The Holak Scale lists five anti-patterns in a single section. After dozens of conversations with teams we see it’s not enough - and that the “one line + bullet” format doesn’t help anyone get out of them.

This article expands the list to seven, adds a recognition test per pattern, describes who gets stuck most often, and shows a concrete way out.

We wrote it together because each of us sees a different side of these stories. Konrad - from a delivery-team perspective at Sii. Grzegorz - from an AI-ambassador perspective at banks and fintechs.

1. “I use AI daily” - stuck at 1

What it is: daily questions to ChatGPT or Claude, but always in “ask - receive an answer” mode. Nothing more. No iteration, no context, no settings.

Test: “What’s new in your AI workflow from the last month?” - if the answer is “nothing, but I use it daily” → stuck at 1.

Who gets stuck: most often managers who decided adoption = usage. Also engineers who tried a few models and picked one.

Way out:

  1. Note 5 recurring question types from the last week.
  2. Build one prompt template per type - with role, context, format.
  3. For a week, use only the templates. After a week you’re at 2-3.

2. Prompt fetishism - stuck at 3

What it is: a library of “golden prompts” in Confluence or Notion. Each is 100-300 lines. You copy from the doc, substitute variables, send.

Test: “Show me the prompt you use daily. Where do you copy it from?” - if from a document, if it’s long, if you’re proud of the structure → prompt fetishism.

Who gets stuck: teams that started with a prompt-engineering workshop and never moved past it. Also individual engineers who built their own “golden set” solo.

Way out:

  1. Pick the prompt you use most often.
  2. Extract the content that doesn’t change (role, style, constraints) - move it to custom instructions.
  3. Extract project-specific knowledge - move it to AGENTS.md / CLAUDE.md.
  4. Leave only the task goal in the prompt. It should fit in 3-5 lines. You’re at 4+.

3. Documentation graveyard - stuck at 5

What it is: the project has AGENTS.md or CLAUDE.md - written once, six months ago. Nobody remembers what’s in there. The agent still works, but “more from memory than from the file.”

Test: “Open CLAUDE.md and tell me what changed in the last month.” - silence → graveyard. Edit once a month or after any significant project change.

Who gets stuck: teams that introduced context files enthusiastically but never put them in retros / weekly reviews. Also seniors who wrote the file and moved on to other topics.

Way out:

  1. Add a retro question: “does CLAUDE.md still match how we actually work?”
  2. Every meaningful workflow change = PR to the context file.
  3. Every quarter, review with the agent: “list things in CLAUDE.md that don’t match the current code.”

4. Skill bloat - stuck at 7

What it is: dozens of skills, dedicated agents, per-task config files. Nobody remembers what each does. The agent rarely picks the right one.

Test: “Name 5 skills you use weekly and describe how they differ.” - no fluent answer → bloat.

Who gets stuck: teams enamoured with “we have our own skills.” Also team leads who wanted to demonstrate internal output.

Way out:

  1. Audit: for each skill, count invocations over the last 30 days.
  2. Skills with <5 invocations - delete or merge.
  3. Skill with no owner - delete.
  4. Skill with overlapping functionality - consolidate.

Target: <8 skills, each described in 1 sentence, each with an owner.

5. False maturity 8 - faking 8 from 6-7

What it is: the company boasts: “we have MCP, we have connectors, we have 15 integrations.” But every agent action requires a manual click to approve. A human is still leading step by step.

Test: “Show me one agent action from the last week that ran in a real system without your click.” - no such example → false 8.

Who gets stuck: companies with strict security/compliance who added MCP without a permissions framework. Also teams chasing hype - they want to say “we have MCP” in slides.

Way out:

  1. Pick one low-risk action (e.g., creating a Jira ticket).
  2. Build the verifiability + reversibility framework.
  3. Enable autonomous execution for that one action.
  4. Monitor for a week. Scale to the next action.

6. The 4→9 jump - an agent with no context

What it is: “We’ll install an agent and let it run.” No context files, no skills, no MCP. The agent hallucinates, the team loses trust, drops to 2 with the “AI doesn’t work” motto.

Test: “How does the agent know your repo’s conventions?” - if the answer is “it figures it out” → 4→9.

Who gets stuck: teams under deadline pressure. Also managers who saw a demo at a conference and bought the tool the same day.

Way out:

  1. Step back to level 5. This isn’t failure - it’s discipline.
  2. Write CLAUDE.md / AGENTS.md in one weekend.
  3. Give the agent 5 tasks in a sandbox and compare with manual.
  4. Only then enable autonomy on one chosen task.

7. Orchestrating mediocrity - faking 10 without 9

What it is: three or five agents collaborate. None of them handles a single task end-to-end. Result: three times the errors, three times the cost, same outcome as with one agent.

Test: “Does a single agent handle 80% of tasks in this domain?” - if no → don’t build a team, fix the individual.

Who gets stuck: teams inspired by multi-agent frameworks. Also companies that want to write “we have orchestration” in their deck.

Way out:

  1. Stop the multi-agent project.
  2. Pick the weakest agent on the team.
  3. Make it succeed at 80% of its domain as a solo unit.
  4. Repeat for each.
  5. Only then return to orchestration - and check whether it’s still needed.

What they have in common

All seven share the same root: confusing activity with maturity.

  • “I use it daily” ≠ I’m advanced
  • “I have a prompt library” ≠ I’m at 4
  • “I wrote AGENTS.md” ≠ I work with context
  • “I have 15 skills” ≠ I have a strategy
  • “I have MCP” ≠ I have autonomy
  • “I have an agent” ≠ I delegated a goal
  • “I have many agents” ≠ I orchestrate

A test that catches them all: “show me a concrete output from the last week and tell me how much time it saved vs the old workflow.”

No answer = one of the patterns.

What’s next

In the Holak Scale, anti-patterns were a starting point. Here they have exits. Version 3 of the scale will collect these into a single table linked to checklists (we’re still gathering material - if you see one of them in your team, get in touch).

In practice: the biggest gain comes from admitting you’re in one of the patterns. The rest is discipline.