5.2 KiB
00 — Product vision
Product thesis
Software-development agents are capable of substantial work, but results still depend heavily on how a task is framed. Users repeatedly spend time restating repository rules, scope limits, validation commands, safety conditions and final-report expectations. Generic prompt libraries solve only the wording problem. They do not solve task contracting, repository adaptation, repeatability, quality evidence or organizational governance.
DevRunbook turns a development intention into a structured execution contract.
From intent to verified change.
The platform combines a versioned playbook, repository profile, user inputs, autonomy selection, risk controls and validation policy. It then renders a deterministic prompt or Run Pack that can be copied into Codex or, in later phases, handed to a controlled Codex integration.
Product category
DevRunbook should be positioned as an agentic development playbook platform, not as:
- a prompt marketplace;
- a chat interface;
- an IDE replacement;
- a remote code-execution service;
- a generic documentation wiki.
Its closest mental models are an operational runbook system, a policy-aware task composer and a quality registry for reusable agent workflows.
Primary value proposition
For an individual developer or technical operator:
- stop rewriting the same instructions;
- obtain a complete task with explicit done-when criteria;
- adapt proven procedures to each repository;
- choose how autonomous Codex may be;
- export a prompt, AGENTS.md suggestion or structured Run Pack;
- retain a history of exactly what was generated.
For a team:
- encode reviewable engineering standards;
- publish approved playbook versions;
- require safety and validation blocks;
- share repository profiles and command conventions;
- evaluate playbooks against fixtures;
- prove which version and context produced a result.
North-star experience
A user types:
“Clean up this TypeScript monorepo without changing behavior.”
DevRunbook recognizes the likely task, proposes the Repository Hygiene and Dead-Code Cleanup playbook, loads the selected repository profile, detects that the project uses pnpm, Turborepo and Vitest, highlights protected directories, lets the user choose Verify autonomy, and generates a task containing:
- pre-change inventory;
- explicit behavioral-preservation constraints;
- dependency and dead-code workflow;
- package-specific validation commands;
- failure-recovery rules;
- a completion contract;
- a structured final report.
The user can inspect where every generated section came from and export it as a multi-file Run Pack.
Product principles
1. Structured before clever
The platform should prefer a clear task contract over magical prompt rewriting. AI-assisted suggestions may help select or populate a playbook, but the final output remains inspectable and deterministic.
2. Verification is not optional decoration
Implementation playbooks must define how success is demonstrated. “Make it work” is not an acceptance criterion.
3. Context has provenance
Every repository fact included in a generated task records whether it was manually entered, imported from a profile, observed through Gitea or inferred. Imported text is clearly delimited as untrusted evidence.
4. Safe autonomy rather than false control
The user chooses an autonomy level with concrete permissions and behavior, not a vague “agent mode” switch.
5. Quality over catalog size
A smaller set of reviewed and evaluated playbooks is preferable to thousands of near-duplicate community prompts.
6. Local-first trust
The reference product is self-hostable. A user can operate the library, profiles, composer and exports without sending repository content to an additional SaaS service.
7. Honest capability boundaries
The product must distinguish static guidance, imported evidence, actual evaluations and user feedback. It must not label a playbook “verified” merely because its text looks comprehensive.
Success metrics
Initial product metrics:
- median time from task selection to export;
- percentage of generated tasks passing prompt lint without manual correction;
- number of repeated uses per playbook;
- number of follow-up prompts required after a generated task;
- user-reported scope adherence;
- percentage of generated tasks with complete validation evidence;
- repository-profile reuse rate;
- seed-catalog import and render success rate.
Longer-term quality metrics:
- fixture success by playbook version;
- regression rate between playbook versions;
- percentage of changes limited to declared scope;
- false-positive and false-negative rates for audit playbooks;
- rate of blocked unsafe exports;
- reproducibility of generated output from stored snapshots.
Non-goals for the MVP
- executing arbitrary repository commands;
- cloning untrusted repositories into a privileged host;
- acting as a full Git forge;
- replacing code review;
- guaranteeing security or compliance certification;
- automatically publishing community content;
- supporting every coding agent from day one;
- introducing a vector database without demonstrated search need;
- billing, public marketplace or complex SaaS tenancy.