Complex Project Discipline

A universal operational discipline distilled from production-grade solo-engineer execution on a project that simultaneously ran (a) a research line with measured gates and preserved failures, (b) a production web service with tunneled deployment and corpus-driven verification, (c) a funding-readiness application cycle across the alignment-funder ecosystem, and (d) a layered AI-agent harness via Claude Code.

This skill is project-agnostic. Every rule has been formulated as a universal pattern. Adapt names to your project; keep the patterns.

When this skill applies

Load full skill body when ANY of:

Starting a new project and establishing operational rules from day 1.
Inheriting a project that lacks discipline (no canonical state file, no append-only logs, ad-hoc commits).
Solo or small-team execution where you need external structure to substitute for a larger team's review redundancy.
Funding application cycle (alignment funders, public posting, fellowship, paid pilot).
Public-surface UX work needing standards baseline (a11y, security, perf, i18n, brand).
Introducing AI-agent harness (Claude Code settings, hooks, skills, agents) to an existing repo.
Session continuity recovery after long break.
Verification gate before production deploy or external claim.

DO NOT auto-load on every casual conversation. Triggers should be specific (see frontmatter description).

Skill structure (progressive disclosure)

SKILL.md                                          (this file — entry + index)
docs/
  01_TRUTH_HIERARCHY_AND_STATE_FILES.md           Canonical state file pattern. 9-level hierarchy.
  02_BOUNDED_EXECUTION.md                         Task pack format + 10-step execution flow.
  03_VERIFICATION_DISCIPLINE.md                   "No artifact, no claim." Three verification states.
  04_BRANCH_AND_PUSH_DISCIPLINE.md                Main protection + explicit-override push pattern.
  05_APPEND_ONLY_LOGS.md                          Execution-log + weekly-status row format.
  06_CHANGE_CLASS_CLASSIFICATION.md               R0/R1/R2 classes + when-in-doubt rule.
  07_UX_STANDARDS_CANONICAL.md                    10-standard distilled stack + per-surface checklists.
  08_PRODUCTION_INFRA_DISCIPLINE.md               Deploy patterns + DB volume preservation + tunnel rules.
  09_CORPUS_DRIVEN_DEVELOPMENT.md                 7-step corpus pattern + thresholds.
  10_FUNDING_READINESS_DISCIPLINE.md              Owner-operated disclosure + reference qualification.
  11_COMMUNICATION_PATTERNS.md                    Apology + opt-in paths + customer-debt acknowledgment.
  12_CLAUDE_CODE_HARNESS_PATTERN.md               8-step layered harness rollout.
  13_MEMORY_STRUCTURE.md                          Repository + state + top-of-stack continuity.
  14_SESSION_CLOSE_DISCIPLINE.md                  Close gate + 6-condition final answer gate.
  15_ANTI_PATTERNS.md                             What NOT to do (catalogued anti-patterns with counter-patterns).
  16_POLISH_DISCIPLINE.md                         Critical re-read for adversarial review.
templates/
  task_pack.md                                    Reusable 9-section task pack template.
  bounded_packet.json                             Bounded execution packet JSON schema.
  agents_log_row.md                               Single-row format for append-only execution log.
  weekly_status_section.md                        Per-task ## section format.
  reference_prep_brief.md                         Integrity-aligned ground-truth context for referees.
  apology_email_pattern.md                        Customer-relationship debt closure pattern.
  paste_ready_field_map.md                        Form-submission paste-ready field map template.
  claude_settings.json                            Project-level Claude Code settings template.
checklists/
  pre_commit_check.md                             Before any git commit.
  pre_push_check.md                               Before any git push.
  session_close_check.md                          End-of-session full closure.
  final_answer_gate.md                            6-condition gate before declaring "done".
  ux_per_surface_check.md                         Per-public-surface UX checklist.
  deploy_verification_check.md                    Production deploy verification.

The seven core principles (read first)

1. Truth hierarchy is non-negotiable

Every project needs a single canonical live-state file. Lower levels of documentation can give context but cannot override higher levels. When no artifact backs a claim, mark it UNVERIFIED. See docs/01_TRUTH_HIERARCHY_AND_STATE_FILES.md.

2. Bounded execution beats heroic execution

One task = one reason. One commit = one logical change. Acceptance criteria mandatory. Rollback plan mandatory for state changes. Audit before edit, even on "obvious" tasks. See docs/02_BOUNDED_EXECUTION.md.

3. No artifact, no claim

PARTIAL is allowed. UNVERIFIED is required when proof is missing. SUCCESS requires (a) measured artifact, (b) live-state updated, (c) memory updated, (d) verifier returns the new state, (e) shell-valid verification commands. See docs/03_VERIFICATION_DISCIPLINE.md.

4. Append-only logs are the spine

Old entries are never edited. New entries always appended at end. Format is rigid (pipe-separated row + section header). The discipline of writing append-only forces honesty in retrospect. See docs/05_APPEND_ONLY_LOGS.md.

5. Disclosure-first integrity beats hidden conflicts

Hidden conflicts of interest are catastrophic when discovered. Disclosed conflicts are signals of integrity. This rule applies equally to research claims, funding references, owner-operated testbeds, and any relationship that could appear as a conflict of interest. See docs/10_FUNDING_READINESS_DISCIPLINE.md.

6. Preserved failures are a competitive advantage

When an experiment fails, document it as canonical FAIL — do not silently rewrite. The discipline of preserving failure is the cleanest empirical demonstration of where your method's boundary actually is. See docs/09_CORPUS_DRIVEN_DEVELOPMENT.md + docs/03_VERIFICATION_DISCIPLINE.md.

7. Main branch is sacred until explicit unlock

Working branch (e.g. test, dev, feature branches) is for active work. Main branch (or production) is protected and only updated via explicit user-issued unlock command (e.g. MERGE_MAIN_NOW). Mechanical enforcement via permissions.deny in tooling settings. No git push --force, no reset --hard, no rebase -i. See docs/04_BRANCH_AND_PUSH_DISCIPLINE.md.

Quick-start for a new project

If you are starting from scratch and want this discipline applied:

Create canonical state file — name it consistently (CANON.md, LIVE_STATE.md, project_truth.md). Single source of truth at the top of hierarchy. See docs/01.
Create append-only logs — execution journal (pipe-separated rows) and weekly status (per-task ## sections). See docs/05 + templates/agents_log_row.md.
Create memory structure — memory.md with Repository / Active Scope / Current State / Top-of-Stack sections. See docs/13.
Define operational rules — equivalent of .claude/CLAUDE.md / OPERATING_RULES.md / AGENTS.md. Sections: Language, Identification, Mandatory Startup, Source-of-Truth Hierarchy, Execution Rules, Forbidden Actions, Logging & Continuity, Class Classification, Task Completion Checklist, Session Close Checklist. See docs/14.
Set up AI-agent harness (if using Claude Code) — .claude/settings.json with allow/deny lists, SessionStart hook for canon load, layered rollout via 3 push checkpoints. See docs/12.
Adopt UX standards canonical (if public-surface project) — distill 10 international standards into hard rules + soft rules + measurable thresholds. See docs/07.
Establish bounded execution flow — task pack format for any non-trivial work, 10-step flow (audit → findings → implement → tests → verifier → commit → restart → live verify → brief → push). See docs/02.

Quick-start for a funding application cycle

If you are preparing funding applications for an existing project:

Audit funding pack — list all draft applications, deadlines, requirements. See docs/10.
Verify disclosures — owner-operated systems explicitly marked, family/partner relationships flagged, no hidden conflicts. See docs/10.
Reference qualification — for each candidate referee: independent? engaged with methodology? willing? See templates/reference_prep_brief.md.
Strategic skip when integrity at risk — refusing to apply with conflicted refs is itself an alignment-discipline signal. Document the skip openly in remaining apps. See docs/10.
Paste-ready field map before submit — survey the form, map content to fields, polish for adversarial review, char budget audit, submit only after explicit user OK on each consent. See templates/paste_ready_field_map.md + docs/16.
Append to submission log — date, fund, ask, status, decisions made. See docs/05.

Quick-start for production deploy

Pre-deploy check — clean git tree, tests pass, corpus pass, verifier known state. See checklists/pre_commit_check.md.
Deploy via documented pattern — never docker compose down (preserves DB volumes); always up -d --build --force-recreate on the app container only. See docs/08.
Health check after every restart — curl /v1/health or equivalent.
Live verify on a real URL — not just unit tests; real HTTP request to public surface. See docs/03.
Update infra state doc if state changed. See docs/08.
Append to execution log — record deploy with commit hash and verification commands.

Operational language rule (universal pattern)

For multilingual projects: pick one user-facing language and one machine-actionable language; never mix in artifacts that get parsed. ZIP file contents, CSV headers, JSON keys → always one canonical language (typically English). Email body, UI copy, brief reports to user → user-facing language. See docs/07_UX_STANDARDS_CANONICAL.md and docs/11_COMMUNICATION_PATTERNS.md.

What this skill does NOT replace

A real human team (this is solo / small-team discipline, not a substitute for review/redundancy).
Domain expertise (you still need to know your research / product / market field).
Legal / accounting / tax advice (the funding-readiness module assumes you'll get those separately).
A real product/market fit (discipline doesn't make a bad idea good).

Update rule

This skill is a snapshot of working discipline patterns. As your project diverges, fork this skill into a project-specific copy — do not modify the canonical version. Date your fork and document divergences explicitly.

This canonical version is intentionally project-agnostic and contains no internal data, partner names, or specific identifiers from any single project.

complex-project-discipline

SKILL.mdskills/complex-project-discipline/SKILL.md