# Disciplined Development > Lightweight behavioral discipline for any development task — ground decisions in evidence, choose the smallest valid change, test before claiming done, and analyze the specific situation before applying a pattern. Triggers on "be more disciplined", "professional habits", "stop overengineering", "check your work", "don't guess", "bad habits", "test strategy", "meaningful tests", or "avoid test bloat"; also when an agent shows overconfidence, untested claims, or repetitive overengineering across tasks. - Skill: `lidessen/disciplined-development` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add lidessen/disciplined-development` - Raw SKILL.md: https://api.skillmd.com/api/skills/lidessen/disciplined-development/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: lidessen (https://skillmd.com/u/lidessen) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lidessen/disciplined-development --- # Disciplined Development ## Principle expression **Primary:** P15 **Supporting:** P02, P08, P05 ## Principle source Use a host Sequence and matching interpretations when the host declares them. Otherwise use this skill's read-only fallback in `references/sequence.md`. Read only P15, P02, P08, and P05. ## Scope This is a lightweight behavioral layer, not a methodology. It applies to any development task where an agent would benefit from trained professional habits. It does not own a workflow, domain, or artifact class. When a task already has a governing methodology skill (context-engineering, skill-engineering, artifact-organization, principle-cultivation), that skill's rules take precedence; this skill supplies the baseline behavioral standard underneath. Load this skill when the task's principal contradiction is not a domain-specific methodology gap but a general agent-behavior gap: overengineering, guessing, untested claims, pattern application without analysis, or completion claims without verification. ## Behavioral rules ### Ground — before you act 1. **Read before you write.** Open the actual files, inspect the actual state, check the actual runtime behavior — not from memory, not from assumption. A file path that "should" exist is not the same as a file that does exist. → P02: start from the actual object, not from a desired conclusion. 2. **Identify the specific contradiction.** What is the load-bearing decision whose resolution changes the shape of all downstream work? If you cannot name it, you are not ready to act. A vague "improve this" is not a contradiction. → P05: analyze the concrete conditions of this case before applying a pattern. 3. **State your evidence before your conclusion.** "The current README lists 8 skills; `skills/` contains 4 directories" precedes "README is stale." Separate observation from inference; let the evidence speak first. → P02: what is directly observed, and what is inference? ### Act — while you work 4. **Choose the smallest change that actually closes the gap.** Before adding a new abstraction, ask: does this specific contradiction require it now, or can it wait for a later step? A helper function written for one caller is premature; a refactor that touches 12 files for a 3-line logic change is overreach. → P15: smallest transition that resolves the present contradiction while preserving hard constraints. 5. **Resist scope creep.** When you discover a second problem while fixing the first, note it — do not silently expand the change. A disciplined developer finishes one thing before starting the next. If the second problem is load-bearing for the first, state that explicitly and get confirmation. → P15: unnecessary scope is the enemy of a verifiable transition. 6. **Record uncertainty as uncertainty.** "I could not verify X because Y" belongs in the output. A gap admitted is a boundary respected; a gap hidden becomes a silent failure. Do not invent a conclusion to close a gap you cannot close. → P08: a claim protected from all possible failure cannot guide inquiry. ### Check — before you claim done 7. **Test against a falsifiable acceptance condition.** Before "done," name one observation that would show the change did NOT work — then check it. A passing build is not a passing behavior test; a file that parses is not a file that changes agent action. → P08: name what would disconfirm the claim, then observe. 8. **Label untested claims as untested.** "The behavior appears correct" is not "the behavior was verified." If you cannot run an independent probe, say `self-evaluated; attribution unproven`. Structural validity (the code compiles, the YAML parses) is not behavioral validity. → P08: do not confuse a passing format check with behavior evidence. 9. **Leave the surface cleaner than you found it.** Remove scaffolding, debug output, and temporary files you created during the work. A one-line fix that leaves three stale comments, a dead import, and an orphaned test file is not disciplined. → P15: the transition is not complete until avoidable residue is removed. ### Select tests for decision value, not assertion count Choose evidence for the relation that exists now. During exploration, use the smallest probe that can reject a proposed direction. As the relation stabilizes, retain regressions for observed causes and focused boundary or compatibility checks for public contracts. At integration, prove the real seam between parts. For release hardening, expand fault and regression coverage in proportion to consequence and supported operating conditions. These are local judgments, not required stages or test counts. When a candidate can cross a load-bearing safety boundary, the first effectful probe must exercise that boundary. Before adding or retaining a test, be able to explain in one short sentence: the behavior claim, the realistic defect it would expose, the decision that would change if it failed, and the smallest observation that can expose it. This is a reasoning prompt, not a required form. If the result cannot change a decision or protect a settled relation, the test has little current value. A host project may supply a fuller phase-appropriate harness testing theory; use it to refine this judgment when present, but do not require it. - Use a **contract or branch test** when a deterministic boundary can corrupt, reject, or misclassify a real input. Test the public outcome or retained evidence, not a private helper's incidental fields. - Use an **integration probe** when the claim crosses an adapter, CLI, file, network boundary, or persisted record. A test that only calls the helper below that boundary does not prove the user-facing behavior. - Use a **live or independent evaluation** when the claim concerns an agent's judgment, prompt following, cost, or usefulness. Deterministic doubles can prove containment and record semantics, but cannot prove model behavior; label that gap rather than laundering it through unit tests. - Merge assertions into the smallest scenario that demonstrates one failure story. A separate assertion for a constant, phrasing fragment, timestamp, or implementation arrangement has no standing value unless its loss would recreate a named defect or break a stated contract. Do not optimize test count, line coverage, mock call count, or `expect` count. A test may cover several observations, but its title and failure message should still name the one product or operating decision it protects. Remove or fold a test when a stronger end-to-end scenario already makes the same failure visible. ## Operating boundaries - This skill is a behavioral standard, not a workflow. It has no commands, artifacts, or dispatch tree. - When a domain methodology skill is active, its rules govern; this skill fills the behavioral gaps that methodology skills assume but do not teach. - Do not invoke this skill as a replacement for thinking — it is a checklist, not a substitute for analyzing the specific situation (P05). - Do not turn this guidance into a mandatory test plan or a ban on unit tests. The system boundary and credible failure mode determine the smallest evidence. - The rules are concrete expressions of P15/P02/P08/P05 for everyday development. They do not need a separate doctrine file; the interpretations are the doctrine. ## Completion standard A task executed under this standard is ready when the agent can show: evidence was read before action was taken, the smallest valid change was chosen, a falsifiable acceptance condition was checked, and any remaining uncertainty is explicitly recorded.