Audience: evaluators (lead engineer, PM, founder) deciding whether to adopt Mycelium for a project or team. Time to read: 5 min, or 1h to actually do the evaluation. Last updated: 2026-05-08.
This page is anti-promotional by design. The cost of a bad fit is high (a quarter of frustration with ceremony that does not match your work); the cost of saying so up front is low.
Mycelium imposes structure before the agent is allowed to start coding. That structure is theory-grounded discipline — 30+ frameworks, theory gates that block progression on insufficient evidence, a canvas of YAML files that tracks every assumption. The structure costs you something:
- Upfront opinionation tax.
/interviewis 15–60 minutes (depending on Phase 0 path) before the first useful output. Sprint mode and inline mode reduce but do not eliminate this. - Vocabulary cost. Diamond, scale, leaf, gate, OST, JTBD — a learning curve before fluency. The glossary compresses it; the curve is real.
- Conformance cost. The framework is opinionated about format (canvas YAML schema, decision-log structure, correction shape). If your team has its own discovery vocabulary, there is friction at the seams.
What you get for the tax:
- Wrong-build risk drops. Two of three Mycelium dogfood projects shipped; one was killed in L0 with evidence. The kill produced 10 framework features. See macos-fileviewer case.
- Onboarding cost on inherited projects drops. The canvas IS the spec. A new team member reads the canvas, runs
/diamond-assess, and has product context without a meeting. - Decision audit improves. Every theory gate fires with a citation. The decision log is contrastive (per-alternative
why_not), so a future reader can see which paths were rejected and why. - Agent drift gets caught. The reflexion loop, the pre-task protocol, the gates — these catch the agent jumping ahead. The discipline is mostly invisible until the moment the agent would have committed to the wrong scope.
The framework is most valuable when deciding what to build is the hard part. If your decision is settled and you only need execution velocity, this is the wrong tool.
See README "Who it's for". Tl;dr: solo developers and small teams using AI agents on real products where the wrong-build cost is non-trivial.
See README "Who it's not for". Tl;dr: triage-lane work, pure execution, projects with no real wrong-build risk. Listen to the ceremony-feels-heavier-than-value signal — it is a fit signal.
A focused hour beats a half-engaged afternoon. Five 10-minute slices, in order:
Open docs/receipts/. Read three case files: macos-fileviewer, drew-hoskins-takehome, framework-self-correction. The case files are the framework's evidence — if the receipts do not match the kind of friction your team hits, the framework probably is not for you.
Look for: are the corrections concrete (named commits, named bugs, named people) or vague? If concrete, the framework's claim that "Mycelium gets smarter with each cycle" stands on something. If vague, push back.
Clone Mycelium into a throwaway directory. Pick a real project you've been chewing on (a side project, a decision you're about to make at work, a thing you almost started). Run /interview Phase 0 → sprint mode (the 15-min path). Watch what the agent does.
Look for: does the agent ask questions that surface things you had not considered? Or does it feel like form-filling? The first is the framework working. The second is the framework over-fitting.
3. Read philosophy.md (10 min)
The why-opinionated rationale. If you disagree with the load-bearing claims (in-loop preventive > post-run evaluative; theory-grounded > ad-hoc; dogfood is required), the rest of the framework will feel like ceremony. If you nod along, the rest will feel like infrastructure.
Pick the two or three skills from docs/skills/ that match the gates relevant to your kind of work. If you build customer-facing products: /threat-model, /privacy-check, /a11y-check. If you build AI-containing products: add /xai-check, /regulatory-review. If you ship to enterprise: add /dora-check.
Look for: do the SKILL.md files give you a recipe you would actually run, or marketing prose? Mycelium's bias is recipes.
Open .claude/canvas/jobs-to-be-done.yml (the framework's own JTBD canvas, Mycelium dogfooded on Mycelium). Read it as a product spec. The evidence arrays are the load-bearing part — note how every claim is sourced (source_class, captured_at, confidence).
Look for: does the canvas pattern survive your team's evidence discipline? If your team's evidence is mostly post-hoc rationalization, Mycelium's evidence-grounding will rub. If your team values "show me the source", this pattern fits.
Concrete signals:
- Your team has been burned by building the wrong thing recently
- You have a junior or new-team member who needs onboarding leverage
- You build at the intersection of design / engineering / product (the canvas is shared documentation)
- You ship products with regulatory or trust surface (security, privacy, accessibility, AI Act exposure)
- You want decision audit to survive personnel turnover
Concrete signals:
- Your work is execution-heavy with bounded scope (build the API endpoint, fix the bug, ship the feature)
- Your team already has strong discovery discipline and you would replace one vocabulary with another
- Your runway is short and you cannot afford an upfront opinionation tax
- You operate in a domain where Mycelium's theory choices do not match your domain (research-heavy science, regulated drug development, etc.)
- The agent harness layer (Claude Code) is incompatible with your tooling
If after the 1h evaluation you want to try it on a real project: see README Quick start. Set the project up, run /interview with sprint or full mode, and commit to one full diamond cycle (1–2 weeks). Most of the value is in the first complete cycle; abandoning halfway through is a worse signal than abandoning at the start.
If after the evaluation you want to push back: open an issue on the Mycelium repo. Pushback that lands a friction surfaces a corrections.md entry; you get named credit per CONTRIBUTORS.md "How to get listed".