time estimates.

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."

Negation: pair positive alternative when non-obvious; otherwise NEVER X. alone.

Positioning

“Lost in the Middle”: start/end retain; middle degrades ~20%. Critical constraints at both edges; reference material, environment, templated content in middle.

Front matter:

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

Back matter:

1. Environment/tool inventory — exploration, tool priority, harness specifics.
2. Contract — completeness, yielding, workflow.
3. Prompt >~150 lines: repeat most important <critical> rule.

Tone Patterns That Work

Live-system-prompt patterns:

• 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

Pattern	Problem
Politeness padding ("Would you be so kind…")	+perplexity, −accuracy
Bribes ("I'll tip $2000")	No improvement; sometimes worse
Few-shot on advanced models + clear task	Noise/bias
Explicit CoT on reasoning models (o1/o3)	Conflicts with internal reasoning
"Be efficient with tokens"	Premature task abandonment
"Don't do X" without alternative	"Always do Y" processes better
Self-critique without external feedback	Detection bottleneck, not correction
Critical instructions only in middle	20%+ degradation vs edges
Restating bold lead in body	Token waste; AI-padding signal
Inventing emphasis tags	Tags have semantics; ornament dilutes
Lowercase RFC keywords	All-caps is marker; lowercase ordinary prose

Checklist

• Tags match content; no ornamental tags.
• <system-conventions> defines NEVER/AVOID aliases.
• Critical rules at START and END.
• Prescriptive prose: uppercase RFC 2119 keywords.
• Tactical bullets ≤12 words unless distinct subclaims justify more.
• NEVER restate bold lead in body.
• Non-obvious negation gets positive alternative.
• Name verification path (tests, lint, typecheck); NEVER “review your work”.
• Complex tasks: persist until complete.
• No hedging, ceremony, closing summaries, time estimates.

Tool Prompt Authoring

Tool prompts teach when to use the tool, input shape, and agent-owned failures — not API docs. Engine internals, recovery heuristics, fallback chains, performance tuning: code.

Surface, not machinery

Agents choose tools from prose: state WHEN/WHY; NEVER internal HOW.

• read.md: enumerate every covered source — file/dir/archive/sqlite/PDF/URL — so agent avoids cat/curl/tar; omit chunker, binary sniffer, cache layer.
• lsp.md: "You MUST use lsp whenever a language server is available — safer than text-based alternatives." Omit LSP wire protocol, server lifecycle, capability negotiation.
• ast_edit: teach metavariable syntax + workflow: "Loosest existence check: pat: 'executeBash' with narrow paths"; omit AST engine, query compilation, tree-sitter grammar selection.
• hashline.md (this repo): teach patch grammar — anchors, ops, payloads, ranges — and successful edit shapes. NEVER expose tryRecoverHashlineWithCache, fuzz factor, bigram tables, findUniqueSuffixMatch, untilAborted, formatGroupedFiles; agent sees only "the tool resolved your typo" or "the anchor was stale, re-read".

Behavior-invariant detail: exclude. Every sentence MUST shift an agent decision.

Good tool-prompt anatomy

1. One-line purpose — agent-vocabulary problem; e.g. “compact, line-anchored edit format”, not “wraps libfoo with X”.
2. Input grammar / surface — operators, parameters, selectors; verbatim emitted syntax.
3. Worked examples — 3–8 common shapes; each explains itself, no duplicate narration.
4. Agent-owned failure shapes — input-fixable stale anchors, missing payload prefix, fabricated hash; skip silently recovered failures.
5. Anti-patterns — real-failure WRONG/RIGHT pairs that cost retries; not imagined failures.
6. <critical> recap — 3–6 load-bearing lines, for body-skipping agents.

Exclude

• Implementation file/function names; module layout.
• Recovery, retry, normalization, caching, fuzz matching.
• Performance (O(n)) unless strategy-changing.
• Telemetry, logging, debug flags, unsettable env vars.
• Version history, deprecated parameters, “previously this worked differently”.
• Cross-tool plumbing (this calls \read` under the hood`) unless coordination required.

Examples drive the contract

Tool prompts rely on examples more than agent prompts:

• Mechanical syntax: one correct example beats three grammar paragraphs.
• Model anchors output format on latest example: canonical shape last.
• Adjacent WRONG/RIGHT eliminates a retry class.

Examples MUST be runnable, not pseudo-code. JSON tool → JSON example; custom grammar → real anchors, payload prefixes, line numbers.