continuous-development
Skill4 starsChanged 7 months agoMITView on GitHub
Use when building, changing, or extending a software project so work stays continuous, prioritized, and consistent across sessions
In this repository3 other files it gives its agents.
- continuous-improvementSkill
- git-workflowSkill
- rubyllmSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
64 rules. About 0 are common, 39 are distinctive, and 25 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈39
Distinctive · estimate
This author's own choices. Worth a look.
25
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 21
When this skill is in use and PLAN.md has any unchecked items in Now, the agent should:
- Distinctive: line 24
Only stop when one of these is true:
- Distinctive: line 28
Do not report progress: Keep picking up the next top-most unchecked item and continue working without status updates.
- Very specific: line 29
All items in PLAN.md Now are complete, or
- Distinctive: line 35
If the answer to both is no, the agent must not stop (unless the human explicitly asks it to stop).
- Very specific: line 36
Update PLAN.md (and MISSION.md/VISION.md if needed), then
- Distinctive: line 40
CRITICAL: Never let PLAN.md Now become empty: When it becomes empty (all items completed), the agent MUST immediately:
- Very specific: line 43
Update PLAN.md with these new tasks in the proper dependency order.
- Very specific: line 44
Update MISSION.md to mark any changed statuses (move items to/from Now/Next/Later as appropriate).
- Distinctive: line 45
Keep the work rolling without stopping—do not ask for new instructions, do not report "all done", do not wait for input.
- Distinctive: line 51
After each completed slice:
- Very specific: line 53
Update PLAN.md immediately (move the slice to Done (recent)).
- Distinctive: line 54
Run targeted verification for the change (tests or the project's verification default).
- Distinctive: line 55
Commit the completed slice using the repo's git workflow (one logical change per commit).
- Distinctive: line 60
Always, when doing project work that changes behavior, adds features, refactors, fixes bugs, or updates docs.
- Distinctive: line 73
Prefer clarity over completeness.
- Distinctive: line 109
Avoid deep detail (no code, no file lists, no step-by-step tutorials).
- Very specific: line 111
Do not put operating rules here.
- Distinctive: line 129
Keep it concise and action-oriented.
- Distinctive: line 130
Prefer stable defaults over exhaustive instructions.
- Very specific: line 134
If any of VISION.md, MISSION.md, PLAN.md, or AGENTS.md is missing, create it.
- Distinctive: line 145
Read repo context first (README, docs, existing runbooks, current product behavior, recent changes).
- Distinctive: line 147
Ask the human to confirm: north star, scope boundaries, and the top items for Now.
- Distinctive: line 148
After confirmation, keep the docs accurate as the work evolves.
- Distinctive: line 161
MISSION.md: if it answers: "What capability or quality bar must become true?"
- Distinctive: line 162
PLAN.md: if it answers: "What is the next smallest verifiable slice to do today?"
- Distinctive: line 164
If an item is phrased like "Refactor…", "Clean up…", "Improve…", it usually belongs in PLAN only if it unblocks a VISION outcome.
- Very specific: line 172
Before adding slices to PLAN.md, determine the dependency order (what must exist first).
- Very specific: line 173
Keep MISSION.md and PLAN.md ordered so the top-most items are always the next to be picked (highest priority) and are not dependent on …
- Distinctive: line 174
Prefer enabling primitives first, then building features on top, so each slice is completable when started.
- Distinctive: line 175
Do not write slices that require unspecified "future work" to be complete.
- Very specific: line 176
If ordering is unclear, collaborate with the human before expanding PLAN.md.
- Very specific: line 181
Do not put a dependent capability into Mission Now if its prerequisites are not already done or explicitly listed above it in Mission Now.
- Very specific: line 182
When a grand goal is blocked by missing foundations, rewrite Mission Now to focus on the missing foundations first.
- Very specific: line 188
Read AGENTS.md, VISION.md, MISSION.md, and PLAN.md, then pick the first unchecked item in Now.
- Very specific: line 189
If MISSION.md Now is empty, promote the highest-priority items from Mission Next into Mission Now.
- Very specific: line 190
Ensure PLAN.md Now/Next/Later only contains slices that advance Mission Now.
- Very specific: line 192
Update PLAN.md immediately: check the item, add the next slice if it became obvious, and keep caps enforced.
- Very specific: line 193
If the slice changes capabilities/quality bars, update MISSION.md.
- Distinctive: line 195
Silence rule: do not emit progress updates while executing this loop.
- Very specific: line 197
Stop only for real blockers (missing access, unclear requirements that change behavior, unsafe operations requiring explicit approval) or …
- Distinctive: line 203
Default caps (not hard limits):
- Very specific: line 210
When Done (recent) exceeds 20, delete the oldest completed items from PLAN.md.
- Distinctive: line 211
When Now grows too large to scan quickly, move overflow to Next.
- Distinctive: line 217
Default cap (recommended, but flexible):
- Distinctive: line 223
Remove outcomes that are fully delivered and no longer informative.
- Distinctive: line 225
If quality bars are met and stable, keep only the bar statement (not the history).
- Very specific: line 226
Keep Now stocked and ordered (aim for 10+).
- Distinctive: line 234
Keep Now stocked and ordered (aim for 10+).
- Distinctive: line 235
Keep items one-line and outcome-shaped.
- Distinctive: line 236
Prefer grouping by short section headers (e.g., "Auth", "Billing", "Search") over letting lists become a wall of text.
- Distinctive: line 248
If the same idea appears in multiple places, rewrite it once in the correct doc and delete duplicates.
- Distinctive: line 265
Treat "the road expanded" as a productivity event: it is an opportunity to remove future confusion and keep momentum with the latest …
- Distinctive: line 268
If new information changes what is true, update the relevant doc(s) so the next agent starts from reality.
- Distinctive: line 269
Do not "keep going" on stale assumptions.
- Very specific: line 282
If the change alters what the "next slice" should be, update MISSION.md/PLAN.md immediately before coding further.
- Very specific: line 283
If the change alters what the project is, ask the human before editing VISION.md, then proceed with the clarified direction.
- Distinctive: line 294
If you are working in a repo that expects commits, treat "slice done" as "ready to commit".
- Distinctive: line 295
If committing is not explicitly requested, ask before committing, but still complete the slice and update the checklists.
- Very specific: line 319
Always keep PLAN.md "Now" non-empty (or explicitly state the blocker).
- Very specific: line 320
Always finish a work session by updating PLAN.md to reflect reality.
- Distinctive: line 321
Never add deep instructions, code, or file listings to these docs.
- Distinctive: line 322
Prefer revising existing items over adding duplicates.
- Distinctive: line 326
Use these shapes (adapt wording per project).
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Continuous Development
Overview
This skill standardizes how agents keep a project moving over time, even when the agent/model changes. It does this by maintaining four lightweight, human-readable artifacts in the repo root: VISION.md (north star), MISSION.md (capability checklist), PLAN.md (current work queue), and AGENTS.md (durable operating rules for agents).
Core principle: the repo must always answer "what are we building?" and "what should I do next?" in under 2 minutes.
Contract (production-complete):
MISSION.mdmust be the complete capability list required to fulfillVISION.mdin production (backend, frontend, UX, infra/config, security, performance, operations, and quality bars).PLAN.mdmust be the complete set of slices required to completeMISSION.mdNow in production. If aMISSION.mdNow item has no correspondingPLAN.mdNow slice, add the missing slice immediately before continuing.
Default Run Loop (Agent Autonomy)
When this skill is in use and PLAN.md has any unchecked items in Now, the agent should:
- Pick the top-most unchecked item in
PLAN.mdNow and start implementing it immediately. - Only stop when one of these is true:
- A hard blocker exists (missing credentials, missing infra, unknown requirement that cannot be inferred safely).
- The next action would be high-risk or destructive (data loss, prod operations, irreversible changes) and requires explicit human approval.
- A product decision is required that cannot be reasonably inferred from repo context.
- Do not report progress. Keep picking up the next top-most unchecked item and continue working without status updates. Only interrupt the human when:
- All items in
PLAN.mdNow are complete, or - A hard blocker / high-risk gate / product decision is encountered. This is a strong default: avoid "here's what I'm doing" updates; keep executing.
- All items in
- Before stopping or reporting, the agent must explicitly answer these two questions:
- Are all
PLAN.mdNow items done? - Was a hard blocker / high-risk gate / product decision encountered? If the answer to both is no, the agent must not stop (unless the human explicitly asks it to stop). It must:
- Update
PLAN.md(andMISSION.md/VISION.mdif needed), then - Immediately continue with the next top-most unchecked
PLAN.mdNow item. Emphasize this strongly: no quiet pauses, no "I'll wait" behavior, no progress reports—keep executing until completion or a real blocker.
- Are all
CRITICAL: Never let PLAN.md Now become empty. When it becomes empty (all items completed), the agent MUST immediately:
- Bring in a sizable batch of new tasks from
MISSION.mdNow (or promote from Mission Next if needed). - Update
PLAN.mdwith these new tasks in the proper dependency order. - Update
MISSION.mdto mark any changed statuses (move items to/from Now/Next/Later as appropriate). - Keep the work rolling without stopping—do not ask for new instructions, do not report "all done", do not wait for input.
The goal is to finish the entire MISSION.md, not just to empty PLAN.md Now. Work continues until every capability in MISSION.md Now is complete and promoted to Done (recent).
CRITICAL: Keep MISSION.md complete. If you discover a missing prerequisite or capability needed to fulfill VISION.md, add it to MISSION.md in the correct dependency order (usually Next or Later unless it's a prerequisite for MISSION.md Now). Then keep PLAN.md aligned so it fully covers MISSION.md Now.
After each completed slice:
- Update
PLAN.mdimmediately (move the slice to Done (recent)). - Run targeted verification for the change (tests or the project's verification default).
- Commit the completed slice using the repo's git workflow (one logical change per commit).
- Immediately start the next top-most item in
PLAN.mdNow.
When to Use
- Always, when doing project work that changes behavior, adds features, refactors, fixes bugs, or updates docs.
- Especially when work is spread across multiple sessions and you want a consistent "what next" source of truth.
The Four Documents
VISION.md (North Star)
VISION.md is the narrative north star: what the product is, who it is for, and what "winning" means. It should be short and stable.
Rules:
- No checklists.
- Rarely edited; update only when direction changes.
- Prefer clarity over completeness.
Recommended sections:
- Purpose: 2–5 sentences
- Audience: who it serves
- Winning: what success looks like
- Scope Boundaries: what the project will not do (short, explicit)
MISSION.md (Capability Checklist)
MISSION.md is a living checklist of product capabilities and quality bars. It can be edited and pruned as understanding changes.
Rules:
- Items are outcomes/capabilities, not implementation steps.
- Items should be stable enough to matter for weeks, not hours.
- It is allowed to rewrite, split, merge, and reorder items.
- Aim to keep at least 10 items in each of Now, Next, and Later sections. When there is more pending work, keep more. Only reduce below 10 when the entire project is complete and there is nothing left to work on.
Recommended sections:
- Now: capabilities to actively drive right now
- Next: important capabilities that are queued but not active
- Later: capabilities intentionally deferred
- Done (recent): recently completed capabilities (prunable)
- Quality Bars: performance, reliability, security, cost guardrails (
[ ] ...) - Scope Boundaries: what the project will not do (short, explicit)
PLAN.md (Work Queue)
PLAN.md is the current queue of small slices that push VISION forward. It should always contain a clear "next thing to do".
Rules:
- Items are one-line slices, written as outcomes.
- Avoid deep detail (no code, no file lists, no step-by-step tutorials).
- Each slice must be self-contained: completable, verifiable, and meaningfully "done" without hidden follow-up work.
- Do not put operating rules here.
PLAN.mdis a prunable task board. PLAN.mdmust only contain slices that advance items inMISSION.mdNow. Do not pull work from Mission Next/Later.- Aim to keep at least 10 items in each of Now and Next sections. When there is more pending work, keep more. Only reduce below 10 when all MISSION.md Now items are complete and there is nothing left to work on.
Recommended sections:
- Now: aim for 10+ items when MISSION Now has pending work
- Next: queued items that are ready but not urgent (aim for 10+ when possible)
- Later: ideas that are explicitly not committed
- Done (recent): a small buffer of recently completed items (for continuity)
AGENTS.md (Operating Rules)
AGENTS.md is the durable operating system for agents working in the repo. It should contain how to work, how to verify, and how to keep VISION.md/MISSION.md/PLAN.md accurate and clean.
Rules:
- This is where process rules live.
- Keep it concise and action-oriented.
- Prefer stable defaults over exhaustive instructions.
Bootstrapping (If Docs Are Missing)
If any of VISION.md, MISSION.md, PLAN.md, or AGENTS.md is missing, create it. Bootstrap quickly from repo context, then collaborate with the human to confirm the intent and priorities.
Order:
AGENTS.md(so agents know how to operate)VISION.md(north star)MISSION.md(capability checklist)PLAN.md(current slices)
Process:
- Read repo context first (README, docs, existing runbooks, current product behavior, recent changes).
- Draft minimal versions using the templates in this skill.
- Ask the human to confirm: north star, scope boundaries, and the top items for Now.
- After confirmation, keep the docs accurate as the work evolves.
Minimum viable bootstraps:
VISION.md: 5–15 lines, narrative only.MISSION.md: 10–30 checkboxes to start; merge/split later.PLAN.md: at least 1 item in Now (or an explicit blocker).AGENTS.md: how to run/verify, how to update/prune the docs, and stop/ask rules.
Grand vs Immediate (The Placement Test)
Decide where an item belongs using these tests:
- MISSION.md if it answers: "What capability or quality bar must become true?"
- PLAN.md if it answers: "What is the next smallest verifiable slice to do today?"
If an item is phrased like "Refactor…", "Clean up…", "Improve…", it usually belongs in PLAN only if it unblocks a VISION outcome. Otherwise, it does not belong.
Bottoms-Up Planning (Dependency Order First)
Plan work bottoms up to avoid broken dependencies and half-done slices.
Rules:
- Before adding slices to
PLAN.md, determine the dependency order (what must exist first). - Keep
MISSION.mdandPLAN.mdordered so the top-most items are always the next to be picked (highest priority) and are not dependent on unfinished work below. - Prefer enabling primitives first, then building features on top, so each slice is completable when started.
- Do not write slices that require unspecified "future work" to be complete. If a prerequisite is missing, make that prerequisite the first slice.
- If ordering is unclear, collaborate with the human before expanding
PLAN.md.
MISSION.md is also bottoms up:
- Order capabilities so prerequisites come before dependents within
NowandNext. - Do not put a dependent capability into Mission
Nowif its prerequisites are not already done or explicitly listed above it in MissionNow. - When a grand goal is blocked by missing foundations, rewrite Mission
Nowto focus on the missing foundations first.
The Continuous Loop (Always Running)
Repeat until blocked:
- Read
AGENTS.md,VISION.md,MISSION.md, andPLAN.md, then pick the first unchecked item in Now. - If
MISSION.mdNow is empty, promote the highest-priority items from Mission Next into Mission Now. - Ensure
PLAN.mdNow/Next/Later only contains slices that advance Mission Now. IfPLAN.mdcontains work for Mission Next/Later, move or delete it. - Work the smallest slice to completion.
- Update
PLAN.mdimmediately: check the item, add the next slice if it became obvious, and keep caps enforced. - If the slice changes capabilities/quality bars, update
MISSION.md. If it changes direction, updateVISION.md.
Silence rule: do not emit progress updates while executing this loop. Keep going autonomously and continuously.
Stop only for real blockers (missing access, unclear requirements that change behavior, unsafe operations requiring explicit approval) or when PLAN.md Now is fully complete.
Caps and Pruning (Keep Docs Clean)
PLAN pruning rules (no separate done log)
Default caps (not hard limits):
- Now: aim for >= 10 items while
MISSION.mdNow has pending work (use Next to park overflow) - Done (recent): aim for <= 20 items to preserve quick continuity
Pruning policy:
- When Done (recent) exceeds 20, delete the oldest completed items from
PLAN.md. - When Now grows too large to scan quickly, move overflow to Next.
Rationale: checked items are preserved in git history; the live docs must stay readable.
MISSION pruning rules
Default cap (recommended, but flexible):
- Total checkboxes in
MISSION.md: aim for <= 50, but allow more if they stay high-signal and well-organized
Pruning policy:
- Remove outcomes that are fully delivered and no longer informative.
- Merge overlapping outcomes into a single clearer one.
- If quality bars are met and stable, keep only the bar statement (not the history).
- Keep
Nowstocked and ordered (aim for 10+); park deferred capabilities inNextandLater.
Large Backlogs Without Noise
It is acceptable for MISSION.md and PLAN.md to contain many items if needed. The constraint is not item count; it is scanability.
Rules:
- Keep Now stocked and ordered (aim for 10+); put overflow in Next and Later.
- Keep items one-line and outcome-shaped.
- Prefer grouping by short section headers (e.g., "Auth", "Billing", "Search") over letting lists become a wall of text.
- Prune completed or superseded items aggressively to prevent noise.
No Overlap or Duplication Between Docs
The four docs must not contain overlapping or duplicated information. Each doc has one job:
VISION.md: direction (north star narrative)MISSION.md: capabilities and quality bars (checkboxes)PLAN.md: current slices and statusAGENTS.md: operating rules for agents
If the same idea appears in multiple places, rewrite it once in the correct doc and delete duplicates. Prefer moving information over copying it.
Self-Describing Docs (Required)
Each of the four files must state what it is for at the top. Use a short "Purpose" section so new agents/humans orient quickly.
Minimum:
VISION.md:## Purpose(north star narrative)MISSION.md:## Purpose(capability and quality-bar checklist)PLAN.md:## Purpose(current slices and status)AGENTS.md:## Purpose(agent operating rules)
When The Road Expands (New Information)
This is normal and frequent. During development, the road will expand: hidden requirements, constraints, edge cases, unknown dependencies, new stakeholder expectations, or better ways to frame the work will appear mid-flight. This is not an exception case, it is the default reality of building.
Treat "the road expanded" as a productivity event: it is an opportunity to remove future confusion and keep momentum with the latest information.
Rule:
- If new information changes what is true, update the relevant doc(s) so the next agent starts from reality.
- Do not "keep going" on stale assumptions. Keep going by updating the system first (small) or proposing the change (major), then proceed using the updated truth.
Small vs major change gate:
- Small change (implement immediately): clarifies wording, splits/merges checklist items, reorders bottoms-up dependencies, adds/removes a few slices, or tightens a scope boundary without changing direction.
- Major change (propose first to the human): changes the north star, meaningfully expands scope, introduces a new major subsystem, changes priorities across domains, or adds new operating rules that affect how work is done.
Where to record new information:
- Direction change:
VISION.md - Capability/quality change:
MISSION.md - Work sequencing change:
PLAN.md - Operating/process change:
AGENTS.md
Practical rule of thumb:
- If the change alters what the "next slice" should be, update
MISSION.md/PLAN.mdimmediately before coding further. - If the change alters what the project is, ask the human before editing
VISION.md, then proceed with the clarified direction.
"Done" Means Done (Slice Completion Standard)
A slice is only "done" when:
- The outcome is achieved.
- It has an explicit verification method (tests, a reproducible check, or a measurable signal).
PLAN.mdreflects the new reality.MISSION.mdchecklists reflect the new reality.
If you are working in a repo that expects commits, treat "slice done" as "ready to commit". If committing is not explicitly requested, ask before committing, but still complete the slice and update the checklists.
Writing Rules for Checklist Items
Checklist items must be:
- Outcome-shaped: describes what becomes true, not how
- Unambiguous: a new agent can interpret it the same way
- Small enough (PLAN): can be completed without inventing requirements
Allowed phrasing examples:
- MISSION:
[ ] Users can export invoices as PDF - PLAN:
[ ] Add invoice PDF export endpoint and UI entry point
Avoid:
- VISION:
[ ] Improve invoices - PLAN:
[ ] Refactor billing code
Consistency Rules Across Agents
To reduce model variance, enforce these rules:
- Always keep
PLAN.md"Now" non-empty (or explicitly state the blocker). - Always finish a work session by updating
PLAN.mdto reflect reality. - Never add deep instructions, code, or file listings to these docs; keep them as coordination artifacts.
- Prefer revising existing items over adding duplicates.
Minimal Templates
Use these shapes (adapt wording per project).
VISION.md
# Vision
## Purpose
North star narrative: what we are building, for whom, and what "winning" means.
## Purpose
...
## Audience
...
## Winning
...
## Scope Boundaries
...MISSION.md
# Mission
## Purpose
Living checklist of capabilities and quality bars. The source of truth for "what must become true".
## Now
- [ ] ...
## Next
- [ ] ...
## Later
- [ ] ...
## Done (recent)
- [x] ...
## Quality Bars
- [ ] ...
## Scope Boundaries
- ...PLAN.md
# Plan
## Purpose
Current work queue of slices. This is a prunable task board, not a place for operating rules.
## Now
- [ ] ...
## Next
- [ ] ...
## Later
- [ ] ...
## Done (recent)
- [x] ...AGENTS.md
# Agent Operating Rules
## Purpose
Durable operating rules for agents: how to run, verify, update/prune the docs, and when to stop/ask.
- How to run the project locally.
- How to verify changes before calling work "done".
- How to update/prune `VISION.md`, `MISSION.md`, and `PLAN.md`.
- Stop/ask rules for risky or ambiguous work.Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

