iterion

Prompt-engineering cursors

Cursors are framing dials, not gates. They adjust the tone and emphasis of an LLM node’s system prompt — ambition, depth, rigor, escalation threshold — without forking the underlying prompt: block. Goodhart-resistance still lives in the judge architecture and the scanner/coverage gates documented in workflow_authoring_pitfalls.md; cursors are calibration, not enforcement.

Read that warning twice. The most common cursor misuse is “lower rigor so the failing judge passes.” That is a façade. Fix the issue or adjust the gate.

What a cursor looks like

A cursor declares a name plus either an enum of named values or a numeric band map over the unit interval [0, 1]. Each value or band carries a prompt fragment that gets appended to the agent’s system prompt under a ## Calibration section when the cursor is activated.

cursor ambition:
  description: "How aggressively to expand scope beyond the stated request."
  values:
    cautious: "Stick strictly to the stated request; do not propose extensions."
    balanced: "Address the request; surface one obvious adjacent improvement."
    ambitious: "Proactively identify 2-3 adjacent improvements; rank by leverage."

cursor depth:
  description: "Investigation thoroughness."
  bands:
    "0.0..0.33": "Skim the surface; cite primary sources only."
    "0.34..0.66": "Examine main code paths and one layer of dependencies."
    "0.67..1.0":  "Trace all call sites and edge cases."

A cursor declares either values: (enum) or bands: (numeric) — not both. The IR compiler rejects the both-form with C085.

Activating cursors on a node

agent reviewer:
  model: "anthropic/claude-sonnet-4-6"
  system: review_system
  user: review_user
  cursors:
    enabled: true
    ambition: ambitious      # enum lookup
    depth: 0.7               # numeric → matches the 0.67..1.0 band
    rigor: ${ITERION_RIGOR}  # env-substitution allowed

When enabled: false (or the block is absent), no ## Calibration section is emitted — cache-stable, zero cost.

The runtime resolves each setting like so:

Expand ${VAR} against the process env (same rules as reasoning_effort).
If the value parses as a float, take the numeric path. Otherwise take the enum path.
Numeric path: clamp to [0, 1]. If the cursor has bands:, return the prompt of the first band whose [lo, hi] contains the value. If it has values: only, snap to the enum position by floor(v * len).
Enum path: look up the value by name against the cursor’s values: list. Misses fail at compile time (C084).

Resolved fragments are sorted alphabetically by cursor name and appended under a ## Calibration section as **<CursorName>:** <fragment> lines. Sorting is deterministic so the resulting prompt is prompt-cache-friendly across runs.

Diagnostics

Code	Severity	Meaning
`C083`	warning	Agent/judge references a cursor name that is not declared at workflow scope.
`C084`	error	Cursor value is invalid: not in the enum, outside `[0, 1]`, or no matching band. Skipped for `${VAR}` values (runtime handles substitution).
`C085`	error	Cursor declaration is malformed: neither `values:` nor `bands:`, both forms set, bands overlap, or a range falls outside `[0, 1]`.
`C086`	error	Duplicate cursor name within a single workflow source.

Risks and guardrails

Goodhart on numerics

Operators will treat depth: 0.9 as “more = better” and ratchet. Resist this. Band fragments must stay qualitative: describe what the model should do ("Trace all call sites"), not what it should output ("Produce at least 5 references"). Quantitative outputs belong to deterministic scanners, not to framing prompts.

Prompt bloat

Four enabled cursors add roughly 4 × 60 tokens to every node call. The soft cap is 3 cursors per node. Beyond that, the calibration section starts to dominate the base prompt — either consolidate cursors or accept the cost intentionally.

Cache-toggling

Flipping enabled or changing any setting value invalidates the prompt-cache for that node. The runtime sorts cursor fragments alphabetically so identical activations across runs share their cache entry, but cross-run drift of values still costs cache hits.

False precision in hybrid mode

When a numeric cursor (bands:) is invoked with values like 0.51, the difference between 0.51 and 0.49 is a band boundary, not a gradient. Bands are coarse zones. For workflows reviewed by humans, prefer enum cursor names (depth: deep) over numeric values — they travel better through code review and diff.

Cross-node non-transitivity

Cursors on a judge do not propagate to the agent it judges. Each node sees only the cursors set on its own declaration. A “demanding reviewer” judge does not by itself make the upstream implementer agent more demanding — you’d activate the matching cursor on the implementer too.

The façade temptation

Setting rigor: loose to make a stuck judge pass is a misuse. The judge is signaling something. Investigate, fix the issue, or adjust the gate. Calibration is a framing dial; it is not authorization to ignore validation.

Reference catalogue

Four reference cursors live at examples/cursors/cursors.bot:

ambition — how aggressively to expand scope beyond the stated request
depth — investigation thoroughness, 0.0 (skim) → 1.0 (exhaustive)
rigor — verification standard, 0.0 (accept) → 1.0 (cite every claim)
autonomy — escalation threshold (gated → autonomous)

These are copy-paste-ready snippets, not a runnable workflow. examples/cursors/sample.bot is a minimal validatable workflow that exercises two of them.

Authoring guidance

Lead each fragment with an imperative verb (Trace, Surface, Stick). Avoid hedge words and “consider X” phrasing — the LLM reads “consider” as “ignore.”
Cursor names are lowercase, snake_case if multi-word. They become capitalized labels in the ## Calibration section.
Numeric cursors use string-keyed ranges ("0.0..0.33"). The lo and hi are inclusive on both ends; bands cannot overlap.
Cursor declarations are per-file. There is no shared catalogue imported across workflows. If you find yourself copying the same cursor into many .bot files, that is a sign the workflow should share a bundle, not that cursors should grow an import system.

This site is open source. Improve this page.