# Lenses — operating guide > Companion to [Plasticity](/docs/plasticity.md). Raw Markdown: > [/docs/lenses.md](/docs/lenses.md) · rendered: [/docs/lenses](/docs/lenses) **In founder terms:** lenses are how you stop every agent from asking the same cold questions — tuned methodology you keep, export, and hand to a peer as markdown (not a second proprietary format). A **lens** is a typed prompt that shapes how Bloom samples and interrogates structural holes. Lenses are product configuration: system templates merge with tenant overrides, gate by dose tier, and freeze into the run at start. Tune lenses in the Lens Console ([app.jackiq.com/plasticity](https://app.jackiq.com/plasticity)) or via the plasticity HTTP/MCP surfaces. Edits write `lens_revision` rows; in-flight runs keep their frozen `lens_set`. ## The five categories | Category | Job | Typical hooks | | -- | -- | -- | | `selection` | Which candidates are drawn | structural-hole · orphan-rescue · cross-domain · temporal-drift · contradiction-hunt | | `interrogation` | What is asked of each pair | what-if-same-problem · what-would-have-to-be-true · second-order-effects · steelman-the-opposite | | `domain` | Business perspective | positioning · financial-model · gtm · product · ops · fundraise · risk · customer | | `discipline` | Guardrails | evidence-required · null-preferred · provenance-strict · charter-aware | | `judgment` | Verification | correctness · feasibility · reputational · financial · does-it-reproduce | Order in the console is selection → interrogation → domain → discipline → judgment. Unknown categories fall back under selection. **Judgment lenses must be genuinely distinct perspectives**, not N identical skeptics. Diversity catches failure modes redundancy cannot. **Domain lenses** should inherit the founder's methodology stages (the `/doctrine/*` shape) rather than invent a parallel framework. ## How to write a lens (enforced-output form) Write lenses in **enforced-output form**, not prose essays. Consolidation Run 02 showed prose doctrine produced no lift; enforced-output form transferred. ### Template ```text ROLE You are the lens (). Operate only on the pair and evidence provided. Do not invent store facts. TASK REQUIRED OUTPUT (emit exactly these fields, nothing else) - finding: - evidence: - blast_radius: - confidence: <0.0–1.0> - kill_if: CONSTRAINTS - Prefer null over a weak finding. - If evidence is missing, set finding to "null" and say so in evidence. - Do not rewrite charter, doctrine, or settled decisions — propose only. ``` ### Rules that stay true 1. **Name the fields.** Free-form paragraphs do not compose under Peak. 2. **Prefer null.** Discipline lenses exist so empty is a valid, valuable answer. 3. **Cite or abstain.** Provenance-strict lenses should refuse undated certainty. 4. **Stay inside category.** A selection lens picks candidates; it does not render a GTM essay. 5. **Set `min_tier` honestly.** Sharper interrogation lenses gate above Standard (`t3`). Locked lenses stay visible as "unlocks at …" — never hide the ceiling. ### Fields the system stores When creating or overriding a lens: - `slug` — stable id (kebab-case) - `name` — founder-facing label - `category` — one of the five above - `prompt` — enforced-output text - `weight` — 0–1 relative influence when assembling the run - `enabled` — soft toggle; disabled lenses are excluded from the freeze - `min_tier` — `t1`…`t5` (must be unlocked by current dose) Tenant overrides win on merge by `slug`. The system row remains the diff base ("Reset to system default" deletes the override). ## The system library Every entitled tenant starts from the shipped library (~25 lenses across the five categories). It is product IP — methodology in executable form. **Selection (ship with):** structural-hole, orphan-rescue, cross-domain, temporal-drift, contradiction-hunt **Interrogation:** what-if-same-problem, what-would-have-to-be-true, second-order-effects, steelman-the-opposite **Domain:** positioning, financial-model, gtm, product, ops, fundraise, risk, customer **Discipline:** evidence-required, null-preferred, provenance-strict, charter-aware **Judgment:** correctness, feasibility, reputational, financial, does-it-reproduce Exact prompts evolve; treat live `GET /plasticity/lenses` (or the console board) as source of truth for wording and weights. Ablation may cut lenses that show no measurable effect — do not assume every shipped slug is permanent. ## How tuning affects a run 1. **Before start** — weights, toggles, and prompts merge (system + tenant), filter by dose tier, normalize weights, and **freeze** into `operator_run.lens_set`. 2. **During the run** — further edits do not change the frozen set. The UI and run record say so explicitly. 3. **What `plan_run` shows** — lenses excluded by the chosen dose appear in the plan so the founder sees what their ceiling did not unlock. 4. **Return / history** — kills keep their reasons. That kill record is how you learn which lenses earned their weight. 5. **Overrides** — raise weight on `contradiction-hunt` when the intention is about conflict; enable a domain lens that matches the question; disable noisy lenses rather than rewriting them into mush. ### Practical tuning loop for agents 1. Read current lenses and the founder's intention. 2. Propose minimal overrides (weight / enable / prompt) tied to that intention. 3. `plan_run` and confirm the freeze preview. 4. After Return, summarize kill reasons → suggest the next weight change. 5. Never claim a mid-run prompt edit affected the active run. ## Related guides - [Plasticity](/docs/plasticity.md) — operators, phases, containment - [Dose](/docs/dose.md) — which tiers unlock which lenses - [Resource index](/llms.txt) ## Derivation (optional) Lens design and the enforced-output ruling live with the plasticity formal spine and ADR 0017 (operator registry / methodology as executable lenses). Search the Company Brain for `enforced-output` and `system lens library` when you need the eval trail, not just the operating rules.